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 .
This action is in response to the application filed on 03/28/2024. Claims 1-20 are pending in the application and have been examined.
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless – (a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claims 1-16; 18-20 are rejected under 35 U.S.C. 102 as being anticipated by Wang et al. (“DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-Scale Learning to Rank Systems”, as disclosed in IDS, hereinafter “Wang”).
Regarding Claim 1,
Wang discloses A recommendation method, wherein the method comprises: obtaining a target feature vector, wherein the target feature vector is obtained by performing feature extraction on attribute information of both a target user and a target article; (Wang [Section 3]; “ This section describes a novel model architecture — DCN-V2 — to learn both explicit and implicit feature interactions. DCN-V2 starts with an embedding layer, followed by a cross network containing multiple cross layers that models explicit feature interactions, and then combines with a deep network that models implicit feature interactions. The improvements made in DCN-V2 are critical for putting DCN into practice for highly-optimized production systems. DCN-V2 significantly improves the expressiveness of DCN[50] in modeling complex explicit cross terms in web-scale production data, while maintaining its elegant formula for easy deployment.”
Wang [Section 3.1]; “ The embedding layer takes input as a combination of categorical (sparse) and dense features, and outputs x0 ∈ R𝑑. For the 𝑖-th categorical feature, we project it from a high-dimensional sparse space to a lower-dimensional dense space via xembed,𝑖 =𝑊embed,𝑖e𝑖, where e𝑖 ∈ {0,1}𝑣𝑖; 𝑊 ∈ R𝑒𝑖×𝑣𝑖 is a learned projection matrix; xembed,𝑖 ∈ R𝑒𝑖 is the dense embedded vector; 𝑣𝑖 and 𝑒𝑖 represents vocab and embedding sizes respectively. For multivalent features, we use the mean of the embedded vectors as the final vector. The output is the concatenation of all the embedded vectors and the normalized dense features:x0 = [xembed,1;. . . ;xembed,𝑛;𝑥dense]. Unlike many related works [13, 16, 26, 34, 35, 46] which requires 𝑒𝑖 = 𝑒𝑗 ∀𝑖, 𝑗, our model accepts arbitrary embedding sizes. This is particularly important for industrial recommenders where the vocab size varies from 𝑂(10) to 𝑂(105). Moreover, our model isn’t limited to the above described embedding method; any other em bedding techniques such as hashing could be adopted” wherein the embedding layer taking as input the combination of categorical and dense features implicitly reads on obtained target feature vectors containing attribute information of target users and target articles across the explicit feature interactions)
processing the target feature vector by using a recommendation model, to obtain recommendation information, wherein the recommendation information indicates a probability that the target user selects the target article (Wang [Section 3.2];
PNG
media_image1.png
338
340
media_image1.png
Greyscale
Wherein the model comprising the combined deep and cross network outputting the predicted binary label of click rate thus reads on obtained recommendation information indicative of a probability that the target user selects the target article (click trough rate prediction))
the recommendation model comprises a cross network, a deep network, and a target network, the cross network comprises a first cross layer and a second cross layer, and the deep network comprises a first deep layer and a second deep layer (Wang [Figure 1];
PNG
media_image2.png
277
337
media_image2.png
Greyscale
Wherein the target layer is interpreted as the top layer of Fig. 1b)
the target network is used to perform fusion processing on a first intermediate output that is output by the first cross layer and a second intermediate output that is output by the first deep layer, to obtain a first fusion result, and the target network is further used to: (Wang [Figure 1];
PNG
media_image2.png
277
337
media_image2.png
Greyscale
Wherein the concatenation of the cross and deep network intermediate outputs is performed to obtain a first fusion result)
process the first fusion result to obtain a first weight corresponding to the first cross layer and a second weight corresponding to the first deep layer, and weight the first fusion result with the first weight and the second weight separately, to obtain a first intermediate input and a second intermediate input and the first intermediate input is input data of the second cross layer, and the second intermediate input is input data of the second deep layer; (Wang [Figure 1];
PNG
media_image2.png
277
337
media_image2.png
Greyscale
Wherein the plurality of fusion nodes in the uppermost fusion layer of the network obtained through the plurality of cross and deep layers thus reads on at least a first fusion result and its associated weight obtained correspondent to the plurality of deep and cross layers in the cross and deep networks; wherein the fusion results obtained through parallel structure combination of the cross network and deep network thus implicitly read on the first fusion results being weighted separately by the weights of the cross and deep to obtain some intermediate inputs respective of the cross and deep layers)
and when the recommendation information meets a preset condition, determining to recommend the target article to the target user (Wang [Section 3.1];
PNG
media_image3.png
91
338
media_image3.png
Greyscale
Wherein the recommender output usage of the embodiment implicitly reads on some recommendation information meeting a preset condition comprising recommending of the target article to a target user)
Regarding Claim 2,
Wang discloses the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Wang further discloses wherein the fusion processing comprises one of point-wise addition, Hadamard product, concatenation, and attention mechanism based pooling (Wang [Section 3.4];
PNG
media_image4.png
275
336
media_image4.png
Greyscale
Wherein the fusion processing comprises at least concatenation).
