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 .
DETAILED ACTION
The instant application having Application No. 17319579 has a total of 20 claims pending in the application, of which claims 5, 13, and 19 have been cancelled.
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 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1: Claims 1 and 9 are process type claims. Claim 15 is a machine type claim. Therefore, claims 1-20 are directed to either a process, machine, manufacture or composition of matter.
As per claim 1,
2A Prong 1:
“A data set comprising features of a plurality of users associated with content items of the plurality of content items that correspond to the plurality of users, wherein each respective user of the plurality of users is associated with a respective first set of content items” A salesman mentally or with pencil and paper matches the products used by his customers (i.e. plurality of users)).
“to output corresponding vector representations of relevant content items for users based on features of the users and corresponding content identifiers for the relevant content items…” The salesman mentally or with pencil and paper determines new products his customers might like based on their previous purchases.
“Wherein the training data set comprises, for each respective user of the plurality of users, the features of the respective user associated with respective labels indicating that:” The salesman keeps track of features of the products the customer bought and identifies items with similar features that could be relevant to future recommendations.
“the first set of content items correspond to the respective user;” The salesman mentally or with pencil and paper keeps track of content items associated with each of his customers.
“and a second set of content items, having embedding representations that are within a threshold distance of embedding representations of the first set of conte items, correspond to the respective user” The salesman looks for items with similar features to include when considering what to recommend to their customers.
2A Prong 2: This judicial exception is not integrated into a practical application.
Additional elements:
“training a machine learning model”, “Training the machine learning model, using a training data set”, “wherein training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings and a second component corresponding to content identifiers, wherein the first component is parameterized to have a larger impact on loss than the second component” (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: Machine learning model is claimed generically with no additional limitations or details beyond a generic, off the shelf machine learning model.
“Providing features of a plurality of content items as inputs to an embedding model”, “receiving a dataset” (Adding insignificant extra-solution activity to the judicial exception - see MPEP 2106.05(g)).
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
“training a machine learning model”, “Training the machine learning model, using a training data set”, “wherein training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings and a second component corresponding to content identifiers, wherein the first component is parameterized to have a larger impact on loss than the second component” (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: Machine learning model is claimed generically with no additional limitations or details beyond a generic, off the shelf machine learning model.
“Providing features of a plurality of content items as inputs to an embedding model”, “receiving a dataset” (MPEP 2106.05(d)(II) indicate that merely “Receiving or transmitting data” is a well‐understood, routine, conventional function when it is claimed in a merely generic manner (as it is in the present claim). Thereby, a conclusion that the claimed providing and receiving steps are well-understood, routine, conventional activity is supported under Berkheimer).
As per claim 2, this claim contains additional generic machine learning aspects similar to claim 1, and is rejected for similar reasons.
As per claim 3, this claim contains similar mental steps to claim 1, and is rejected similarly to claim 1.
As per claim 4, and 6-8, this claim contains similar mental steps and generic machine learning to claim 1, and is rejected for similar reasons to claim 1.
As per claim 9,
2A Prong 1:
“to output embeddings of content items based on user features and corresponding content identifiers for the relevant content items” A salesman mentally or with pencil and paper determines new products his customers might like based on their previous purchases.
“A training data set that was generated based on identifying, by comparing a similarity of embeddings of a first set of content items associated with a particular user of a plurality of users to embeddings of other content items, a second set of conte items that correspond to the particular user” The salesman mentally or with pencil and paper keeps track of features of the products the customer bought or interacted with along with similar items to those they’ve enjoyed and uses them to recommend other products to the customer.
“receiving … in response to the inputs, an output indicating one or more embedding of content items” The salesman mentally or with pencil and paper decides on products to recommend for his customers based on the data they’ve used to learn their customers interests.
2A Prong 2: This judicial exception is not integrated into a practical application.
Additional elements:
“to a machine learning model that has been trained”, “wherein the training involves a training data set”, “wherein training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings and a second component corresponding to content identifiers, wherein the first component is parameterized to have a larger impact on loss than the second component” (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: Machine learning model is claimed generically with no additional limitations or details beyond a generic, off the shelf machine learning model.
“Providing a plurality of features of a user as inputs” (Adding insignificant extra-solution activity to the judicial exception - see MPEP 2106.05(g)).
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
“to a machine learning model that has been trained”, “wherein the training involves a training data set”, “wherein training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings and a second component corresponding to content identifiers, wherein the first component is parameterized to have a larger impact on loss than the second component” (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: Machine learning model is claimed generically with no additional limitations or details beyond a generic, off the shelf machine learning model.
