Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
The amendments filed 02/09/2026 have been entered. Claims 1-16 remain pending in the application.
Applicant’s amendment, with respect to the claim rejection(s) of claim 1-16 under 35 U.S.C 101 filed 10/15/2025 have been considered and are not persuasive. Therefore, the previous rejections as set forth in the previous office action will be maintained.
The applicant argues that the rejection under 35 U.S.C. § 101 is improper because the amended independent claims are not directed to a mental process, mathematical concept, or generic computer implementation. Applicant asserts that the claims recite a specific computer-implemented knowledge graph embedding training framework, including learning a first knowledge graph embedding model from input knowledge data, extracting embedding vectors from the learned first model, extracting prior knowledge from the embedding vectors, and using the extracted prior knowledge, including a virtual type of an entity, to train a second knowledge graph embedding model by initializing embedding vectors and/or transforming input knowledge data. Applicant contends that these operations cannot practically be performed in the human mind or with pen and paper and therefore should not be characterized as mental processes.
Applicant further argues that, even if the claims are considered to recite a judicial exception, the claims integrate the exception into a practical application because the extracted virtual-type prior knowledge is applied to modify the training of the second knowledge graph embedding model. According to Applicant, this provides a technical improvement to knowledge graph embedding model training and improves model performance, rather than merely collecting, analyzing, or displaying information. Applicant also contends that the ordered combination of claim elements amounts to significantly more than any alleged abstract idea because it recites a particular two-stage model-training pipeline using prior knowledge extracted from a first model to guide initialization and/or transformation for a second model. Applicant relies on USPTO eligibility guidance and Federal Circuit precedent to support the position that the claims are directed to eligible software-based improvements in computer or technical functionality.
The examiner respectfully disagrees. The Applicant’s arguments have been considered but are not persuasive. As discussed in the rejection, the amended limitations are still reasonably characterized as reciting a judicial exception. The claim limitations directed to learning from input knowledge data, extracting embedding vectors, extracting prior knowledge from the embedding vectors, determining or using a virtual type of an entity as prior knowledge, and initializing embedding vectors or transforming input knowledge data based on the prior knowledge/virtual type are directed to mental processes and/or mathematical concepts. Although Applicant argues that the claimed operations are computer-implemented and involve knowledge graph embedding models, the claims recite the operations at a high level of generality and do not set forth any particular unconventional training algorithm, specific model architecture, or technological implementation that changes the character of the claim from an abstract idea into a patent-eligible application.
Applicant’s reliance on the fact that the claim involves “embedding vectors” and “knowledge graph embedding models” does not overcome the rejection. The embedding vectors are numerical representations, and the claimed extraction, use, initialization, and transformation based on those vectors amount to mathematical manipulation and evaluation of information. As recited in the Specification at paragraph 64-65, equation 5 and paragraph 70-77, equation 6, the process of determining a virtual type of an entity to be utilized as the prior knowledge is configured as a mathematical equation at equation 5, and the process of transforming input/initializing embedding vector to learn the model are configured as calculation/mathematical equation 6. The amended recitation that the prior knowledge includes a “virtual type of an entity” further identifies the type of information being used, but does not change the nature of the limitation. Determining and using such type information as prior knowledge remains an abstract mental evaluation or mathematical/data calculation process. The claim does not recite a specific technical manner by which the virtual type improves the operation of the computer itself or improves a particular machine-learning architecture beyond using the information as input for further learning.
Further, the recitation of a first knowledge graph embedding model and a second knowledge graph embedding model does not integrate the judicial exception into a practical application. The models are recited as tools used to perform the abstract learning, extracting, initializing, and transforming steps. The claim does not provide specific details showing how either model is trained in a non-conventional manner, how the model architecture is improved, or how the computer itself operates differently. Rather, the claim uses generic computer implementation and generic machine-learning model terminology to apply the abstract idea to knowledge graph embedding data. Merely applying the abstract idea in the environment of knowledge graph embedding models is insufficient to show integration into a practical application.
Applicant’s assertion that the claimed process improves knowledge graph embedding model performance is also not persuasive because the claim does not recite a specific technological improvement that achieves such alleged improvement. Any improvement appears to result from the abstract use of extracted prior knowledge, including virtual-type information, to guide later training. Such improvement is an improvement to the result of information processing or model output, not a claimed improvement to computer functionality or to a specific technical mechanism for training machine-learning models. The claim therefore remains analogous to collecting, analyzing, communicating and using information to generate or improve a result, without reciting a practical application that imposes a meaningful limit on the judicial exception.
The claim also does not amount to “significantly more” under Step 2B. The additional elements, considered individually and as an ordered combination, merely use a generic computer and generic knowledge graph embedding models to perform the abstract learning, extraction, communication, initialization, and transformation operations. The ordered combination does not add any unconventional computer component, specific technical implementation, or inventive concept beyond the abstract idea itself. Accordingly, the amended claim still recites an abstract idea without integrating the exception into a practical application and without significantly more. Therefore, the rejection under 35 U.S.C. § 101 is maintained.
Applicant’s amendment and argument, with respect to the claim rejection(s) of claim 1-16 under 35 U.S.C 103 filed 10/15/2025 have been considered and are not persuasive. Therefore, the previous rejections as set forth in the previous office action will be maintained.
The applicant argues that the amended independent claim 1, 9 and 16 are patentable because the cited combination of Wang and Costabello does not teach or suggest the newly emphasized feature that the extracted prior knowledge includes a “virtual type of an entity” and that this virtual type is then used to initialize embedding vectors or transform input knowledge data when learning a second knowledge graph embedding model. Applicant contends that Wang is directed to multi-teacher knowledge distillation, where knowledge is transferred from multiple teacher models to a student model by aligning output values/loss functions, but Wang does not teach generate a virtual type entity from embedding results or use such generated type information as input for a subsequent embedding -learning process.
Applicant further argues that Costabello does not cure the deficiency because the art merely uses embeddings and clustering for analysis or classification, not as a cyclic fed back to initialize or transform data for learning another knowledge graph embedding model. For the dependent claim, Applicant similarly argues that Banerji and Jordan only teach general k-means clustering concepts, Xie uses pre-existing hierarchical type information rather than generated virtual types, and Aslan related to knowledge transfer/joint learning among models, but none of these references cure the alleged deficiencies of Wang and Costabello. Therefore, Applicants submits that all 103 rejections should be withdrawn.
The examiner respectfully disagrees. Applicant’s arguments have been considered but are not persuasive. Applicant argues that Wang does not teach extracting prior knowledge comprising a virtual type of an entity from embedding vectors and using the virtual type to initialize embedding vectors or transform input knowledge data for learning a second knowledge graph embedding model. However, the rejection does not rely on Wang alone for this feature. Wang teaches learning a student/junior KGE model using knowledge transferred from teacher KGE models. Costabello teaches generating cluster-derived user/entity information from embedding -based clustering and further using that cluster-derived information in a later training/retraining process.
Costabello discloses at paragraph 32 that clusters of users in relation to products and/or services may be used as additional training data for retraining one or more models, including a knowledge graph embedding model. Thus, Costabello’s clusters are not merely final clustering results or post-processing labels; rather, the cluster-derived information is fed back into further model learning. Costabello also discloses at paragraph 42 that a machine learning model may learn patterns from observations without labeling or supervision and may provide output indicating such patterns by using clustering and/or association to identify related groups of items. A person of ordinary skill in the art would have understood that the further learning/retraining may include such unsupervised learning process, such that the clustered user information may be used to transform input observations into clustered-based output.
The clusters of users in relation to products and/or services corresponds to the claimed prior knowledge comprising a virtual type of the entity because they are derived categories of user/entities generated from embedding-based clustering. Furthermore, those clusters may be used as additional training data for retraining the KGE model, which teaches or at least suggests transforming the input knowledge/training data based on the virtual type of the entity, as claimed.