Regarding Claim 3,
Wang discloses the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Wang further discloses wherein the target network comprises a first feature adaptation network, and the first feature adaptation network is a fully- connected network, a squeeze-and-excitation network, an attention network, an SENet, or a gatenet; and the processing the first fusion result comprises: processing the first fusion result by using the first feature adaptation network (Wang [Page 9 Paragraph 1]; “DCN-V2, on the other hand, consistently outperforms DNN. It successfully leveraged both the explicit and implicit feature inter actions. We attribute this to the balanced number of parameters between the cross network and the deep network (expressive), as well as the simple structure of cross net which eased the opti mization (easy-to-optimize). It’s worth noting that the high-level 8 structure of DCN-V2 shares a similar spirit of the self-attention mechanism adopted in AutoInt, where each feature embedding at tends to a weighed combination of other features. The difference is that during the attention, higher-order interactions were modeled explicitly in DCN-V2 but implicitly in AutoInt.” Wherein the DCN-V2 architecture utilizes a fully connected self-attention network for processing of the fusion results; wherein the first feature adaptation network of the cross network is an attention network).
Regarding Claim 4,
Wang discloses the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Wang further discloses wherein the first intermediate output comprises M first feature vectors, the second intermediate output comprises M second feature vectors, the first fusion result comprises M third feature vectors, the first weight comprises a first weight value corresponding to each first feature vector, and the second weight comprises a second weight value corresponding to each second feature vector (Wang [Figure 3b];
PNG
media_image5.png
143
176
media_image5.png
Greyscale
Wherein the disclosure of gating layer outputs using a mixture of low-rank experts thus implies the application of gating layer outputs between the intermediate layers of the sub-networks (cross and deep) as well as intermediate fusion layer nodes in the final fusion level).
and the weighting the first fusion result with the first weight and the second weight separately comprises: weighting each first feature vector of the M third feature vectors with a corresponding first weight value; and weighting each second feature vector of the M third feature vectors with a corresponding second weight value (Wang [Section 3.2];
PNG
media_image6.png
282
340
media_image6.png
Greyscale
PNG
media_image7.png
189
335
media_image7.png
Greyscale
Wang [Section 3.3-3.4];
PNG
media_image8.png
468
353
media_image8.png
Greyscale
Wherein the plurality of fusion results and their associated combination fusion vectors are interpreted as M third feature vectors; wherein such third feature vectors being dependent on the separate weightings of the deep and cross layer vectors thus on the deep and cross feature vectors being correspondent to the fusion third feature vectors)
Regarding Claim 5,
Wang discloses the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Wang further discloses wherein the cross network further comprises a third cross layer, and the deep network further comprises a third deep layer; the target network is further used to perform fusion processing on a third intermediate output that is output by the second cross layer and a fourth intermediate output that is output by the second deep layer, to obtain a second fusion result, and the target network is further used to: process the second fusion result to obtain a third weight corresponding to the second cross layer and a fourth weight corresponding to the second deep layer, and weight the second fusion result with the third weight and the fourth weight separately, to obtain a third intermediate input corresponding to the second cross layer and a fourth intermediate input corresponding to the second deep layer; and the third cross layer is configured to process the third intermediate input, and the third deep layer is configured to process the fourth intermediate input (Wang [Figure 1];
PNG
media_image2.png
277
337
media_image2.png
Greyscale
Wherein the cross network comprises a plurality of cross layers including at least a third cross layer and its associated plurality of intermediate inputs; wherein the deep network comprises a plurality of deep layers including at least a third deep layer; wherein the outputted fusion layer comprising intermediate outputs of the plurality of deep and cross layers thus reads on third and fourth intermediate outputs to obtain a second fusion result in the fusion level, as well as intermediary third and fourth weights associated with the fusion results).