“Providing a plurality of features of a user as inputs” (MPEP 2106.05(d)(II) indicate that merely “Receiving or transmitting data” is a well‐understood, routine, conventional function when it is claimed in a merely generic manner (as it is in the present claim). Thereby, a conclusion that the claimed providing step is well-understood, routine, conventional activity is supported under Berkheimer).
As per claim 10, this claim contains additional generic receiving and transmitting of data to claim 9, and is rejected for similar reasons.
As per claim 11, this claim contains additional receiving, transmitting, and mental steps similar to claim 9, and is rejected for similar reasons.
As per claim 12 and 14, these claims contain similar mental steps to claim 9, and are rejected for similar reasons.
As per claim 15,
2A Prong 1:
“Receiving embeddings of the plurality of content items as outputs from the embedding model based on the inputs” A salesman mentally or with pencil and paper recommends products to his customers based upon a mental model he uses for recommendations.
“A data set comprising features of a plurality of users associated with content items of the plurality of content items that correspond to the plurality of users, wherein each respective user of the plurality of users is associated with a respective first set of content items” A salesman mentally or with pencil and paper matches the products used by his customers (i.e. plurality of users)).
“to output corresponding vector representations of relevant content items for users based on features of the users and corresponding content identifiers for the relevant content items” The salesman mentally or with pencil and paper determines new products his customers might like based on their previous purchases.
“Wherein the training data set comprises, for each respective user of the plurality of users, the features of the respective user associated with respective labels indicating that:” The salesman keeps track of features of the products the customer bought or interacted with.
“the first set of content items correspond to the respective user;” The salesman mentally or with pencil and paper keeps track of content items associated with each of his customers.
“and a second set of content items, having embedding representations that are within a threshold distance of embedding representations of the first set of conte items, correspond to the respective user” The salesman looks for items with similar features to the products their customer enjoyed and consider them as well when considering products to recommend to their customers in the future.
2A Prong 2: This judicial exception is not integrated into a practical application.
Additional elements:
“one or more processors”, “a memory” (mere instructions to apply the exception using a generic computer component);
“training a machine learning model”, “Train the machine learning model, using a training data set”, “training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings and a second component corresponding to content identifiers, wherein the first component is parameterized to have a larger impact on loss than the second component” (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: Machine learning model is claimed generically with no additional limitations or details beyond a generic, off the shelf machine learning model.
“Provide features of a plurality of content items as inputs to an embedding model”, “receive a dataset” Adding insignificant extra-solution activity to the judicial exception - see MPEP 2106.05(g)).
2B: The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Additional elements:
“one or more processors”, “a memory” (mere instructions to apply the exception using a generic computer component)
“training a machine learning model”, “Train the machine learning model, using a training data set”, “training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings and a second component corresponding to content identifiers, wherein the first component is parameterized to have a larger impact on loss than the second component” (Adding the words “apply it” (or an equivalent) with the judicial exception, or mere instructions to implement an abstract idea on a computer, or merely uses a computer as a tool to perform an abstract idea - see MPEP 2106.05(f) – Examiner’s note: Machine learning model is claimed generically with no additional limitations or details beyond a generic, off the shelf machine learning model.
“Provide features of a plurality of content items as inputs to an embedding model”, “receive a dataset” (MPEP 2106.05(d)(II) indicate that merely “Receiving or transmitting data” is a well‐understood, routine, conventional function when it is claimed in a merely generic manner (as it is in the present claim). Thereby, a conclusion that the claimed providing and receiving steps are well-understood, routine, conventional activity is supported under Berkheimer).
As per claim 16, 18 and 20, this claim contains similar mental steps and generic machine learning models to claim 15, and is rejected for similar reasons.
As per claim 17, this claim contains similar mental steps to claim 15, and is rejected for similar reasons.
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 (i.e., changing from AIA to pre-AIA ) 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 1-5, 9-11, 13-15, and 17-19 are rejected under 35 U.S.C. 103 as being unpatentable over Hsieh et al (US 20210174164 A1) in view of Piroonsup et al (“Analysis of training data using clustering to improve semi-supervised self-training”).
As per claims 1 and 15, Hsieh Discloses, “A method for training a machine learning model, comprising” (pg.9, particularly paragraph 0108; EN: this denotes training a neural network).
“providing features of a plurality of content items as inputs to an embedding model” (Pg.4, particularly paragraph 0051; EN: this denotes taking in input content into the Content neural network and producing important concepts/keywords. The neural network is the embedding model).
“receive embeddings of the plurality of content items as outputs from the embedding model based on the inputs” (Pg.4, particularly paragraph 0051; EN: this denotes taking in input content into the Content neural network and producing important concepts/keywords. The keywords/concepts are the output embeddings).
