DETAILED ACTION
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 .
Applicant Reply
“The claims may be amended by canceling particular claims, by presenting new claims, or by rewriting particular claims as indicated in 37 CFR 1.121(c). The requirements of 37 CFR 1.111(b) must be complied with by pointing out the specific distinctions believed to render the claims patentable over the references in presenting arguments in support of new claims and amendments. . . . The prompt development of a clear issue requires that the replies of the applicant meet the objections to and rejections of the claims. Applicant should also specifically point out the support for any amendments made to the disclosure. See MPEP § 2163.06. . . . An amendment which does not comply with the provisions of 37 CFR 1.121(b), (c), (d), and (h) may be held not fully responsive. See MPEP § 714.” MPEP § 714.02. Generic statements or listing of numerous paragraphs do not “specifically point out the support for” claim amendments. “With respect to newly added or amended claims, applicant should show support in the original disclosure for the new or amended claims. See, e.g., Hyatt v. Dudas, 492 F.3d 1365, 1370, n.4, 83 USPQ2d 1373, 1376, n.4 (Fed. Cir. 2007) (citing MPEP § 2163.04 which provides that a ‘simple statement such as ‘applicant has not pointed out where the new (or amended) claim is supported, nor does there appear to be a written description of the claim limitation ‘___’ in the application as filed’ may be sufficient where the claim is a new or amended claim, the support for the limitation is not apparent, and applicant has not pointed out where the limitation is supported.’)” MPEP § 2163(II)(A).
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
The following is a quotation of the first paragraph of pre-AIA 35 U.S.C. 112:
The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor of carrying out his invention.
Claims 1, 3, and 8 rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
All independent claims substantially recite “train the artificial neural network model randomly or sequentially[.]” Original claim 5 provides verbatim support for this limitation, but nothing in original claim 5 indicates how “randomly” or “sequentially” should modify the training of the ANN. “The written description requirement is not necessarily met when the claim language appears in ipsis verbis in the specification. ‘Even if a claim is supported by the specification, the language of the specification, to the extent possible, must describe the claimed invention so that one skilled in the art can recognize what is claimed. The appearance of mere indistinct words in a specification or a claim, even an original claim, does not necessarily satisfy that requirement.’” MPEP § 2163.03. The closest language in the Specification describes the use of latent factors “arbitrarily or sequentially as an initial embedding vector to train the artificial neural network model.” Since the claims recite training an ANN randomly or sequentially but the closest support in the Specification describes something different – using latent factors arbitrary or sequential as initial embedding vectors – the closet language in the Specification also fails to provide any support for the language.
All dependent claims are rejected as including the material of the claims from which they depend.
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1, 3, and 8 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
All independent claims substantially recite “train the artificial neural network model randomly or sequentially[.]” The closest support for this language in the Specification describes the use of latent factors “arbitrarily or sequentially as an initial embedding vector to train the artificial neural network model.” Since the claims recite training an ANN randomly or sequentially but the closest support in the Specification describes something different – using latent factors arbitrary or sequential as initial embedding vectors – there are at least two similarly reasonable interpretations for the claim language. It is not clear which is correct. In the interest of compact prosecution, note that neither interpretation would be clear even without the alternative because there is no objective measure of when an ANN is trained “randomly or sequentially.” Similarly, there is no objective measure that would allow one of ordinary skill to determine when the use of latent factors is arbitrary or sequential.
All dependent claims are rejected as including the material of the claims from which they depend.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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, 3, and 8 are rejected under 35 U.S.C. 103 as being unpatentable over He (Neural Collaborative Filtering, 2017) and Baraniuk (US 2014/0279727)
1. A learning content recommendation system for predicting a probability of a correct answer using an artificial neural network model based on collaborative filtering, the learning content recommendation system comprising: (He teaches: “we present a general framework named NCF, short for Neural network-based Collaborative Filtering.” He Abstract. “The prediction score ^yui then represents how likely i is relevant to u.” He Page 3 Col. 2, continued onto page 3. Note that “u” refers to the user and “i” refers to the item.
He does not teach that the likelihood (probability) the prediction score is correct is part of a learning content recommendation system.