In view of Wang’s teaching of training a student/junior KGE model using knowledge transferred from teacher KGE models, it would have been obvious to one of ordinary skill in the art to incorporate Costabello’s cluster-derived training information into Wang’s student KGE training process. A person of ordinary skill in the art would have been motivated to do so to provide additional training information for Wang’s student model and improve learning of the student KGE model using derived cluster knowledge from prior embedding/clustering results of the teacher model. The combination would have predictably applied Costabello’s known cluster-based retraining technique in Wang’s known KGE student-model training framework to improve the available training information and perverse learned entity-category/relational information during subsequent KGE model learning. Therefore, Wang in view of Costabello teaches or at least suggest learning the second knowledge graph embedding model by transforming the input knowledge data based on the virtual type of the entity.
Applicant’s argument directed to dependent claims are also not persuasive because the principally rely on the same alleged deficiency argued with respect to the independent claims, namely that the art does not teach the amended virtual-type/prior knowledge limitation. However, as discussed above, Wang in view of Costabello teaches or at least suggests the amended limitation. The additional reference applied to the dependent claims are relied upon for their respective additional dependent features. Accordingly, the dependent claims remain unpatentable for the reasons as set forth in the rejections.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-16 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim 1,
Step 1:
Claim 1 recites a method, one of the four statutory categories of patentable subject matter.
Step 2A, Prong I:
Claim 1 further recites the limitations of:
“performing learning ... based on input knowledge data”. The process of performing learning based on input knowledge data is a mental process. A person can mentally learn from knowledge data such as by learning from a book.
“extracting all embedding vectors ..., and extracting prior knowledge based on the extracted embedding vectors”. The process of extracting all embedding vectors and extracting prior knowledge based on the extracted embedding vectors is a mental process as well as a mathematical concept. A person can mentally extract embedding vectors as embedding vectors are just numeric representation that can be mentally comprehended. The person can mentally extract prior knowledge based on the embedding vector. For example, paragraph 65 of the Specification recites equation 5 to calculate the prior knowledge, which suggest a mathematical equation, and the person can mentally calculate the mathematical equation.
“performing learning of ... through at least one of initialization of the embedding vectors and transform of the input knowledge data based on the extracted prior knowledge” The process of learning through initialization of the embedding vectors and transform of the input knowledge data based on the extracted prior knowledge is a mental process as well as a mathematical concept. A person can mentally perform the calculation to initialize the embedding vectors and transform input data. For example, paragraph 70 of the Specification recites “calculated average vector as an embedding vector initialization value of the entity”, and paragraph 77 of the Specification recites equation 6 to transform the input data. These recitations recite a calculation method and mathematical concepts to calculate the data to perform the learning, wherein the person can mentally perform these calculations and mathematical equations.
“wherein the prior knowledge comprises a virtual type of an entity to be utilized as the prior knowledge,” process of determining a virtual type of an to be utilized as the prior knowledge is a mental process as well as a mathematical concept. A person can mentally perform the calculation to include the virtual type of an entity to be utilized as the prior knowledge. For example, paragraph 64 of the Specification recites “determining a virtual type of an entity to be utilized as the prior knowledge based on the result of clustering as in the following Equation 5”, and paragraph 65 recites equation 5 to include virtual type of an entity within the prior knowledge. These recitations recite a calculation method and mathematical concepts to calculate the data, wherein the person can mentally perform these calculations and mathematical equations
“wherein the performing of the learning ... comprises initializing the embedding vectors, or transforming the input knowledge data based on the virtual type of the entity” The process of learning through initialization of the embedding vectors and transform of the input knowledge data based on the extracted prior knowledge is a mental process as well as a mathematical concept. A person can mentally perform the calculation to initialize the embedding vectors and transform input data. For example, paragraph 70 of the Specification recites “calculated average vector as an embedding vector initialization value of the entity”, and paragraph 77 of the Specification recites equation 6 to transform the input data. These recitations recite a calculation method and mathematical concepts to calculate the data to perform the learning, wherein the person can mentally perform these calculations and mathematical equations.
Step 2A, Prong II:
Claim 1 recites the following additional elements:
“... a method performed by a computer ...” This additional element is a high-level recitation of generic computer components used as a tool, and does not provide integration into a practical application.
“... a first knowledge graph embedding model ...” This additional element recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not provide integration into a practical application. The limitation recites the application of a first knowledge graph embedding model that utilize input data to learn without reciting specific details in how the learning is performed in an unconventional manner or a new machine learning algorithm such that the model can learn. Simply recite the usage a machine learning model that perform the learning based on input data is a demonstration of a conventional black-box application of machine learning model and thus, the additional element does not provide integration into a practical application.
“... the learned first knowledge graph embedding model ...” This additional element recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not provide integration into a practical application. The element simply recites an application of a learned knowledge graph model to obtain a result. The judicial exception of extracting of embedding vector above is based on the learned knowledge graph embedding model, indicate that the knowledge graph is learned. However, the claim does not provide specific details demonstrating how the knowledge graph performs its machine learning in an unconventional manner, or a new learning algorithm for the model. The judicial exception of extracting embedding vector is simply a takeaway of the result from conventional black-box usage of machine learning model and thus, the additional element does not provide integration into a practical application.
“... a second knowledge graph embedding model ...” This additional element recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not provide integration into a practical application. The limitation recites the application of a second knowledge graph embedding model that utilize the learning method of the judicial exception without reciting specific details in how the learning is performed in an unconventional manner or a new machine learning algorithm or a new model’s architecture. Simply recite the usage a machine learning model that perform the learning based on the following judicial without providing further details of the model is just a demonstration of a conventional black-box application of machine learning model and thus, the additional element does not provide integration into a practical application.
Step 2B:
When considered individually or in combination, the additional limitations and elements of claim 1 does not amount to significantly more than the judicial exception for the same reasons discussed above as to why the additional limitations do not integrate the abstract idea into a practical application. The additional elements of outlined in Step 2A performing functions as designed simply accomplishes execution of the abstract ideas.
The additional element “... a method performed by a computer ...” is a high-level recitation of generic computer components used as a tool, and does not amount to significantly more than the judicial exception for the same reasons discussed above as to why the additional limitations do not integrate the abstract idea into a practical application.
The additional element “... a first knowledge graph embedding model ...” recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not amount to significantly more than the judicial exception for the same reasons discussed above as to why the additional limitations do not integrate the abstract idea into a practical application.
The additional element “... the learned first knowledge graph embedding model ...” recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not amount to significantly more than the judicial exception for the same reasons discussed above as to why the additional limitations do not integrate the abstract idea into a practical application.
The additional element “... a second knowledge graph embedding model ...” recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not amount to significantly more than the judicial exception for the same reasons discussed above as to why the additional limitations do not integrate the abstract idea into a practical application.
In conclusions from above for the elements considered as a mental process, and elements reciting a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f) are carried over and do not provide significantly more than the abstract idea. Looking at the limitations in combination and the claims as a whole does not change this conclusion and the claim is ineligible.
Therefore, additional limitations of claim 1 do not amount to significantly more than the judicial exception.
Thus, claim 1 recites abstract ideas with additional elements rendered at a high level of generality resulting in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception.
Therefore, claim 1 is not patent eligible.
Regarding claim 2 depends on claim 1, thus the rejection of claim 1 is incorporated.
Claim 2 recites the element:
“extracting and clustering entity embedding vectors from the learned first knowledge graph embedding model” The limitation recites a mental process. A person can mentally perform the extracting and clustering embedding vectors. For example, paragraph 70 of the specification recites “k entity embedding vectors may be clustered by applying k-means clustering”, which suggests that the clustering can be performed by applying k-means clustering, wherein k-means clustering is a mathematical concept comprises of mathematical equation that can be calculated mentally by the person.
“determining the virtual type of an entity to be utilized as the prior knowledge based on the result of clustering”. The limitation recites a mental process. A person can mentally determine a virtual type of an entity. For example, paragraph 65 of the specification recites equation 5 to calculate type of the entity, which correspond to a mathematical concept comprises of mathematical equation that can be calculated mentally by the person.