“receiving a data set comprising features of a plurality of users associated with content items of the plurality of content items that correspond to the plurality of users wherein each respective user of the plurality of users is associated with a respective first set of content items” (pg.4, particularly paragraph 0057-0059; EN: this denotes user profiles with associated history of content for each user).
“training the machine learning model, using a training data set, to output corresponding vector representations of relevant content items for users based on features of the users and corresponding content identifiers for the relevant content items” (Pg.9, particularly paragraph 0108; EN: this denotes training the based upon a set of users with full interaction data to provide recommendations to the users). “Wherein training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings ana second component corresponding to content identifiers” (pg.9-10, particularly paragraph 0109 and corresponding equations; EN: this denotes a loss function containing both item and user aspects). “wherein the first component is parameterized to have a larger impact on the loss than the second component” (pg.9-10, particularly paragraph 0109 and corresponding equations; EN: this denotes a loss function containing both item and user aspects, the user aspects have a larger impact because they are the deciding factor in relation to the items, as their preferences are what defines the distances to the items being compared to them).
“Wherein the training data set comprises, for each respective user of the plurality of users the features of the respective user” (pg.4, particularly paragraph 0057-0059; EN: this denotes user profiles with features of the user). “associated with respective labels indicating that” (Pg.4, particularly paragraph 0058; EN: this denotes classifications associated with the features).
“the first set of content items correspond to the respective user” (pg.4, particularly paragraph 0057-0059; EN: this denotes user profiles with associated history of content for each user).
However, Hsieh fails to explicitly disclose, “a second set of content items, having embedding representations that are within a threshold distance of embedding represents of the first set of content items, correspond to the respective user.”
Piroonsup discloses, “a second set of content items, having embedding representations that are within a threshold distance of embedding represents of the first set of content items, correspond to the respective user.” (Pg.68, particularly section 4.2; EN: this denotes using clustering to bring in additional training data within a reasonable distance of known training data. Here the threshold is being within the cluster of labeled data).
Hsieh and Piroonsup are analogous art because both involve machine learning.
Before the effective filing date it would have been obvious to one skilled in the art of machine learning to combine the work of Hsieh and Piroonsup in order to determine similar data to use for training in relation to known training data.
The motivation for doing so would be to “apply unlabeled data long with labeled data to train a system” (Piroonsup, Pg.65, section 1) or in the case of Hsieh, allow the system to include similar items to known items when training the prediction system of Hsieh.
Therefore before the effective filing date it would have been obvious to one skilled in the art of machine learning to combine the work of Hsieh and Piroonsup in order to determine similar data to use for training in relation to known training data.
With respect to claim 2, Hsieh discusses wherein training the machine learning model is based on a custom loss function (Para. [0082] of Hsieh discusses a loss function [0109]) that relates to mean squared error and categorical cross-entropy. (Para. [0082] of Hsieh discusses a loss function [0109])
With respect to claim 3, Hsieh discusses wherein determining which respective embeddings of the embeddings correspond to each respective user of the plurality of users based on the data set and similarities among the embeddings comprises identifying at least one embedding of the embeddings that corresponds to a content item that is not included in the data set that corresponds to a given user of the plurality of users based on the similarities among the embeddings. (Para. [0082] of Hsieh discusses a loss function [0041] discusses how “a new movie added to a movie database may not have any existing viewing history but could be incorporated into personalized recommendations for appropriate users through the conceptual mapping of content as used in the system and method”)
With respect to claim 4, Hsieh discusses wherein:
the training data set further comprises a respective content identifier associated with each respective embedding of the embeddings; and (Para. [0100] of Hsieh discusses “Properties of the content data submitted through an API can include content identifier”)
As per claim 9, Hsieh discloses, “A method for recommending content, comprising” (Pg.5, particularly paragraph 0068; EN: this denotes using a NN to provide recommendations to the users).
“Providing a plurality of features of a user as inputs to a machine learning model that has been trained to output embeddings of content items based on user features and corresponding content identifiers for the relevant content items” (Pg.5, particularly paragraph 0068; EN: this denotes training the MmNN based upon a set of users with full interaction data to provide recommendations to the users; Pg.9, particularly paragraph 0108; EN: this denotes the embeddings being both user and item based aspects). “Wherein training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings and a second component corresponding to content identifiers” (pg.9-10, particularly paragraph 0109 and corresponding equations; EN: this denotes a loss function containing both item and user aspects). “Wherein the first component is parameterized to have a larger impact on loss than the second component” (pg.9-10, particularly paragraph 0109 and corresponding equations; EN: this denotes a loss function containing both item and user aspects, the user aspects have a larger impact because they are the deciding factor in relation to the items, as their preferences are what defines the distances to the items being compared to them).