Baraniuk teaches: “In some embodiments, our model represents the probability that a learner provides the correct response to a question in terms of three factors: their understanding of a set of underlying concepts, the concepts involved in each question, and each question's intrinsic difficulty” Baraniuk ¶ 63. “In some embodiments, the key is to integrate textbooks, lectures, and homework assignments into a personalized learning system (PLS) that closes the learning feedback loop by (i) continuously monitoring and analyzing learner interactions with learning resources in order to assess their learning progress and (ii) providing timely remediation, enrichment, or practice based on that analysis.” Baraniuk ¶ 65. “Scheduling: Algorithms that use the results of learning and content analytics to suggest to each learner at each moment what they should be doing in order to maximize their learning outcomes[.]” Baraniuk ¶ 67. “learners may access the server. . . to receive recommendations on further study or questions for further testing. The server may automatically determine the recommendations based on the results of the computational method(s), as variously described herein.” Baraniuk ¶ 69. Baraniuk also teaches use of latent factors. See Baraniuk ¶ 73
It would have been obvious to one of ordinary skill in the art before the effective filing date to combine the teaching of Baraniuk because modifying the system of He to predict correct problems associated with users (instead of preferences associated with users) allows the nonlinear optimization of He to optimize educational content for students to learn more efficiently, faster, and with less human supervision. See e.g. Baraniuk ¶¶ 4-5 and 338. This motivation applies to all citations of Baraniuk.)
a processor; and a memory storing instructions executable by the processor, wherein the processor is configured to collect solution result data for problems solved by a user from a user terminal and store solution result data for problems solved by users (Baraniuk teaches: “In one set of embodiments, a learning system may include a server 110 (e.g., a server controlled by a learning service provider) as shown in FIG. 1.0. The server may be configured to perform any of the various methods described herein. Client computers CC.sub.1, CC.sub.2, . . . , CC.sub.M may access the server via a network 120 (e.g., the Internet or any other computer network). The persons operating the client computers may include learners, instructors, the authors of questions, the authors of educational content, etc. For example, learners may use client computers to access questions from the server and provide answers to the questions. The server may grade the questions automatically based on answers previously provided, e.g., by instructors or the authors of the questions.” Baraniuk ¶ 69. See also Baraniuk ¶673 (“Processing unit 610 . . . stored in memory 612 . . . to implement . . . any combination of the method embodiments described herein.”)
based on performance of the artificial neural network model for predicting the probability of the correct answer, determine a number of latent factors each serving as a basis element for predicting the probability of the correct answer, (Baraniuk teaches: “We initialize W and C with random entries and then iteratively optimize the objective function of (P) for both factors in an alternating fashion.” Baraniuk ¶ 97. “Nevertheless, the use of multiple randomized initialization points can be used to increase the chance of being in the close vicinity of a global optimum, which improves the (empirical) performance of SPARFA-M (see Section I.3.5 for details).” Baraniuk ¶ 115. “The input parameters to SPARFA-M include the number of concepts K and the regularization parameters .gamma. and .lamda.. The number of concepts K is a user-specified value. In practice, cross-validation could be used to select K if the task is to predict missing entries of Y, (see Section I.6.3). The sparsity parameter .lamda. and the l.sub.2-norm penalty parameter .gamma. strongly affect the output of SPARFA-M; they can be selected using any of a number of criteria, including the Bayesian information criterion (BIC) or cross-validation, as detailed in Hastie et al. (2010).” Baraniuk ¶ 121. “Both W and C encode a small number of latent concepts. As we initially noted, the concepts are "abstract" in that they are estimated from the data rather than dictated by a subject matter expert.” Baraniuk ¶ 143. Note that the lowering the number of abstract latent factors also determines the “types” of latent factors eventually learned by the system because at least some will ultimately be different.) obtain the latent factors from the solution result data (He teaches: “We present a neural network architecture to model latent features of users and items and devise a general framework NCF for collaborative