Thus, claim 2 recites abstract ideas in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 2 is not patent eligible.
Regarding claim 3 depends on claim 2, thus the rejection of claim 2 is incorporated.
Claim 3 recites the element:
“The learning method of claim 2, further comprising readjusting a search range of a predetermined parameter for clustering as the learning of the second knowledge graph embedding model is completed” The limitation recites a mental process. A person can mentally readjust a search range of a predetermined. Paragraph 81 of the Specification recites “a search range k of a predetermined parameter may be adjusted for clustering. An optimum k value may differ depending on the property of input data, and it is difficult for a user to set the optimum k value every time”. The recitation indicate that it is difficult for a user to set the optimum k value every time, but does not mention that the user cannot set the k range value or a machine learning model or a computer may set the optimum k value automatically, which suggest that while it is difficult, the user can still mentally set the k value search range.
Thus, claim 3 recites abstract ideas in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 3 is not patent eligible.
Regarding claim 4 depends on claim 2, thus the rejection of claim 2 is incorporated.
Claim 4 recites the element:
“The learning method of claim 2, wherein the extracting of the prior knowledge further comprises extracting relation embedding vectors from the learned first knowledge graph embedding model” The limitation recites a mental process. A person can mentally perform the extracting relation embedding vectors. For example, the Specification recites at paragraph 33 “the TransE technique is a method for searching for vector expressions of head (h), tail (t), and relation (r) by forcing a vector of t so as to make the vector sum of h and r equal to the vector of t (h+r=t) when a knowledge triple (h, r, and t) composed of a relation (r)”, which suggest a mathematical concept of TransE comprises of mathematical equation to extract relation embedding vector r, which can be mentally calculated by the person.
“wherein the extracting and clustering of the entity embedding vectors from the learned first knowledge graph embedding model performs clustering the extracted entity embedding vectors and relation embedding vectors” The limitation recites a mental process. A person can mentally perform clustering various embedding vectors. For example, paragraph 70 of the specification recites “k entity embedding vectors may be clustered by applying k-means clustering”, which suggests that the clustering can be performed by applying k-means clustering, wherein k-means clustering is a mathematical concept comprises of mathematical equation that can be calculated mentally by the person.
Thus, claim 4 recites abstract ideas in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 4 is not patent eligible.
Regarding claim 5 depends on claim 2, thus the rejection of claim 2 is incorporated.
Claim 5 recites the element:
“calculating an average vector of the embedding vectors belonging to the same type of cluster for each cluster generated as the result of the clustering” The limitation recites a mental process. A person can mentally calculate an average vector of the embedding vectors of each cluster. The varage calculation is a mathematical concept comprises of mathematical equation that can be calculated by the person.
“determining the calculated average vector as an embedding vector initialization value of the entity belonging to the cluster of the same type” The limitation recites a mental process. A person can mentally determine the calculated average vector as an embedding vector initialization value of the entity of the cluster.
Thus, claim 5 recites abstract ideas in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 5 is not patent eligible.
Regarding claim 6 depends on claim 2, thus the rejection of claim 2 is incorporated.
Claim 6 recites the element:
“transforming a relation of the input knowledge data based on the determined virtual type of the entity” The limitation recites a mental process. A person can mentally transform the relation of the input knowledge data. For example, paragraph 77 of the Specification recites equation 6 to transform the input data. This recitations recite a mathematical concept comprises of a mathematical equation to calculate the data to perform the learning, wherein the person can mentally perform the of the mathematical equation.
Thus, claim 6 recites abstract ideas in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 6 is not patent eligible.
Regarding claim 7 depends on claim 1, thus the rejection of claim 1 is incorporated.
Claim 7 recites the element:
“The learning method of claim 1, wherein the first knowledge graph embedding model and the second knowledge graph embedding model are the same kinds of knowledge graph embedding models” This additional element recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not provide integration into a practical application. The limitation recites the application of two knowledge graph embedding model that are the same kind, without providing specific architecture or unconventional machine learning algorithm of the model to perform machine learning or improvement to computer element. Simply recite the application of two similar models is just a demonstration of a conventional black-box application of machine learning model and thus, the additional element does not provide integration into a practical application.
Thus, claim 7 recites additional elements rendered at a high level of generality resulting in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 7 is not patent eligible.
Regarding claim 8 depends on claim 1, thus the rejection of claim 1 is incorporated.
Claim 8 recites the element:
“The learning method of claim 1, wherein the first knowledge graph embedding model and the second knowledge graph embedding model are different kinds of knowledge graph embedding models” This additional element recites an additional element of a mere instruction to apply an exception with a recitation of the words "apply it" (or an equivalent) as identified in MPEP 2106.05(f), and does not provide integration into a practical application. The limitation recites the application of two knowledge graph embedding model that are different kinds, without providing specific architecture or unconventional machine learning algorithm of the model to perform machine learning or improvement to computer element. Simply recite the application of two different models is just a demonstration of a conventional black-box application of machine learning model and thus, the additional element does not provide integration into a practical application.
Thus, claim 8 recites additional elements rendered at a high level of generality resulting in claims that do not integrate the abstract idea into a practical application or amount to significantly more than the judicial exception. Therefore, claim 8 is not patent eligible.
Regarding claim 9 which recite a system, one of the four statutory categories of patentable subject matter. Claim 9 is rejected under the same rationale of claim 1 because the claim recites similar limitations and processing steps to claim 1.
Regarding claim 10 depends on claim 9, thus the rejection of claim 9 is incorporated. Claim 10 is rejected under the same rationale of claim 2 because the claim recites similar limitations and processing steps to claim 2.
Regarding claim 11 depends on claim 10, thus the rejection of claim 10 is incorporated. Claim 11 is rejected under the same rationale of claim 3 because the claim recites similar limitations and processing steps to claim 3.
Regarding claim 12 depends on claim 10, thus the rejection of claim 10 is incorporated. Claim 12 is rejected under the same rationale of claim 5 because the claim recites similar limitations and processing steps to claim 5.
Regarding claim 13 depends on claim 10, thus the rejection of claim 10 is incorporated. Claim 13 is rejected under the same rationale of claim 6 because the claim recites similar limitations and processing steps to claim 6.
Regarding claim 14 depends on claim 9, thus the rejection of claim 9 is incorporated. Claim 14 is rejected under the same rationale of claim 7 because the claim recites similar limitations and processing steps to claim 7.
Regarding claim 15 depends on claim 9, thus the rejection of claim 9 is incorporated. Claim 15 is rejected under the same rationale of claim 8 because the claim recites similar limitations and processing steps to claim 8.
Regarding claim 16 which recite a system, one of the four statutory categories of patentable subject matter.
Claim 16 recites additional elements “an input module ...”, “a memory...”, “a processor” which are high-level recitation of generic computer components used as a tool and does not provide integration into a practical application or significantly more than the abstract idea.