“Wherein the training involves a training data set that was generated…” (Pg.9, particularly paragraph 0108; EN: this denotes training the based upon a set of users with full interaction data to provide recommendations to the users). “… embeddings of a first set of content items associated with a particular user of a plurality of users…” (pg.4, particularly paragraph 0057-0059; EN: this denotes user profiles with associated history of content for each user).
“Receiving, form the machine learning model in response to the inputs, an output indicating one or more content items” (Pg.5, particularly paragraph 0068; EN: this denotes training the MmNN based upon a set of users with full interaction data to provide recommendations to the users).
However, Hsieh fails to explicitly disclose, “by comparing a similarity of embeddings of a first set of content items associated with a particular user of a plurality of users to embeddings of other content items, a second set of content items that correspond to the particular user”
Piroonsup discloses, “by comparing a similarity of embeddings of a first set of content items associated with a particular user of a plurality of users to embeddings of other content items, a second set of content items that correspond to the particular user” (Pg.68, particularly section 4.2; EN: this denotes using clustering to bring in additional training data within a reasonable distance of known training data. Here the threshold is being within the cluster of labeled data).
Hsieh and Piroonsup are analogous art because both involve machine learning.
Before the effective filing date it would have been obvious to one skilled in the art of machine learning to combine the work of Hsieh and Piroonsup in order to determine similar data to use for training in relation to known training data.
The motivation for doing so would be to “apply unlabeled data long with labeled data to train a system” (Piroonsup, Pg.65, section 1) or in the case of Hsieh, allow the system to include similar items to known items when training the prediction system of Hsieh.
Therefore before the effective filing date it would have been obvious to one skilled in the art of machine learning to combine the work of Hsieh and Piroonsup in order to determine similar data to use for training in relation to known training data.
With respect to claim 10, Hsieh discusses recommending, via a user interface, the one or more content items to the user. (Para. [0203] of Hsieh discusses “the resulting content items are served to a user interface form where the user interface is updated and changed to display the personalized list of autocomplete content items”)
With respect to claim 11, Hsieh discusses:
receiving data related to a new content item; (Fig. 4 of Hsieh shows the C-NN 120 outputting content embeddings, based on the content data from the content repository 110, into the matchmaker neural network (MmNN) 140)
determining an embedding of the new content item using an embedding model, based on the data related to the new content item; and (Para. [0079] of Hsieh discusses “processing the user embedding through a matchmaking neural network, which is a trained model to map user embeddings and content embeddings to a shared dimensional space, and yielding a user shared-item embedding”)
determining whether to recommend the new content item to the user based on a similarity between the embedding of the new content item and the one or more embeddings of content items indicated in the output from the machine learning model. (Para. [0079] of Hsieh discusses “processing the user embedding through a matchmaking neural network, which is a trained model to map user embeddings and content embeddings to a shared dimensional space, and yielding a user shared-item embedding”. Additionally, para. [0129]of Hsieh discusses “applying analysis of a shared-item embedding in the matchmaking data model S140 may be used for delivering search results, generating content recommendations, generating item pairing recommendations, powering form autocompletion, and generating a report as a partial list of applications”.)
With respect to claim 14, Hsieh discusses wherein determining the plurality of features of the user comprises determining clickstream data and one or more attributes of the user. (Paras. [0058]-[0060] discuss, user profile being processed with machine learning, which includes user attributes, and user interaction data that includes a set of recent interactions including clicking content (i.e. clickstream data). Additionally, Para. [0062] of Hsieh discusses how “the U-NN 130 may operate with the available user related data of a site. In some implementations, this can include creating a user profile for any anonymous or visitor session on a given website or application. As a result, personalization can be customized around the data and interactions of a particular session or visit”. The Examiner notes that this is similar to the data is the information collected about a user while they browse through a website or use a web browser (e.g., clickstream data).)
With respect to claims 17-18, claims 17-18 include similar subject matter as provided in claims 3-5 and vary only in the preamble (e.g., system, method, computer readable medium, etc.). Therefore, claims 17-19 are rejected using the same rationale as applied to claims 3-5.
Claim Rejections - 35 USC § 103
Claims 6-8, 16, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Hsieh et al (US 20210174164 A1) in view of Piroonsup et al (“Analysis of training data using clustering to improve semi-supervised self-training”) and further in view of U.S. Patent No. 10,986,408 to Randhawa et al. (hereinafter “Randhawa”).