filtering based on neural networks.” He Page 2 Col. 2, bullet point 1. “Above the input layer is the embedding layer; it is a fully connected layer that projects the sparse representation to a dense vector. The obtained user (item) embedding can be seen as the latent vector for user (item) in the context of latent factor model. The user embedding and item embedding are then fed into a multi-layer neural architecture, which we term as neural collaborative filtering layers, to map the latent vectors to prediction scores.” He Page 3, Col. 2.) expressed as an MxN matrix including M user rows and N problem columns by using matrix factorization by decomposing the solution result data into user data and problem data thorough the collaborative filtering, (See He P. 2, equation 2. Here the set of vectors pu and the set of vectors qi are mapped to an M x N matrix, where M maps to pu and N maps to Xi. He teaches “MF associates each user and item with a real-valued vector of latent features. Let pu and qi denote the latent vector for user u and item i, respectively; MF estimates an interaction yui as the inner product of pu and qi: [equation 2] where K denotes the dimension of the latent space.” He P. 2. “To explore DNNs for collaborative filtering, we then propose an instantiation of NCF, using a multi-layer perceptron (MLP) to learn the user{item interaction function. Lastly, we present a new neural matrix factorization model, which ensembles MF and MLP under the NCF framework; it unites the strengths of linearity of MF and non-linearity of MLP for modelling the user item latent structures.” He p. 3.)) and train the artificial neural network model randomly or sequentially using the latent factors and determine types of the latent factors based on performance of the trained artificial neural network model for predicting the probability of the correct answer (Baraniuk teaches: “We initialize W and C with random entries and then iteratively optimize the objective function of (P) for both factors in an alternating fashion.” Baraniuk ¶ 97. “Nevertheless, the use of multiple randomized initialization points can be used to increase the chance of being in the close vicinity of a global optimum, which improves the (empirical) performance of SPARFA-M (see Section I.3.5 for details).” Baraniuk ¶ 115. “The input parameters to SPARFA-M include the number of concepts K and the regularization parameters. gamma. and .lamda.. The number of concepts K is a user-specified value. In practice, cross-validation could be used to select K if the task is to predict missing entries of Y, (see Section I.6.3). The sparsity parameter .lamda. and the l.sub.2-norm penalty parameter .gamma. strongly affect the output of SPARFA-M; they can be selected using any of a number of criteria, including the Bayesian information criterion (BIC) or cross-validation, as detailed in Hastie et al. (2010).” Baraniuk ¶ 121. “Both W and C encode a small number of latent concepts. As we initially noted, the concepts are "abstract" in that they are estimated from the data rather than dictated by a subject matter expert.” Baraniuk ¶ 143. Note that the lowering the number of abstract latent factors also determines the “types” of latent factors eventually learned by the system because at least some will ultimately be different.) wherein the latent factors include at least one of problem type, problem difficulty and problem category, and the latent factors include latent factors for the users and latent factors for the problems, (This is obvious over the combination of He and Baraniuk. He teaches: “Each layer of the neural CF layers can be customized to discover certain latent structures of user-item interactions.” He Page 3, Col. 2. (The substitution of latent factors for items preferred by the user with latent factors for the problems is addressed at the beginning of claim 1.)
He does not teach that latent factors include one of problem type, problem difficulty, or problem category.
Baraniuk teaches: “At 4.1.30, the computer system may compute a latent knowledge vector v* for the new learner by estimating a minimum of an objective function with respect to vector argument v, subject to one or more conditions including a norm constraint on the vector argument v. The entries of the latent knowledge vector v* represent the extent of the new learner's knowledge of each of R latent factors (underlying conceptual categories) implicit in the matrix C.” Baraniuk ¶ 658. The claimed factors including problem type and problem category read on the “factors (underlying conceptual categories)” of Baraniuk. Further, the vector taught in Baraniuk is explained as representing the “extent of the new learner’s knowledge,” indicating the vector representing the latent factor also indicates “problem difficulty” for a given learner. The motivation to combine above is applicable to this combination.)