Claim 16 is rejected under the same rationale of claim 1 because the claim recites similar limitations and processing steps to claim 1.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 2, 4, 9, 10, 16 are rejected under 35 U.S.C. 103 as being unpatentable over Wang et.al (NPL: MulDE: Multi-teacher Knowledge Distillation for Low-dimensional Knowledge Graph Embeddings) in view of Costabello et.al (US 20220188850 A1)
Regarding claim 1,
Wang teaches or at least suggests the limitation “performing learning of a first knowledge graph embedding model based on input knowledge data” (page 1 column 1 “we propose MulDE, a novel knowledge distillation framework, which includes multiple low-dimensional hyperbolic KGE models as teachers and two student components, namely Junior and Senior”, page 2 section 1 column 1 “To this end, we determine to employ those low-dimensional hyperbolic models as multiple teachers and integrate their knowledge to train a smaller KGE model in a Knowledge Distillation (KD) process ... In this paper, we pre-train multiple low-dimensional hyperbolic KGE models”, page 2 section 2 column 2 “Knowledge Graph Embeddings aim to represent each entity e ∈ E and each relation r ∈ R as d-dimensional continuous vectors ... For each candidate triple (ein, r, ec), M learns the relation vector r as a transformation between two entity vectors”, and page 4-5 column 1-2 section 3.1 “Taking the RotH model as an example, it uses a d−dimensional Poincaré ball model with trainable negative curvature. Embedding vectors are first mapped into this hyperbolic space, and a relation vector is regarded as a rotation transformation of entity vectors” Wang discloses a novel knowledge distillation framework comprises one or more teacher model and two student model, which are junior and senior model wherein the one or more teacher model corresponds to the first knowledge graph embedding model within the claim. The teacher model is pre-trained based on the preliminaries in section 2 such that given one or more triple (input), the model learns the relation embedding vector between two entity vectors as understood by one of ordinary skilled in the art. Wang provides example of the pre-trained teacher model such as TransH, RotH, etc.)
Wang teaches or at least suggests a part of the limitation “performing learning of a second knowledge graph embedding model ... based on the extracted prior knowledge”. (page 4 section 3.2 column 1 “Our MulDE is a general framework, in which the Junior component can be any existing KGE model. In this paper, we focus on the low-dimensional situation, thus selecting several effective low-dimensional models as Junior, whose initialization follows their original settings, e.g., a random normal distribution”. Wang discloses a novel knowledge distillation framework comprises one or more teacher model and two student model, which are junior and senior model wherein the one or more teacher model corresponds to the first knowledge graph embedding model within the claim and the junior model corresponds to the second knowledge graph embedding model. The junior model is learned based on knowledges distilled from the teacher model that is passed through the senior model for further processing, which suggest the extracted prior knowledge used for learning of the second model, as claimed.)
Wang does not teach the limitation “extracting all embedding vectors from the learned first knowledge graph embedding model, and extracting prior knowledge based on the extracted embedding vectors”. However, Costabello teaches or at least suggests this limitation (paragraph 24 “As shown in FIG. 1D, and by reference number 125, the clustering system may train the sequence embeddings and the knowledge graph embeddings jointly, to generate fine-tuned user embeddings that capture temporal and non-temporal/relational information", and paragraph 28 “As shown in FIG. 1E, and by reference number 130, the clustering system may process the fine-tuned user embeddings, with a clustering model, to determine clusters of the users in relation to products and/or services purchased by the users. The clustering model may include a k-means clustering model, ... In some implementations, the clustering model may identify the clusters of the users in relation to the products and/or services and may assign labels to the clusters. For example, for alcohol products, the clustering model may determine a cluster for casual drinkers, a cluster for beer geeks, a cluster for social drinkers, and/or the like.” Costabello discloses a system and method of jointly learning a knowledge graph embedding model with a LSTM based encoder-decoder model and clustering. Costabello discloses the generation of the fine-tuned user embeddings is based on jointly training the sequence embeddings (results of LSTM based encoder-decoder model based on temporal data) and the knowledge graph embeddings (results of trained knowledge graph embedding model based on non-temporal data associated with the users) as illustrated in fig. 1D, which suggest an extraction of all embedding vectors from the learned knowledge graph embedding model above to generate the fine-tuned user embeddings. Costabello further discloses a clustering process based on the fine-tuned user embeddings, and subsequent processing of those clusters as illustrated in fig. 1E, and fig.1F. The clustering process determine one or more clusters of the users (entity) in relation to products and/or services purchased by the users based on the fine-tuned user embeddings and further use these clusters for retraining or subsequent processing, which corresponds to the process of extracting prior knowledge based on the extracted embedding vectors, as claimed. One of ordinary skilled in the art may configure the embedding knowledge graph model as disclosed herein to correspond to the teacher model as disclosed by Wang above based on the teaching motivation to combine the teachings below.)
Wang does not teach the limitation “wherein the prior knowledge comprises a virtual type of an entity to be utilized as the prior knowledge” However, Costabello teaches or at least suggests this limitation (paragraph 28, and Fig. 1F “As shown in FIG. 1E, and by reference number 130, the clustering system may process the fine-tuned user embeddings, with a clustering model, to determine clusters of the users in relation to products and/or services purchased by the users. The clustering model may include a k-means clustering model, ... In some implementations, the clustering model may identify the clusters of the users in relation to the products and/or services and may assign labels to the clusters. For example, for alcohol products, the clustering model may determine a cluster for casual drinkers, a cluster for beer geeks, a cluster for social drinkers, and/or the like.” Costabello discloses the clustering process to determine one or more clusters of the users in relation to products and/or services purchased by the users based on the fine-tuned user embeddings, which corresponds to the process of extracting knowledge comprising a virtual type of an entity, as claimed. The user corresponds to the claimed entity, and the resulting clusters based on the user with relation to products/services corresponds to the knowledge comprising virtual type of the entity, because it is a derived category of the user/entity generated based on categorizing the clustering of embedding vectors. Furthermore, as illustrated in figure 1F, the system can further use those cluster knowledge to retrain one or more of the KGE model, thereby suggesting that the cluster knowledge corresponds to the prior knowledge in a subsequent training or retraining process. In view of Wang’s teacher/student KGE framework, it would have been obvious to use Costabello’s cluster-derived knowledge in training Wang’s subsequent/student KGE model, such that the student model (second model) is learned based on the categorized cluster knowledge of the teacher model (first model).)
Wang in view of Costabello teaches or at least suggest “performing learning ... through at least one of initialization of the embedding vectors and transformation of the input knowledge data based on the extracted prior knowledge” (Costabello at paragraph 32 “The clustering system may utilize the clusters of the users as additional training data for retraining ..., the knowledge graph embedding model, or the clustering model, thereby increasing the quantity of training data available for training”, paragraph 42 “In some implementations, ... the machine learning model may learn patterns from the set of observations without labeling or supervision, and may provide output that indicates such patterns, such as by using clustering and/or association to identify related groups of items within the set of observations” Costabello discloses the clusters of the users in relation to products and/or services may be used as additional training data for retraining one or more model, which corresponds to the training using the extracted prior knowledge comprising a virtual type of the entity. Furthermore, Costabello discloses the learning process at paragraph 42 that machine learning model may learn patterns from observations without labeling or supervision and may provide output by using clustering. A person of ordinary skill in the art would have understood that the further learning/retraining may include the unsupervised learning process such that the clustered user information may be used to transform input observations into cluster-based output. Accordingly, Costabello teaches or at least suggests performing learning by transforming the input knowledge data based on the virtual type of the entity, as claimed. In view of Wang’s teaching at page 2 section 3 “The proposed Multi-teacher Distillation Embedding (MulDE) frame work utilizes multiple pre-trained low-dimensional models as teachers. Under a novel iterative distillation strategy, it integrates prediction sequences from different teachers, and supervises the training process of a low-dimensional student model”, Wang discloses the student model is trained using knowledge transferred from the teacher models. In view of Costabello’s teaching that cluster-derived user information may be used as additional training data for retraining a KGE model, a person of ordinary skill in the art would have been motivated to incorporate Costabello’s cluster-derived training information into Wang’s student KGE training process to provide additional training information for Wang’s student model and improve the learning of the student KGE model using derived knowledge from prior embedding/clustering results. Accordingly, Wang in view of Costabello teaches or at least suggests learning the second knowledge graph embedding model by transforming the input knowledge data based on the virtual type of entity, as claimed.)