Regarding Claim 6,
Wang discloses the method of Claim 5 (and thus the rejection of Claim 5 is incorporated). Wang further discloses wherein the target network comprises a second feature adaptation network, and the second feature adaptation network is a fully-connected network, a squeeze-and-excitation network, an attention network, an SENet, or a gatenet; and the processing the second fusion result comprises: processing the second fusion result by using the second feature adaptation network (Wang [Page 9 Paragraph 1]; “DCN-V2, on the other hand, consistently outperforms DNN. It successfully leveraged both the explicit and implicit feature inter actions. We attribute this to the balanced number of parameters between the cross network and the deep network (expressive), as well as the simple structure of cross net which eased the opti mization (easy-to-optimize). It’s worth noting that the high-level 8 structure of DCN-V2 shares a similar spirit of the self-attention mechanism adopted in AutoInt, where each feature embedding at tends to a weighed combination of other features. The difference is that during the attention, higher-order interactions were modeled explicitly in DCN-V2 but implicitly in AutoInt.” Wherein the DCN-V2 architecture utilizes a fully connected self-attention network for processing of the fusion results including the second fusion result; wherein the second feature adaptation network of the deep network is an attention network).
Regarding Claim 7,
Wang discloses the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Wang further discloses wherein the target feature vector comprises a first network input and a second network input; and the obtaining a target feature vector comprises: obtaining an initial feature vector, wherein the initial feature vector is obtained by performing feature extraction on the attribute information of both the target user and the target article; and processing the initial feature vector by using a third feature adaptation network, to obtain a fifth weight corresponding to the cross network and a sixth weight corresponding to the deep network, and weighting the initial feature vector with the fifth weight and the sixth weight separately, to obtain the first network input corresponding to the cross network and the second network input corresponding to the deep network, wherein the first network input is used as an input of the cross network, and the second network input is used as an input of the deep network (Wang [Section 3.1]; “ The embedding layer takes input as a combination of categorical (sparse) and dense features, and outputs x0 ∈ R𝑑. For the 𝑖-th categorical feature, we project it from a high-dimensional sparse space to a lower-dimensional dense space via xembed,𝑖 =𝑊embed,𝑖e𝑖, where e𝑖 ∈ {0,1}𝑣𝑖; 𝑊 ∈ R𝑒𝑖×𝑣𝑖 is a learned projection matrix; xembed,𝑖 ∈ R𝑒𝑖 is the dense embedded vector; 𝑣𝑖 and 𝑒𝑖 represents vocab and embedding sizes respectively. For multivalent features, we use the mean of the embedded vectors as the final vector. The output is the concatenation of all the embedded vectors and the normalized dense features:x0 = [xembed,1;. . . ;xembed,𝑛;𝑥dense]. Unlike many related works [13, 16, 26, 34, 35, 46] which requires 𝑒𝑖 = 𝑒𝑗 ∀𝑖, 𝑗, our model accepts arbitrary embedding sizes. This is particularly important for industrial recommenders where the vocab size varies from 𝑂(10) to 𝑂(105). Moreover, our model isn’t limited to the above described embedding method; any other em bedding techniques such as hashing could be adopted” wherein the embedding layer taking as input the combination of categorical and dense features implicitly reads on obtained target feature vectors containing attribute information of target users and target articles across the explicit feature interaction; wherein such target feature vectors comprise sparse and dense feature inputs; wherein the embedding of features into projection space thus reads on feature extraction on the attribute information of the target user and article to determine such embeddings
Wang [Figure 1];
PNG
media_image2.png
277
337