With respect to claim 6, Hsieh discusses wherein training the machine learning model comprises:
providing the features of the plurality of users in the training data set as inputs to the machine learning model; (Fig. 1 of Hsieh shows user profile information as input data which is used to train the matchmaking data model)
comparing outputs received from the machine learning model in response to the inputs to the respective labels in the training data set; and (Para. [0098] of Hsieh discusses “Training a set of neural networks S110, when used for user and content comparisons”)
The combination of Hsieh and Bala fails to explicitly discuss, iteratively adjusting one or more parameters of the machine learning model based on the comparing.
However, in analogous art, Randhawa discloses iteratively adjusting one or more parameters of the machine learning model based on the comparing at column 13, Lines 20-39 of Randhawa, which discusses how “content recommendation platform 102 may be configured to retrain the one or more data models based on the new content viewership data and/or the user feedback data.”
It would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to use the content data association method of Randhawa to modify the content recommendation engine of Hsieh in order to improve user satisfaction by leveraging feedback to enable the systems to know what recommendations were correct or incorrect, while enabling the system to iteratively adjust parameters to perform better in the future as more data is available.
With respect to claim 7, Hsieh discusses:
receiving feedback data indicating whether a content item recommended based on an output from the machine learning model was relevant to a user; (Para. [0049] of Hsieh discusses how “C-NN 120 preferably includes functionality to update and improve item-item similarity through feedback from the system and from outside the system”)
The combination of Hsieh and Bala fails to explicitly discuss:
adding, removing, or modifying one or more training data instances in the training data set based on the feedback data to produce an updated training data set; and
re-training the machine learning model based on the updated training data set.
However, in analogous art, Randhawa discloses adding, removing, or modifying one or more training data instances in the training data set based on the feedback data to produce an updated training data set and re-training the machine learning model based on the updated training data set at column 13, Lines 20-39 of Randhawa, which discusses how “content recommendation platform 102 may be configured to retrain the one or more data models based on the new content viewership data and/or the user feedback data.”
It would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to use the content data association method of Randhawa to modify the content recommendation engine of Hsieh in order to improve user satisfaction by leveraging feedback to enable the systems to know what recommendations were correct or incorrect, while enabling the system to iteratively adjust parameters to perform better in the future as more data is available.
With respect to claim 8, Hsieh discusses:
receiving data related to a new content item; (Fig. 4 of Hsieh shows the C-NN 120 outputting content embeddings, based on the content data from the content repository 110, into the matchmaker neural network (MmNN) 140)
determining an embedding of the new content item using the embedding model, based on the data related to the new content item; (Para. [0079] of Hsieh discusses “processing the user embedding through a matchmaking neural network, which is a trained model to map user embeddings and content embeddings to a shared dimensional space, and yielding a user shared-item embedding”)
determining that the embedding of the new content item is similar to a given embedding in the training data set that corresponds to a given set of user features; (Para. [0079] of Hsieh discusses “processing the user embedding through a matchmaking neural network, which is a trained model to map user embeddings and content embeddings to a shared dimensional space, and yielding a user shared-item embedding”)
adding, to the training data set, an association between the embedding of the new content item and the given set of user features to produce an updated training data set; and (Para. [0118] of Hsieh discusses “the method can include adding content to a data model”)
The combination of Hsieh and Bala fails to explicitly discuss re-training the machine learning model based on the updated training data set.
However, in analogous art, Randhawa discloses re-training the machine learning model based on the updated training data set at column 13, Lines 20-39 of Randhawa, which discusses how “content recommendation platform 102 may be configured to retrain the one or more data models based on the new content viewership data and/or the user feedback data.”
It would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to use the content data association method of Randhawa to modify the content recommendation engine of Hsieh in order to improve user satisfaction by leveraging feedback to enable the systems to know what recommendations were correct or incorrect, while enabling the system to iteratively adjust parameters to perform better in the future as more data is available.
With respect to claim 16, claim 16 includes similar subject matter as provided in claim 8 and varies only in the preamble (e.g., system, method, computer readable medium, etc.). Therefore, claim 16 is rejected using the same rationale as applied to claim 8.
With respect to claim 20, claim 20 includes similar subject matter as provided in claim 6 and varies only in the preamble (e.g., system, method, computer readable medium, etc.). Therefore, claim 20 is rejected using the same rationale as applied to claim 6.
Claim Rejections - 35 USC § 103
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over Hsieh et al (US 20210174164 A1) in view of Piroonsup et al (“Analysis of training data using clustering to improve semi-supervised self-training”) and further in view of Wang et al (US 20220405607 A1).