Wang in view of Costabello teaches or at least suggest “wherein the performing of the learning of the second knowledge graph embedding model comprises initializing the embedding vectors, or transforming the input knowledge data based on the virtual type of the entity” (Costabello at paragraph 32 “The clustering system may utilize the clusters of the users as additional training data for retraining ..., the knowledge graph embedding model, or the clustering model, thereby increasing the quantity of training data available for training”, paragraph 42 “In some implementations, ... the machine learning model may learn patterns from the set of observations without labeling or supervision, and may provide output that indicates such patterns, such as by using clustering and/or association to identify related groups of items within the set of observations” Costabello discloses the clusters of the users in relation to products and/or services may be used as additional training data for retraining one or more model, which corresponds to the training using the extracted prior knowledge comprising a virtual type of the entity. Furthermore, Costabello discloses the learning process at paragraph 42 that machine learning model may learn patterns from observations without labeling or supervision and may provide output by using clustering. A person of ordinary skill in the art would have understood that the further learning/retraining may include the unsupervised learning process such that the clustered user information may be used to transform input observations into cluster-based output. Accordingly, Costabello teaches or at least suggests performing learning by transforming the input knowledge data based on the virtual type of the entity, as claimed. In view of Wang’s teaching at page 2 section 3 “The proposed Multi-teacher Distillation Embedding (MulDE) frame work utilizes multiple pre-trained low-dimensional models as teachers. Under a novel iterative distillation strategy, it integrates prediction sequences from different teachers, and supervises the training process of a low-dimensional student model”, Wang discloses the student model is trained using knowledge transferred from the teacher models. In view of Costabello’s teaching that cluster-derived user information may be used as additional training data for retraining a KGE model, a person of ordinary skill in the art would have been motivated to incorporate Costabello’s cluster-derived training information into Wang’s student KGE training process to provide additional training information for Wang’s student model and improve the learning of the student KGE model using derived knowledge from prior embedding/clustering results. Accordingly, Wang in view of Costabello teaches or at least suggests learning the second knowledge graph embedding model by transforming the input knowledge data based on the virtual type of entity, as claimed.)
Before the effective filing date, it would have been obvious to one of ordinary skilled in the art
to combine the teaching of a novel knowledge distillation framework comprises learning of one or more teacher knowledge graph embedding model and the knowledge graph embedding junior student model by Wang, with the teaching of obtaining the embedding vector from the learned knowledge graph embedding model for fine-tuning and clustering to determine clusters correspond to user and their relations and initialization by Costabello. The motivation to do so is referred to in Costabello’s disclosure (paragraph 15 “the clustering system utilizes machine learning models for data-driven customer segmentation. The clustering system may provide a machine-learning driven system to automatically cluster product customers that authored sequences of product reviews in an online community service. The clustering system may provide a mechanism for determining friendships among the product customers. Furthermore, the clustering system efficiently provides data representation and temporal feature modeling, with unlabeled data (e.g., without a need for human labels). This, in turn, conserves computing resources, networking resources, and/or the like that would otherwise have been consumed in generating incorrect customer segments, generating and implementing ineffective product marketing campaigns based on the incorrect customer segments, generating and implementing ineffective targeted advertising based on the incorrect customer segments, discovering and correcting the incorrect customer segments, and/or the like.”, and paragraph 23 “the clustering system may convert the non-temporal data associated with the users, into a knowledge graph, and process the knowledge graph with a knowledge graph embedding model to capture trained knowledge graph embeddings. A knowledge graph embedding represents entities and relations, in a knowledge graph” Costabello discloses the benefit of obtaining the fine-tuned user embeddings based on the knowledge graph embeddings to determine cluster corresponding to clients (entities) and their activity’s history (prior knowledge), which help a user learn the relationship between multiple clients and efficiently provides data representation and temporal feature modeling while conserves computing resources, networking resources and avoid incorrect marketing toward incorrect client segments. In other word, the clustering based on knowledge graph embedding provide several benefits in data manipulation of real-world client. One of ordinary skilled in the art would have been motivated to modify the knowledge distillation framework by Wang to incorporate the clustering system by Costabello, such that the embedding vectors resulted from learning at the teacher model may be clustered by the clustering system for further utilizing for subsequent learning of the student model, thereby improve the overall framework. Furthermore, while Wang discloses the initialization of the junior student model but in a random manner, the teacher model by Wang may incorporate the LSTM model by Costabello to perform the initialization of embedding vectors, wherein one of ordinary skilled in the art would have been motivated to configure the initialized embedding vector of the teacher model as part of the knowledge distilled to the junior student model using the knowledge distillation framework such that the junior student model may begin learning with initialized embedding vectors of entities whose relational structure have been captured by the teacher model, thereby enabling faster convergence and preserve relationships as would be understood by one of ordinary skilled in the art.)
Regarding claim 2 depends on claim 1, thus the rejection of claim 1 is incorporated.
Costabello teaches the limitation “extracting and clustering entity embedding vectors from the learned first knowledge graph embedding model” (paragraph 24 “As shown in FIG. 1D, and by reference number 125, the clustering system may train the sequence embeddings and the knowledge graph embeddings jointly, to generate fine-tuned user embeddings that capture temporal and non-temporal/relational information", and paragraph 28 “As shown in FIG. 1E, and by reference number 130, the clustering system may process the fine-tuned user embeddings, with a clustering model, to determine clusters of the users in relation to products and/or services purchased by the users. The clustering model may include a k-means clustering model, ... In some implementations, the clustering model may identify the clusters of the users in relation to the products and/or services and may assign labels to the clusters. For example, for alcohol products, the clustering model may determine a cluster for casual drinkers, a cluster for beer geeks, a cluster for social drinkers, and/or the like”. Costabello discloses the generation of the fine-tuned user embeddings is based on jointly training the sequence embeddings (results of LSTM based encoder-decoder model based on temporal data associated with the users) and the knowledge graph embeddings (results of trained knowledge graph embedding model based on non-temporal data associated with the users) as illustrated in fig. 1D, which suggest an extraction of all embedding vectors from the learned knowledge graph embedding model above to generate the fine-tuned user embeddings. Costabello further discloses a clustering process based on the fine-tuned user embeddings as illustrated in fig. 1E. The clustering process determine one or more clusters of the users (entity) in relation to products and/or services purchased by the users (prior knowledge) based on the fine-tuned user embeddings of the learned knowledge graph embedding model, which is analogous to the clustering entity embedding vectors from the learned first knowledge graph embedding model.
Costabello teaches the limitation “determining a virtual type of an entity to be utilized as the prior knowledge based on the result of clustering” (paragraph 28, and figure 1F “The clustering model may include a k-means clustering model, ... In some implementations, the clustering model may identify the clusters of the users in relation to the products and/or services and may assign labels to the clusters. For example, for alcohol products, the clustering model may determine a cluster for casual drinkers, a cluster for beer geeks, a cluster for social drinkers, and/or the like.” Costabello discloses the clustering process to provide different clusters in relation with a purchased products and/or services of users (entity), which suggests the virtual type of the entity utilized as the prior knowledge based on the result of clustering, as claimed, wherein after determining clusters of users from embeddings, the system can further use those cluster knowledge to retrain one or more of the KGE model, thereby suggesting that the cluster knowledge corresponds to the prior knowledge in a subsequent training or retraining process of a KGE model. In view of Wang’s teacher/student KGE framework, it would have been obvious to use Costabello’s cluster-derived knowledge in training Wang’s subsequent/student KGE model, such that the student model (second model) is learned based on the categorized cluster knowledge of the teacher model (first model).)
Regarding claim 4 depends on claim 2, thus the rejection of claim 2 is incorporated.
Costabello teaches the limitation “The learning method of claim 2, wherein the extracting of the prior knowledge further comprises extracting relation embedding vectors from the learned first knowledge graph embedding model” (paragraph 23 “A knowledge graph embedding represents entities and relations”, and paragraph 24 “In some implementations, the clustering system may jointly train the sequence embeddings and the knowledge graph embeddings to generate the fine-tuned user embeddings that capture both temporal and non-temporal/relational information.” Costabello discloses a knowledge graph embedding represents entities and relations, suggesting the learned knowledge graph embedding model provide embedding vectors represent entities and relations. The fine-tuned user embeddings that capture both temporal and non-temporal/relational information from the knowledge graph embedding suggesting embeddings of relations is obtained.)