media_image2.png
Greyscale
Wherein processing the embedding layer’s initial feature vector through the combined cross and deep network thus is interpretable as processing the initial embedding feature vector through a third feature adaptation network (combined cross, deep network) to obtain a plurality of weights (including a fifth and sixth weight separately) that are used as input into the cross and deep layers)
Regarding Claim 8,
Wang discloses the method of Claim 7 (and thus the rejection of Claim 7 is incorporated). Wang further discloses wherein the third feature adaptation network is a fully-connected network, a squeeze-and-excitation network, an attention network, an SENet or a gatenet (Wang [Page 9 Paragraph 1]; “DCN-V2, on the other hand, consistently outperforms DNN. It successfully leveraged both the explicit and implicit feature inter actions. We attribute this to the balanced number of parameters between the cross network and the deep network (expressive), as well as the simple structure of cross net which eased the opti mization (easy-to-optimize). It’s worth noting that the high-level 8 structure of DCN-V2 shares a similar spirit of the self-attention mechanism adopted in AutoInt, where each feature embedding at tends to a weighed combination of other features. The difference is that during the attention, higher-order interactions were modeled explicitly in DCN-V2 but implicitly in AutoInt.” Wherein the DCN-V2 architecture utilizes a fully connected self-attention network for processing of the fusion results including the second fusion result; wherein the third feature adaptation network interpreted as the combination cross/deep network is thus a self-attention network).
Regarding Claim 9,
Wang discloses the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Wang further discloses wherein the attribute information of the target user comprises at least one of the following: a gender, an age, an occupation, income, a hobby, and an education degree (Wang [Section 7.1.1];
PNG
media_image9.png
121
329
media_image9.png
Greyscale
Wherein the target user attribute information from the publicly available MovieLen-1M comprises at least some user-features information comprising Male or Female as well as age).
Regarding Claim 10,
Wang discloses the method of Claim 1 (and thus the rejection of Claim 1 is incorporated). Wang further discloses wherein the attribute information of the target article comprises at least one of the following: an article name, a developer, an installation package size, a category, and a good rating (Wang [Section 7.1.1];
PNG
media_image9.png
121
329
media_image9.png
Greyscale
Wherein the target article information from the publicly available MovieLen-1M comprises at least a good rating).
Claims 11-16 recite the exact method of Claims 1-6 respectively aside from the following additional limitation present in Claim 11 that Wang already discloses:
determining a loss based on the recommendation information and an actual selection result of the target user for the target article, and updating the first recommendation model based on the loss, to obtain a second recommendation model (Wang [Section 3.4]
PNG
media_image10.png
176
331
media_image10.png
Greyscale
)
Thus, Claims 11-16 are rejected for reasons set forth in the rejection of Claim 1-6 respectively.
Claims 18-19 recite a computing device comprising a memory and processor storing instructions to perform the exact method of Claims 1-2 respectively. Thus, Claims 18-19 are rejected for reasons set forth in the rejection of Claim 1-2 respectively.
Claims 20 recites a non-transitory computer storage medium comprising instructions to execute the exact method of Claim 1 respectively. Thus, Claim 20 is rejected for reasons set forth in the rejection of Claim 1 respectively.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable by Wang et al. (“DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-Scale Learning to Rank Systems”, as disclosed in IDS, hereinafter “Wang”) in view of Hu et al. (“Squeeze-and-excitation networks”, as disclosed in IDS, hereinafter “Hu”).