wherein the processor is further configured to (i) perform embedding on the user data of the solution result data for problems solved by the users to generate initial user embedding vectors graspable by the artificial neural network model by using the latent factors for the users, and perform embedding on the problem data of the solution result data to generate initial problem embedding vectors graspable by the artificial neural network model by using the latent factors for the problems, (This reads on creating embedding vectors for the user and for the solution result data. All three steps are taught in the art cited after the third step because the reference teaches all three steps together. For clarity, note that this step is taught in He p. 3 col. 2. (“Above the input layer is the embedding layer; it is a fully connected layer that projects the sparse representation to a dense vector. The obtained user (item) embedding can be seen as the latent vector for user (item) in the context of latent factor model. The user embedding and item embedding are then fed into a multi-layer neural architecture[.]”)) (ii) input the initial user embedding vectors and the initial problem embedding vectors into the artificial neural network model, obtain final user embedding vectors by weight-adjusting the initial user embedding vectors, and obtain final problem embedding vectors by weight-adjusting the initial problem embedding vectors, and (iii) train the artificial neural network model using the final user embedding vectors and the final problem embedding vectors, (This language reads on merely inputting embedding vectors and training, then using the trained values as new embedding vectors. He teaches: “Above the input layer is the embedding layer; it is a fully connected layer that projects the sparse representation to a dense vector. The obtained user (item) embedding can be seen as the latent vector for user (item) in the context of latent factor model. The user embedding and item embedding are then fed into a multi-layer neural architecture, which we term as neural collaborative filtering layers, to map the latent vectors to prediction scores. Each layer of the neural CF layers can be customized to discover certain latent structures of user-item interactions. The dimension of the last hidden layer X determines the model's capability. The final output layer is the predicted score ^yui, and training is performed by minimizing the pointwise loss between ^yui and its target value yui.” He Page 3, Col. 2. “This is the objective function to minimize for the NCF methods, and its optimization can be done by performing stochastic gradient descent (SGD).” He Page 4, Col. 1. “The derivative of the model w.r.t. each model parameter can be calculated with standard back-propagation, which is omitted here due to space limitation.” He Page 5, Col. 2. “Due to the non-convexity of the objective function of NeuMF, gradient-based optimization methods only find locally-optimal solutions. It is reported that the initialization plays an important role for the convergence and performance of deep learning models [7]. Since NeuMF is an ensemble of GMF and MLP, we propose to initialize NeuMF using the pretrained models of GMF and MLP. We first train GMF and MLP with random initializations until convergence. We then use their model parameters as the initialization for the corresponding parts of NeuMF's parameters.” He Page 5, Col. 1-2. “wui is a hyper-parameter denoting the weight of training instance (u; i).” He Page 3 Col. 2. “More precisely, the MLP model under our NCF framework is defined as . . . where Wx, bx, and ax denote the weight matrix, bias vector, and activation function for the x-th layer's perceptron, respectively.” He Page 4, Col. 2. See also Equation 10 on page 4 of He. With respect to training the model using the final embedding vectors, He teaches: “After feeding pre-trained parameters into NeuMF, we optimize it with the vanilla SGD[.]” He P.5 Col.2.)
wherein each of the weight-adjusting of the initial user embedding vectors and the initial problem embedding vectors comprises comparing a predictive value obtained by inputting each of the initial user embedding vectors and the initial problem embedding vectors to the artificial neural network model with an actual value and performing weight-adjustment on each of the initial user embedding vectors and the initial problem embedding vectors in a direction to reduce an error between the predictive value and the actual value, and the weight adjustment is repeatedly performed until the error is less than or equal to a preset value; (Note that this language appears to largely repeat the previous limitations. Specifically, this reads on weight adjusting embedding vectors based on a loss function. Refer to the sections of He cited under the previous limitation. Further, note that He teaches: “Above the input layer is the embedding layer; it is a fully connected layer that projects the sparse representation to a dense vector. The obtained user (item) embedding can be seen as the latent vector for user (item) in the context of latent factor model. The user embedding and item embedding are then fed into a multi-layer neural architecture, which we term as neural collaborative filtering layers, to map the latent vectors to prediction scores. Each layer of the neural CF layers can be customized to discover certain latent structures of user-item interactions. The dimension of the last hidden layer X determines the model's capability. The final output layer is the predicted score ^yui, and training is performed by minimizing the pointwise loss between ^yui and its target value yui.” He Page 3, Col. 2. “This is the objective function to minimize for the NCF methods, and its optimization can be done by performing stochastic gradient descent (SGD).” He Page 4, Col. 1. “The derivative of the model w.r.t. each model parameter can be calculated with standard back-propagation, which is omitted here due to space limitation.” He Page 5, Col. 2. “It is reported that the initialization plays an important role for the convergence and performance of deep learning models [7]. Since NeuMF is an ensemble of GMF and MLP, we propose to initialize NeuMF using the pretrained models of GMF and MLP. We first train GMF and MLP with random initializations until convergence. We then use their model parameters as the initialization for the corresponding parts of NeuMF's parameters.” He Page 5, Col. 1-2. Note that training until convergence implies stopping at a convergence a threshold.) and wherein the processor is further configured to predict a probability of a correct answer of the user for an arbitrary problem through the artificial neural network model trained through the final user embedding vectors and the final problem embedding vectors, (See rejection above, explaining that the prediction of a correct answer of a user based the ANN trained as claimed is obvious over the combination of cited art)
wherein the processor is further configured to adjust the number of the latent factors by (i) determining performances of the artificial neural network model for different numbers of latent factors and (ii) comparing the determined performances of the artificial neural network model, (The closest support for this limitation in found in the Specification is: “The latent factor calculation unit 220 may adjust the number of latent factors in consideration of the performance of the artificial neural network. The number of latent factors may be arbitrarily adjusted with hyperparameters or may be adjusted by finding an optimal value through cross validation. For example, the artificial neural network may have better performance when the number of latent factors is set to N+1 rather than when the number of latent factors is set to N. In this case, the latent factor calculation unit 220 may temporarily set the number of latent factors to N+1, and compare a case having N+1 latent factors with a case N+2 latent factors to find the number of latent factors having the optimal performance.” Spec. P.9, Col.19 - P.10, Col.4.