Costabello teaches the limitation “wherein the extracting and clustering of the entity embedding vectors from the learned first knowledge graph embedding model performs clustering the extracted entity embedding vectors and relation embedding vectors” (paragraph 28 “the clustering system may process the fine-tuned user embeddings ... The clustering model may include a k-means clustering model, ... In some implementations, the clustering model may identify the clusters of the users in relation to the products and/or services and may assign labels to the clusters. For example, for alcohol products, the clustering model may determine a cluster for casual drinkers, a cluster for beer geeks, a cluster for social drinkers, and/or the like.” Costabello discloses the clustering process identify the clusters of the users in relation to the products and/or services, which utilize fine-tuned user embeddings that include relation and entity embeddings as explained above.)
Regarding claim 9 which recite a system, one of the four statutory categories of patentable subject matter. Claim 9 is rejected under the same rationale of claim 1 because the claim recites similar limitations and processing steps to claim 1.
Regarding claim 10 depends on claim 9, thus the rejection of claim 9 is incorporated. Claim 10 is rejected under the same rationale of claim 2 because the claim recites similar limitations and processing steps to claim 2.
Regarding claim 16,
Costabello teaches the limitation “an input module configured to receive input knowledge data” (paragraph 62 “The input component 440 enables the device 400 to receive input, such as user input and/or sensed inputs.” Costabello discloses the input component to receive input, wherein the input can be knowledge input from user for knowledge graph embedding learning.)
Costabello teaches the limitation “a memory configured to store therein a program ...” (paragraph 63 “For example, a non-transitory computer-readable medium (e.g., the memory 430) may store a set of instructions (e.g., one or more instructions, code, software code, program code, and/or the like) for execution by the processor 420. The processor 420 may execute the set of instructions to perform one or more processes described herein.” Costabello discloses a non-transitory computer-readable medium such as a memory to store the program code that can be executed by the processor, wherein one of ordinary skilled in the art can configure a computer with the memory and processor by Wang below to perform the combination method of claim 1.)
Wang teaches the limitation “a processor configured to perform ...” (page 6 section 4.1 “Implementation Details.All experiments are performed on Intel Core i7-7700K CPU @ 4.20GHz and NVIDIA GeForce GTX1080 Ti GPU, and are implemented in Python using the PyTorch deep learning framework” Wang discloses using a processor CPU to implement the method of learning knowledge graph embedding models.)
Claim 16 is further rejected under the same rationale of claim 1 because the claim recites similar limitations and processing steps to claim 1. The motivation to combine the teaching of Wang with Costabello is similar to the motivation in claim 1.
Claims 3, 11 are rejected under 35 U.S.C. 103 as being unpatentable over Wang et.al (NPL: MulDE: Multi-teacher Knowledge Distillation for Low-dimensional Knowledge Graph Embeddings) in view of Costabello et.al (US 20220188850 A1), further in view of Banergi et.al (NPL: Selecting Number Of Clusters in K-Mean Clustering)
Regarding claim 3 depends on claim 2, thus the rejection of claim 2 is incorporated.
Wang/Costabello does not teach the limitation “The learning method of claim 2, further comprising readjusting a search range of a predetermined parameter for clustering as the learning of the second knowledge graph embedding model is completed”. However, Banergi teaches this limitation (Page 2“In this blog, we will discuss the most important parameter i.e., the ways by which we can select number of clusters (K)”, page 3-4 section 1 “Elbow Curve Method The elbow method runs k-means clustering on the dataset for a range of values of k (say 1 to 10) ... The curve looks like an elbow. In the above plot, elbow is at k=3(i.e. Sum of squared distances falls suddenly) indicating the optimal k for this dataset is 3” Banergi discloses a method of selecting number of clusters in K-Mean Clustering. Within the disclosure, Banergi discloses the Elbow Curve Method, which utilize the k-means clustering algorithm with an initial range of cluster and calculate the sum of the squared distances between each data point and its assigned cluster's centroid to plot a curve to determine a good estimate for the optimal number k of clusters. One of ordinary skilled in the art would have been able to configure a new range based on the k-value to further refine the clustering based on evaluating the plot of the elbow curve. The implementation of the elbow curved method that indicate a new k value for clustering based on the initial range to further determine a new range is analogous to the claimed readjusting a search range of a predetermined parameter for clustering. Furthermore, one of ordinary skilled in the art would have been able to configure the learning of the junior student knowledge graph embedding model by Wang in view of the clustering process by Costabello to determine clusters based on learned embeddings of the model using K-mean clustering technique and further apply the elbow curved method to indicate a range of the cluster based on the motivation to determine the range below.)
Wang/Costabello does not teach the limitation “wherein the extracting and clustering of the entity embedding vectors from the learned first knowledge graph embedding model performs the clustering based on the adjusted parameter” (Page 2“In this blog, we will discuss the most important parameter i.e., the ways by which we can select number of clusters (K)”, page 3-4 section 1 “Elbow Curve Method The elbow method runs k-means clustering on the dataset for a range of values of k (say 1 to 10) ... The curve looks like an elbow. In the above plot, elbow is at k=3(i.e. Sum of squared distances falls suddenly) indicating the optimal k for this dataset is 3” Banergi discloses the Elbow Curve Method to determine the optimal number k of cluster for the k-mean clustering process. One of ordinary skilled in the art would have been able to configure the framework by Wang, wherein knowledge can be transferred from the junior student model to the teacher model as illustrated in Fig.2, wherein the transferred knowledge may be configured to include the k value of optimal number of clusters for the teacher model to improve the learning at the teacher model.)
Before the effective filing date, it would have been obvious to one of ordinary skilled in the art
to combine the teaching of a novel knowledge distillation framework comprises learning of one or more teacher knowledge graph embedding model and the knowledge graph embedding junior student model by Wang, and the teaching of obtaining the embedding vector from the learned knowledge graph embedding model for fine-tuning and clustering to determine clusters correspond to user and their relations and initialization by Costabello, with the teaching of the Elbow Curved method that determine a range of k value for clustering by Banergi. The motivation to do so is referred to in Banergi’s disclosure (page 1-2 “K-means clustering is an unsupervised algorithm. In unsupervised algorithm, we are not interested in making predictions (since we don’t have target/output variable). There are certain factors that can impact the efficacy of the final clusters formed when using k-means clustering. So, we must keep in mind the following factors when solving business problems using K-means clustering algorithm. 1. Number of clusters (K): The number of clusters you want to group your data points into, has to be predefined. ... we will discuss the most important parameter i.e., the ways by which we can select number of clusters (K)”. Banergi discloses the importance of factors that affect the efficacy of the final clusters formed when using k-means clustering such as the number of clusters K. While the teaching by Wang in view of Costabello utilize K-mean clustering for cluster the embeddings as disclosed by Costabello at paragraph 28 “As shown in FIG. 1E, and by reference number 130, the clustering system may process the fine-tuned user embeddings, with a clustering model, to determine clusters of the users in relation to products and/or services purchased by the users. The clustering model may include a k-means clustering model”. Therefore, the teaching by Wang/Costabello can further incorporate the method by Banergi to further evaluate the cluster for more precise clustering.)
Claims 5, 12 are rejected under 35 U.S.C. 103 as being unpatentable over Wang et.al (NPL: MulDE: Multi-teacher Knowledge Distillation for Low-dimensional Knowledge Graph Embeddings) in view of Costabello et.al (US 20220188850 A1) further in view of Jordan et.al (NPL: Grouping data points with k-means clustering.)
Regarding claim 5 depends on claim 2, thus the rejection of claim 2 is incorporated.