With respect to claim 12, Hsieh fails to explicitly discuss wherein the similarity between the embedding of the new content item and the one or more embeddings of the content items indicated in the output from the machine learning model is determined using cosine similarity.
However, in analogous art, Wang discloses the similarity between the embedding of the new content item and the one or more embeddings of the content items indicated in the output from the machine learning model is determined using cosine similarity at para. [0110] of Wang, which discusses “the similarity between the user feature vector of the target user and the tag feature vector of the content tag may be a dot product value, a Euclidean distance, and a cosine similarity between the user feature vector and the tag feature vector.”
It would have been obvious to one having ordinary skill in the art before the effective filing date of the claimed invention to use the user profile analysis method of Wang to modify the content recommendation engine of Hsieh in order to improve the precision of content recommendation by associating the attribute information and the historical behavior data of the user across different applications with content tags of multiple content. This may enable data gathered across multiple different application platforms to be used by a target application in a meaningful way.
Response to Arguments
In pg.9-10, Applicant argues in regards to the rejection under U.S.C. 101,
Here, the claims are directed to a technical solution to problems arising in the field of using machine learning to determine content to provide to a user. For example, the Application explains:
While some existing techniques involve determining relevant content to provide to users based on past user interactions with content, these techniques are not easily adapted to new content that has not yet been interacted with by users. Furthermore, these techniques may require significant amounts of historical data to produce accurate results, and may involve significant processing resources for analyzing the historical data. As such, there is a need in the art for improved techniques of determining relevant content to provide to users. Application as Filed, I [0003] (emphasis added).
The claimed methods and systems provide a solution to these problems. For example, "by utilizing embeddings of content items instead of merely identifiers of content items to train a model for content prediction, techniques described herein allow for more dynamic and flexible determinations of relevant content items that is based on meaning associated with content items instead of being based only on rigid past associations between users and particular content items." Id., I [0029]. Additionally, "embodiments of the present disclosure allow for the identification of new content items that may be relevant to a user, even if the new content items were not included in the training data for the model, based on similarities between embeddings." Id. Thus, accurate recommendations can be generated even for content that has not been interacted with by users (e.g., content that is not associated with any user). Furthermore, "by training a content prediction model based on embeddings determined using an embedding model, the knowledge of the embedding model is transferred into the content prediction model. Thus, while the embedding model may be large and require significant amounts of processing and storage resources, a smaller and more efficient content prediction model may be able to leverage the transferred knowledge of the embedding model in a more efficient model architecture, which allows for the content prediction model to use less energy and consume less space, and thus run on a wider range of devices. Accordingly, embodiments of the present disclosure provide improved machine learning techniques, and allow for improved content recommendation." Id., I [0031] (emphasis added).
In response the Examiner maintains the rejection as shown above. Applicant’s argument appears to be that they have improved the technology of machine learning by improving content recommendation via machine learning. However, using generic machine learning models and computer equipment to perform an abstract idea does not improve the underlying hardware or machine learning model. Merely making the abstract idea improved or more efficient is an improvement to the abstract idea, not the machine learning model, and therefore the rejection is maintained as shown above.
In pg.11, Applicant further argues in regards to the rejection under U.S.C. 101,
Applicant submits that the claims as currently presented do not recite a mental process. Example 39 of the subject matter eligibility examples establishes that generating a training data set and training a machine learning model are not abstract ideas. Examples, pp. 8-9. In Example 39, the claim recites "training the neural network in a first stage using the first training set" and "training the neural network in a second stage using the second training set." Examples, pp. 8-9. Also, Example 39 recites "collecting a set of digital facial images from a database (i.e., receiving data) and creating first and second training data sets. Id. Notably, Example 39 states that:
the claim does not recite any mathematical relationships, formulas, or calculations. While some of the limitations may be based on mathematical concepts, the mathematical concepts are not recited in the claims. Further, the claim does not recite a mental process because the steps are not practically performed in the human mind. Finally, the claim does not recite any method of organizing human activity such as a fundamental economic concept or managing interactions between people. Thus, the claim is eligible because it does not recite a judicial exception. 2019 Subject Matter Eligibility Examples, p. 9 (emphasis added).
In response, the Examiner maintains the rejection as shown above. Example 39 does not apply to these claims, as Example 39 was found to not have an abstract idea, and therefore be statutory under U.S.C. 101. However, as the current claims have a clearly indicated abstract idea of recommending content items to a user, Example 39 does not apply to the current claims. Therefore the rejection is maintained as shown above.