He teaches using neural networks to express and generalize matrix factorization. See He P.3, Col.2 (“We then show that MF can be expressed and generalized under NCF. To explore DNNs for collaborative filtering, we then propose an instantiation of NCF, using a multi-layer perceptron (MLP) to learn the user{item interaction function. Lastly, we present a new neural matrix factorization model, which ensembles MF and MLP under the NCF framework; it unifies the strengths of linearity of MF and non-linearity of MLP for modelling the user-item latent structures.”)
He does not expressly teach that the NCF framework using neural models determines the number of latent factors by comparing determined performances of the models.
Baraniuk teaches: “The number of latent, abstract concepts K is small relative to both the number of learners N and the number of questions Q. This implies that the questions are redundant and that the learners' graded responses live in a low-dimensional space. The parameter K dictates the concept granularity. Small K extracts just a few general, broad concepts, whereas large K extracts more specific and detailed concepts. Standard techniques like cross-validation (Hastie et al. (2010)) can be used to select K.” Baraniuk ¶ 85. “We use two performance metrics to evaluate the performance of these algorithms, namely (i) the prediction accuracy, which corresponds to the percentage of correctly predicted unobserved responses, and (ii) the average prediction likelihood of the unobserved responses, as proposed in Gonzalez-Brenes and Mostow (2012), for example.” Baraniuk ¶214. “Furthermore, we see from FIG. 1.10 that the prediction performance varies little over different values of K, meaning that the specific choice of K has little influence on the prediction performance within a certain range. This phenomenon agrees with other collaborative filtering results (see, e.g., Koren et al. (2009); Koren and Sill (2011)). Consequently, the choice of K essentially dictates the granularity of the abstract concepts we wish to estimate. We choose K=5 in the real data experiments of Section I.6.2 when we visualize the question-concept associations as bipartite graphs, as it provides a desirable granularity of the estimated concepts in the datasets.” Baraniuk ¶217. See also Baraniuk Fig. 1.10. Note that, even though the value of K had little effect on the model within most of the range, the choice of K=5 was selected in response to the performances of models shown in Figure 1.10, including avoiding the value of 0 for K, which was shown to be suboptimal in Figure 1.10.
It would have been obvious to one of ordinary skill in the art before the effective filing date to combine the teaching of Baraniuk for this limitation because choosing a value for abstract latent concepts K can affect the prediction likelihood and prediction accuracy of the model.