Wang/Costabello does not teach the limitation “calculating an average vector of the embedding vectors belonging to the same type of cluster for each cluster generated as the result of the clustering”. However, Jordan teaches this limitation (page 1 “K-means clustering is a simple method for partitioning data points in groups, or clusters. Essentially, the process goes as follows: 1. Select centroids. These will be the center point for each segment. centroid 2. Assign data points to nearest centroid. centroid 3. Reassign centroid value to be the calculated mean value for each centroid cluster. 4. Reassign data points to nearest centroid. centroid 5. Repeat until data points stay in the same cluster”, and page 6 “After each data point is assigned to a cluster, reassign the centroid value for centroid each cluster to be the mean value of all the data points within the cluster.” Jordan discloses the method of k-mean clustering, including a step of calculating the mean value of all the data points within the cluster to generate a new centroid value that represent each cluster for further processing, wherein the calculation of the mean value is based on each cluster from the clustering process with an initial centroid selected based on a number of cluster k determined. Jordan further provides illustration of grouping data points with k-means clustering, suggesting the data point to be vector data point. The calculation of the mean value to represent the centroid of each cluster based on the clustering is analogous to the calculating an average vector of the embedding vectors for each cluster generated as the result of the clustering within the claim.)
Wang/Costabello does not teach the limitation “determining the calculated average vector as an embedding vector initialization value of the entity belonging to the cluster of the same type”. However, Jordan teaches this limitation (page 6 “Reassign data points to new clusters This is where the iterative process begins. Follow the same process for initially assigning data points to clusters, this time with new centroid values. Iteratively calculate new centroid values and reassign data points until you converge on a cluster segmentation that stops changing” Jordan discloses the iterative process of the k-mean clustering method, in which the mean value of centroid for each cluster is now considered as a replacement for the initial centroid selected. This centroid mean value may be iteratively calculated until convergence on cluster segmentation is achieved. The iterative process in which a new calculated centroid value is iteratively utilized to determine optimal cluster segmentation is analogous to the determination of the calculated average vector as an embedding vector initialization value within the claim, because the calculated centroid mean value is reused as the initial centroid value such that distance between data point within the cluster is calculated using this new centroid value until optimal cluster segmentation is achieved.)
Before the effective filing date, it would have been obvious to one of ordinary skilled in the art to combine the teaching of a novel knowledge distillation framework comprises learning of one or more teacher knowledge graph embedding model and the knowledge graph embedding junior student model by Wang, and the teaching of obtaining the embedding vector from the learned knowledge graph embedding model for fine-tuning and clustering to determine clusters correspond to user and their relations and initialization by Costabello, with the teaching of K-mean clustering method including the calculation of the centroid mean value as the initial value in an iterative process by Jordan. The motivation to do so is referred to in Jordan’s disclosure (page 1 “K-means clustering is a simple method for partitioning data points in groups, or clusters. Essentially, the process goes as follows: 1. Select centroids. These will be the center point for each segment. centroid 2. Assign data points to nearest centroid. centroid 3. Reassign centroid value to be the calculated mean value for each centroid cluster. 4. Reassign data points to nearest centroid. centroid 5. Repeat until data points stay in the same cluster”, Jordan discloses a simple method of K-means clustering for partitioning data points into clusters. Jordan further provides the explanation of how the k-mean clustering method is performed, in which a mean value is calculated to represent the centroid of each cluster, which is then further used as the initial centroid value to further determine the best cluster. One of ordinary skilled in the art would have been able to incorporate the teaching by Jordan into the teaching by Wang in view of Costabello to implement the k-mean clustering method and calculation to obtain clusters of embedding vectors for entities and relations that are learned from the teacher model. The cluster would provide a user easier understanding of data for further evaluation of the learning of the knowledge graph embedding model. Costabello also discloses a k-mean clustering algorithm to cluster data. Therefore, one of ordinary skilled in the art would have been motivated to incorporate the teaching by Jordan into the teaching of Wang in view of Costabello to further explain and perform the k-mean clustering method as explained by Jordan.)
Regarding claim 12 depends on claim 10, thus the rejection of claim 10 is incorporated. Claim 12 is rejected under the same rationale of claim 5 because the claim recites similar limitations and processing steps to claim 5.
Claims 6, 13 are rejected under 35 U.S.C. 103 as being unpatentable over Wang et.al (NPL: MulDE: Multi-teacher Knowledge Distillation for Low-dimensional Knowledge Graph Embeddings) in view of Costabello et.al (US 20220188850 A1) further in view of Xie et.al (NPL: Representation Learning of Knowledge Graphs with Hierarchical Types)
Regarding claim 6 depends on claim 2, thus the rejection of claim 2 is incorporated.
Wang teaches a part of the limitation “performing the learning of the second knowledge graph embedding model ...” (page 4 section 3.2 column 1 “Our MulDE is a general framework, in which the Junior component can be any existing KGE model. In this paper, we focus on the low-dimensional situation, thus selecting several effective low-dimensional models as Junior, whose initialization follows their original settings, e.g., a random normal distribution”. Wang discloses a novel knowledge distillation framework comprises one or more teacher model and two student model, wherein the junior model corresponds to the second knowledge graph embedding model. The junior model is learned based on knowledge distilled from the teacher model that is passed through the senior model for further processing. One of ordinary skilled in the art would have been able to configure the learning of the junior model using the transformed input based on the teaching of ..., and the teaching motivation to use the transformed input below.)
Wang/Costabello does not teach the limitation “transforming a relation of the input knowledge data based on the determined virtual type of the entity”. However, Xie teaches this limitation (page 3 section 3.4 “negative sampling as objective for training ... E(h, r, t) is the energy function score of positive triple and E(h’ , r’ , t’ ) is that of negative triple ... we also add relation replacements to negative sampling for better performances in relation prediction. Moreover, the new triples after replacements will not be considered as negative samples if they are already in T”, and page 4 section 3.5 column 1 “Besides we can use type information not only as projection matrices but also as constraints with the help of relation-specific type information.”, Xie discloses a method of Type-embodied Knowledge Representation Learning to take advantages of hierarchical entity types. Within the disclosure, Xie discloses a technique of generating negative triple by replacing the relation embedding vector r into a negative relation embedding vector r’, where in the replacement of the positive relation r with the negative relation r’ is analogous to transforming a relation of the input knowledge data based within the claim. The negative relation determination is based on the entity type as the type information can be used as constraints with the help of relation-specific type information to determine a valid but incorrect negative triple.)
Wang/Costabello does not teach a part of the limitation “performing the learning... based on the input knowledge data of which the relation has been transformed” However, Xie teaches this limitation (page 4 section 4.1 “In evaluation, we implement TransE and TransR for comparison. For TransE, we improve their dissimilarity measure with ... replace relations as well as entities during negative sampling” Xie discloses the experiment, in which two knowledge graph embedding model such as TransE and TransR is utilized for learning with the negative set of triples, including the negative relation embedding vector. One of ordinary skilled in the rat would have been motivated to similarly use the negative triple set resulted from the original triple for the learning of the junior knowledge graph embedding model based on the combination with the teaching by Xie and the motivation below.)
Before the effective filing date, it would have been obvious to one of ordinary skilled in the art to combine the teaching of a novel knowledge distillation framework comprises learning of one or more teacher knowledge graph embedding model and the knowledge graph embedding junior student model by Wang, and the teaching of obtaining the embedding vector from the learned knowledge graph embedding model for fine-tuning and clustering to determine clusters correspond to user and their relations and initialization by Costabello, with the teaching of a technique of generating negative triple by replacing the relation embedding vector for knowledge graph embedding learning by Xie. The motivation to do so is referred to in Xie’s disclosure (page 3 section 3.4 “negative sampling as objective for training ... E(h, r, t) is the energy function score of positive triple and E(h’ , r’ , t’ ) is that of negative triple ... we also add relation replacements to negative sampling for better performances in relation prediction” Xie discloses the replacements to negative sampling create better performance in relation prediction. One of ordinary skilled in the art would have recognized the benefit of the negative sampling to evaluate the learning of a knowledge graph embedding model for better performance. For example, the junior student model in Wang may utilize the technique by Xie to create a new triple with negative relation embedding vector for evaluation through performing prediction of missing link and obtain a result that indicate performance of the learned model.)