In pg.13, the Applicant further argues in rejection to the claims under U.S.C. 101,
Applicant's claims as currently amended integrate any alleged abstract idea into a practical application by improving machine learning technologies. In particular, Applicant's claimed solution allows for "the identification of new content items that may be relevant to a user, even if the new content items were not included in the training data for the model, based on similarities between embeddings." Application as Filed, I [0029]. Thus, accurate recommendations can be generated even for content that has not been interacted with by users (e.g., content that is not associated with any user). Furthermore, "by training a content prediction model based on embeddings determined using an embedding model, the knowledge of the embedding model is transferred into the content prediction model. Thus, while the embedding model may be large and require significant amounts of processing and storage resources, a smaller and more efficient content prediction model may be able to leverage the transferred knowledge of the embedding model in a more efficient model architecture, which allows for the content prediction model to use less energy and consume less space, and thus run on a wider range of devices. Accordingly, embodiments of the present disclosure provide improved machine learning techniques, and allow for improved content recommendation." Id., I [0031].
In response, the Examiner maintains the rejection as shown above. Applicant once again describes improvements to the abstract idea, not the machine learning model. The use of a generic machine learning model to perform an abstract idea is not an improvement to the machine learning model. Requiring less processing or using less data is not an improvement to the model itself, its an improvement to the abstract idea. Therefore the rejection is maintained as shown above.
In pg.13-14, the Applicant further argues to the rejection under U.S.C. 101,
Applicant next turns the Examiner's attention to a recent rehearing decision (Ex parte Desjardins, Appeal No. 2024-000567 (Decision on Request for Rehearing) (September 26, 2025)) written by Director Squires. The claims at issue in Ex parte Desjardins relate to training a machine learning model in a way that avoids the problem of catastrophic forgetting, where subsequent training steps override the training performed in previous training steps. The claims in Ex parte Desjardins were held to recite an abstract idea (specifically, at least computing a probability distribution). However, Director Squires vacated the 101 rejection, reasoning that the claim features "constitute[] an improvement to how the machine learning model itself operates, and not, for example, the identified mathematical calculation." Similar to the claims in Ex parte Desjardins, Applicant's claims recite an improvement to the functioning of machine learning models, as discussed above (e.g., by transferring the knowledge of the embedding model to the recommendation model and enabling the use of content items that have yet to be interacted with as training data).
Furthermore, in Ex parte Desjardins, the rehearing panel stated that the reasoning of the PTAB in introducing the 101 rejection was "overbroad." Desjardins, p. 9. For example, the PTAB reasoned that the machine learning model is "an unpatentable 'algorithm' and the remaining additional elements [are] 'generic computer components,' without adequate explanation." Desjardins, p. 9. Similar to the PTAB's analysis in Desjardins, the Office Action in the present case rejects the claims by dismissing the limitations as either an abstract idea or mere instructions to apply an exception. Office Action, pp. 5-6.
In response, the Examiner maintains the rejection as shown above. Desjardins disclosed an improvement to the machine learning model – solving the problem of catastrophic forgetting for the machine learning model. Merely improving the recommendation of content to a user is not an improvement to the machine learning model, it is an improvement to the abstract idea of recommending content to a user. Therefore the rejection is maintained as shown above.
In pg.14-15, the Applicant further argues in regards to the rejection under U.S.C. 101,
During the interview of December 15, 2025, the Examiner suggested that the claims of the present case may be distinguishable from the claims of Desjardins in that the claims of the present case improve "the abstract idea of content recommendation" as opposed to the functioning of a machine learning model. In response to this line of reasoning, Applicant submits that the end use case for all machine learning models - including the machine learning model in Desjardins - could be interpreted as being an abstract idea under similar reasoning. For example, the specification of Desjardins states that the machine learning model can be used for "classification tasks, such as image processing tasks, speech recognition tasks, natural language processing tasks, or optical character recognition tasks." Desjardins specification, paragraph [28]. Classification, speech recognition, natural language processing, character recognition, etc. are all tasks that could arguably be performed in the human mind. However, rejecting the claims of Desjardins under Section 101 because the model of Desjardins would be used to perform a mental process would be exactly the type of overbroad reasoning that the Desjardins decision warned would render AI innovations unpatentable.
In response, the Examiner maintains the rejection as shown above. Applicant appears to be arguing that because the discussion of Desjardins stated that the model used in that case could be used for many things, that this means that their claims should be statutory under U.S.C. 101. As stated above, Desjardins had an improvement to the issue of catastrophic forgetting for machine learning models. The improvement was, therefore, for the machine learning model. By improving the machine learning model, that machine learning model could potentially be used for many things, including potential abstract ideas! However, the instant claims do not provide an improvement to the machine learning model, it describes improvements to the abstract idea of providing content recommendation. Since there is no similar improvement to the machine learning model as seen in Desjardins, the rejection is maintained as shown above.