wherein the number of the latent factors is K and the user data is expressed as an MxK matrix in which each row of the MxK matrix indicates a user among the M users and each column of the MxK matrix indicates a latent factor among the K latent factors for the user data, and the problem data is expressed as a KxN matrix in which each row of the KxN matrix indicates a latent factor among the K latent factors for the problem data and each column of the KxN matrix indicates a problem among the N problems, and (See He P. 2, equation 2. Here the set of vectors pu and the set of vectors qi are mapped to an M x N matrix, where M maps to pu and N maps to Xi. He teaches “MF associates each user and item with a real-valued vector of latent features. Let pu and qi denote the latent vector for user u and item i, respectively; MF estimates an interaction yui as the inner product of pu and qi: [equation 2] where K denotes the dimension of the latent space.” He P. 2.) wherein the learning content recommendation system is configured to: (i) perform natural language processing on learning content data to generate a learning content vector, (Baraniuk teaches “Inspired by the recent success of modern text processing algorithms, such as latent Dirichlet allocation (LDA) [3], we posit that the text associated with each question can potentially reveal the meaning of the estimated latent concepts without the need of instructor-provided question tags.” Baraniuk ¶587. See also Baraniuk ¶¶593-594 including equation 2 (“Assume that we observe the word-question occurrence matrix B.epsilon..sup.Q.times.V, where V corresponds to the size of the vocabulary, i.e., the number of unique words that have occurred among the Q questions. Each entry B.sub.i,j represents how many times the .nu..sup.th word occurs in the associated text of the i.sup.th question . . . where t.sub..nu..epsilon..sub.+.sup.K is a non-negative column vector that characterizes the expression of the .nu..sup.th word in every concept.”).) (ii) determine learning content to be recommended to the user based on the learning content vector and the predicted probability of the correct answer of the user, (He teaches “In this work, we strive to develop techniques based on neural networks to tackle the key problem in recommendation – collaborative filtering – on the basis of implicit feedback.” “Popularized by the Netix Prize, MF has become the de facto approach to latent factor model-based recommendation.” He P. 1. Col. 2. “This work addresses the aforementioned research problems by formalizing a neural network modelling approach for collaborative filtering.” He P. 2. Col. 1. “We then show that MF can be expressed and generalized under NCF. To explore DNNs for collaborative filtering, we then propose an instantiation of NCF, using a multi-layer perceptron (MLP) to learn the user-item interaction function. Lastly, we present a new neural matrix factorization model, which ensembles MF and MLP under the NCF framework; it unifies the strengths of linearity of MF and non-linearity of MLP for modelling the user-item latent structures.” He P. 3 col. 1. “We now formulate the NCF's predictive model as [equation 3] where where P E RMxK and Q E RNxK, denoting the latent factor matrix for users and items, respectively; and theta_f denotes the model parameters of the interaction function f.” He P. 3 col. 2.) and (iii) provide the determined learning content to the user. (“In this patent we disclose, among other things, (a) a new model and algorithms for machine learning-based learning analytics, which estimate a learner's knowledge of the concepts underlying a domain, and (b) content analytics, which estimate the relationships among a collection of questions and those concepts. In some embodiments, our model represents the probability that a learner provides the correct response to a question in terms of three factors: their understanding of a set of underlying concepts, the concepts involved in each question, and each question's intrinsic difficulty.” Baraniuk ¶63. “Furthermore, learners may access the server to determine (e.g., view) their estimated concept-knowledge values for the concepts that have an extracted by the computational method(s), and/or, to view a graphical depiction of question-concept relationships determined by the computational method(s), and/or, to receive recommendations on further study or questions for further testing.” Baraniuk ¶69.)
3. The learning content recommendation system of claim 1, wherein the processor is further configured to, upon identifying that a learning efficiency of the user has improved in response to the artificial neural network model being trained with a specific latent factor assigned a weight, weight the specific latent factor to generate an initial embedding vector for the user. (With respect to the language “for the user,” intended use language is explained in MPEP §§ 2103 and 2111.02. “Claim scope is not limited by claim language that suggests or makes optional but does not require steps to be performed, or by claim language that does not limit a claim to a particular structure.” MPEP § 2111.04. See rejection of claim 1. See also Section 3.1.1. of He explaining regression with squared loss as comparing observed interaction with a prediction score (i.e. comparing y with ^y as y is used in that section.))
Claim 8 is rejected for reasons given in the rejection of claim 1.
Response to Arguments
Applicant's arguments filed 02/10/2026 have been fully considered but they are not persuasive.
Rejections under § 101
All rejections under this section are withdrawn. As indicated by Applicant, the invention is directed to a technique for more efficiently finding the global minima of the problem space using machine learning.
Rejections under § 103
Applicant states that the references fail to teach “determining the number of latent factors based on performance of the AI model” and “determining types of latent factors based on performance of the trained AI model.” Notably absent is any specific argument distinguishing the sections of the prior art cited as teaching these aspects of the claimed invention in the previous action. See rejection above.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to PAUL M KNIGHT whose telephone number is (571) 272-8646. The examiner can normally be reached Monday - Friday 9-5 ET.
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, Michelle Bechtold can be reached on (571) 431-0762. 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.
PAUL M. KNIGHTExaminerArt Unit 2148
/PAUL M KNIGHT/Examiner, Art Unit 2148