Regarding claim 13 depends on claim 10, thus the rejection of claim 10 is incorporated. Claim 13 is rejected under the same rationale of claim 6 because the claim recites similar limitations and processing steps to claim 6.
Claims 7, 8, 14, 15 are rejected under 35 U.S.C. 103 as being unpatentable over Wang et.al (NPL: MulDE: Multi-teacher Knowledge Distillation for Low-dimensional Knowledge Graph Embeddings) in view of Costabello et.al (US 20220188850 A1) further in view of Aslan et.al (US 20170132528 A1)
Regarding claim 7 depends on claim 1, thus the rejection of claim 1 is incorporated.
Wang/Costabello does not teach the limitation “The learning method of claim 1, wherein the first knowledge graph embedding model and the second knowledge graph embedding model are the same kinds of knowledge graph embedding models”. However, Aslan teaches the limitation (paragraph 45 “Notably, the multiple models that are jointly trained can be of the same, or similar, ... yet the architecture can be optimized in at least one of the models for deployment purposes”, and paragraph 47 “some implementations, various ensembles of teacher models and/or ensembles of student models can be utilized with the joint training techniques and systems described herein. FIG. 2 is a schematic diagram of an example technique for joint training of multiple machine learning models involving an ensemble of N “teacher” models 200 ... the student model 202 is to be jointly trained in parallel with the N teacher models 200” Aslan discloses a joint model training method, in which the first model and the second model can be jointly trained, with the first model and the second model can be of the same, or similar type but the second model may have an optimized architecture.)
Before the effective filing date, it would have been obvious to one of ordinary skilled in the art to combine the teaching of a novel knowledge distillation framework comprises learning of one or more teacher knowledge graph embedding model and the knowledge graph embedding junior student model by Wang, and the teaching of obtaining the embedding vector from the learned knowledge graph embedding model for fine-tuning and clustering to determine clusters correspond to user and their relations and initialization by Costabello, with the teaching of joint model training with the first model and the second model being the same. The motivation to do so is referred to in Aslan’s disclosure (paragraph 9 “Moreover, the techniques and systems described herein improve the technical field of machine learning by providing more flexibility in model training, as compared to current training methods. For example, the techniques and systems described herein allow for “transforming” a machine learning model from one type to another type by training a particular type of machine learning model to mimic another type of machine learning model”, paragraph 26 “For example, the first model 100 can learn from the training data 104, and the training of the second model 102 can be influenced by what the first model 100 is learning from the training data 104 while the first model 100 is being trained, and/or before the first model 100 completes its training. In this sense, the second (student) model 102 can be considered to be learning from the first (teacher) model 100 as the first model 100 learns”, and paragraph 44 “the second model 102 can be trained to mimic the much larger first model 100 (through learning how to approximate the function learned by the first model 100) without significant loss in accuracy of the second model's 102 output. Because the smaller second model 102 take much less memory to maintain and can operate faster on less processing power at runtime, the second model 102 can be a compressed form of the larger first model 100 such that the second model 102 can be more readily deployed on computing devices with limited resources” Aslan discloses the benefit of the joint training model method, in which one or more first (teacher) model and second (student) model can be trained jointly. The first and second model can be of the same type, wherein the second model can be trained to mimic the much larger first model, such that the second model can take much less memory to maintain and can operate faster on less processing power at runtime, and can be more readily deployed on computing devices with limited resources. Furthermore, Wang also discloses a similar framework of one or more teacher model and junior model (student model), which correspond to the first (teacher) model and second (student) model by Aslan, and Wang further discloses at page 4 section 3.2 column 1 “the Junior component can be any existing KGE model”. Therefore, one of ordinary skilled in the art can further incorporate the teaching by Aslan into the teaching by Wang such that the junior knowledge graph embedding model can be a similar type of model that mimic the teacher knowledge graph embedding model, such that the junior model can be further optimized and enhanced based on the benefit as demonstrated by Aslan above.)
Regarding claim 8 depends on claim 1, thus the rejection of claim 1 is incorporated.
Wang/Costabello does not teach the limitation “The learning method of claim 1, wherein the first knowledge graph embedding model and the second knowledge graph embedding model are different kinds of knowledge graph embedding models.”. However, Aslan teaches the limitation (paragraph 20 “two machine learning models of the same, or similar, size can be jointly trained, wherein the two machine learning models differ in terms of their architectures or some other model attribute”, and paragraph 45 “Notwithstanding the utility of the joint training techniques for use in model compression, it is to be appreciated that other applications for the use of joint training are contemplated where, more generally, one type of machine learning model can be “transformed” into another type of machine learning model. For instance, the first model 100 and the second model 102 can differ in their architectures” Aslan discloses a joint model training method, in which the first model and the second model can be jointly trained, with the first model and the second model can differ in terms of their architectures as one type of machine learning model can be “transformed” into another type of machine learning model.)
Before the effective filing date, it would have been obvious to one of ordinary skilled in the art to combine the teaching of a novel knowledge distillation framework comprises learning of one or more teacher knowledge graph embedding model and the knowledge graph embedding junior student model by Wang, and the teaching of obtaining the embedding vector from the learned knowledge graph embedding model for fine-tuning and clustering to determine clusters correspond to user and their relations and initialization by Costabello, with the teaching of joint model training with the first model and the second model being different. The motivation to do so is referred to in Aslan’s disclosure (paragraph 9 “Moreover, the techniques and systems described herein improve the technical field of machine learning by providing more flexibility in model training, as compared to current training methods. For example, the techniques and systems described herein allow for “transforming” a machine learning model from one type to another type by training a particular type of machine learning model to mimic another type of machine learning model”, paragraph 26 “For example, the first model 100 can learn from the training data 104, and the training of the second model 102 can be influenced by what the first model 100 is learning from the training data 104 while the first model 100 is being trained, and/or before the first model 100 completes its training. In this sense, the second (student) model 102 can be considered to be learning from the first (teacher) model 100 as the first model 100 learns”, and paragraph 45 “the first model 100 can comprise a deep neural net (DNN) and the second model 102 can comprise a boosted decision tree—with one having a computational advantage over the other in a given scenario. Perhaps the first DNN model 100 is best suited for accurately learning from the original training data 104, but it is not the type of model that is best to deploy in a particular scenario. Instead, the second model 102 that can be trained in parallel with the first model 100 according to the techniques and systems described herein can be easily deployable and can learn from information passed to it from the first model 100” Aslan discloses the benefit of the joint training model method, in which one or more first (teacher) model and second (student) model can be trained jointly. The first and second model can be different in their architecture as the second model may provide a computational advantage over the first model in a given scenario. Furthermore, Wang also discloses a similar framework of one or more teacher model and junior model (student model), which correspond to the first (teacher) model and second (student) model by Aslan, and Wang further discloses at page 4 section 3.2 column 1 “the Junior component can be any existing KGE model”. Therefore, one of ordinary skilled in the art can further incorporate the teaching by Aslan into the teaching by Wang such that the junior knowledge graph embedding model can have a different architecture to the teacher knowledge graph embedding model, such that the junior model can provide a computational advantage over the teacher model in a given scenario as demonstrated by Aslan above.)
Regarding claim 14 depends on claim 9, thus the rejection of claim 9 is incorporated. Claim 14 is rejected under the same rationale of claim 7 because the claim recites similar limitations and processing steps to claim 7.
Regarding claim 15 depends on claim 9, thus the rejection of claim 9 is incorporated. Claim 15 is rejected under the same rationale of claim 8 because the claim recites similar limitations and processing steps to claim 8.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DUY TU DIEP whose telephone number is (703)756-1738. The examiner can normally be reached M-F 8-4:30.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Alexey Shmatov can be reached at (571) 270-3428. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/DUY T DIEP/Examiner, Art Unit 2123
/ALEXEY SHMATOV/Supervisory Patent Examiner, Art Unit 2123
Prosecution Insights