In pg. 16-17, the Applicant further argues in regards to the rejection under U.S.C. 101,
The Specification here provides such sufficient details to support that the claimed features improve the technology of machine learning-based content recommendation. See, e.g., Application as Filed III [0029]-[0031]. For example, the Specification explains how "by utilizing embeddings of content items instead of merely identifiers of content items to train a model for content prediction, techniques described herein allow for more dynamic and flexible determinations of relevant content items that is based on meaning associated with content items instead of being based only on rigid past associations between users and particular content items." Application as Filed, I [0029]. Additionally, "embodiments of the present disclosure allow for the identification of new content items that may be relevant to a user, even if the new content items were not included in the training data for the model, based on similarities between embeddings." Id. Thus, accurate recommendations can be generated even for content that has not been interacted with by users (e.g., content that is not associated with any user). Furthermore, "by training a content prediction model based on embeddings determined using an embedding model, the knowledge of the embedding model is transferred into the content prediction model. Thus, while the embedding model may be large and require significant amounts of processing and storage resources, a smaller and more efficient content prediction model may be able to leverage the transferred knowledge of the embedding model in a more efficient model architecture, which allows for the content prediction model to use less energy and consume less space, and thus run on a wider range of devices. Accordingly, embodiments of the present disclosure provide improved machine learning techniques, and allow for improved content recommendation." Id., I [0031]. These improvements are accomplished by specific features recited in the claims, such as "providing features of a plurality of content items as inputs to an embedding model: receiving embeddings of the plurality of content items as outputs from the embedding model based on the inputs; receiving a data set comprising features of a plurality of users associated with content items of the plurality of content items that correspond to the plurality of users, wherein each respective user of the plurality of users is associated with a respective first set of content items; and training the machine learning model, using a training data set, to output corresponding vector representations of relevant content items for users based on features of the users and corresponding content identifiers for the relevant content items, wherein training the machine learning model is based on a custom loss function comprising a first component corresponding to embeddings and a second component corresponding to content identifiers, wherein the first component is parametrized to have a larger impact on loss than the second component, and wherein the training data set comprises, for each respective user of the plurality of users, the features of the respective user associated with respective labels indicating that: the first set of content items correspond to the respective user; and a second set of content items, having embedding representations that are within a threshold distance of embedding representations of the first set of content items, correspond to the respective user."
In response, the Examiner maintains the rejection as shown above. As discussed previously, improvements to content recommendation is not improvement to the underlying hardware or machine learning model. Using less data with a machine learning model does not improve the machine learning model. The machine learning model remains the same. A “custom loss function” that merely contains data related to the subject matter the machine learning model will be using is not an improvement to the machine learning model, it is an improvement to the abstract idea of weighting the data using for content recommendation. Since there is no improvement to the underlying model or computer hardware, the rejection is maintained as shown above.
Applicant's arguments with respect to claims <CLAIMS> have been considered but are moot in view of the new ground(s) of rejection.
In pg.18, the Applicant argues in regards to the rejection under U.S.C. 103,
As agreed upon during the interview, the cited combination of references does not teach, suggest, or otherwise render obvious the features of Applicant's claims as currently amended. Accordingly, Applicant respectfully requests withdrawal of the rejections under Section 103 of Claims 1, 9, and 15, as well as all claims dependent thereon. See M.P.E.P. § 2143.03 ("If an independent claim is nonobvious under 35 U.S.C. 103, then any claim depending therefrom is nonobvious. In re Fine, 837 F.2d 1071, 5 USPQ2d 1596 (Fed. Cir. 1988)").
In response, the Examiner maintains the rejection as shown above. After reviewing the interview summary and interview Agenda, it is clear the Agenda and the corresponding interview failed to disclose that the amendments provided in the Agenda (filed 12/17/25 under Office Action Appendix in relation to the Interview Summary) that the material added to claim 1 was from dependent claims 4-5. Had Applicant clearly pointed out that all of the material was from claims already rejected under U.S.C. 103, the Examiner would not have said they overcome the art of record. Since the limitations can be clearly met by the references as shown above, the rejection is maintained as shown above using the references from the previous office action.
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 BEN M RIFKIN whose telephone number is (571)272-9768. The examiner can normally be reached Monday-Friday 9 am - 5 pm.
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.
/BEN M RIFKIN/Primary Examiner, Art Unit 2123