Regarding Claim 17,
Wang discloses the method of Claim 11 (and thus the rejection of Claim 11 is incorporated). Wang already discloses wherein the target feature vector comprises a first network input and a second network input; and the obtaining a target feature vector comprises: obtaining an initial feature vector, wherein the initial feature vector is obtained by performing feature extraction on the attribute information of both the target user and the target article; and processing the initial feature vector by using a third feature adaptation network, to obtain a fifth weight corresponding to the cross network and a sixth weight corresponding to the deep network, and weighting the initial feature vector with the fifth weight and the sixth weight separately, to obtain the first network input corresponding to the cross network and the second network input corresponding to the deep network, wherein the first network input is used as an input of the cross network, and the second network input is used as an input of the deep network (Wang [Section 3.1]; “ The embedding layer takes input as a combination of categorical (sparse) and dense features, and outputs x0 ∈ R𝑑. For the 𝑖-th categorical feature, we project it from a high-dimensional sparse space to a lower-dimensional dense space via xembed,𝑖 =𝑊embed,𝑖e𝑖, where e𝑖 ∈ {0,1}𝑣𝑖; 𝑊 ∈ R𝑒𝑖×𝑣𝑖 is a learned projection matrix; xembed,𝑖 ∈ R𝑒𝑖 is the dense embedded vector; 𝑣𝑖 and 𝑒𝑖 represents vocab and embedding sizes respectively. For multivalent features, we use the mean of the embedded vectors as the final vector. The output is the concatenation of all the embedded vectors and the normalized dense features:x0 = [xembed,1;. . . ;xembed,𝑛;𝑥dense]. Unlike many related works [13, 16, 26, 34, 35, 46] which requires 𝑒𝑖 = 𝑒𝑗 ∀𝑖, 𝑗, our model accepts arbitrary embedding sizes. This is particularly important for industrial recommenders where the vocab size varies from 𝑂(10) to 𝑂(105). Moreover, our model isn’t limited to the above described embedding method; any other em bedding techniques such as hashing could be adopted” wherein the embedding layer taking as input the combination of categorical and dense features implicitly reads on obtained target feature vectors containing attribute information of target users and target articles across the explicit feature interaction; wherein such target feature vectors comprise sparse and dense feature inputs; wherein the embedding of features into projection space thus reads on feature extraction on the attribute information of the target user and article to determine such embeddings
Wang [Figure 1];
PNG
media_image2.png
277
337
media_image2.png
Greyscale
Wherein processing the embedding layer’s initial feature vector through the combined cross and deep network thus is interpretable as processing the initial embedding feature vector through a third feature adaptation network (combined cross, deep network) to obtain a plurality of weights (including a fifth and sixth weight separately) that are used as input into the cross and deep layers)
Wang fails to explicitly disclose but Hu discloses updating the first recommendation model based on the loss, to obtain a second recommendation model comprises: updating the first recommendation model and the-a second feature adaptation network based on the loss, to obtain the second recommendation model and an updated second feature adaptation network (Hu [Page 8 Column 2 Paragraph 1]; “the learning rate is initialised to 0.1 and training continues until the validation loss plateaus2 (∼300 epochs in total). The learning rate is then reduced by a factor of 10 and then this process is repeated (three times in total). Label-smoothing regularisation [20] is used during training.” Wherein the training of the recommendation model until the validation loss plateaus thus reads on updating the first recommendation model based on the loss to obtain a second (post-training) recommendation model and associated updated feature adaptation network)
It would have been obvious to use Wang’s determined loss based on the recommendation information as a threshold metric upon which Hus additional training is conducted until Wang’s loss plateaus. One would have been motivated to do so in order “To allow us to study the upper limit of performance for each variant” (Hu [Page 8 Column 2 Paragraph 1]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
“Method and Apparatus for Training Online Prediction Model, Device and Storage Medium” (US 20210248513 A1) which discloses training of a recommendation prediction model based on some loss
“Method, device, equipment and medium for recommending articles based on sequence recommendation model” (CN 111626827 A) which discloses a target article recommendation model trained through obtained user information and determined through feature interaction layers
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JONATHAN J KIM whose telephone number is (571)272-0523. The examiner can normally be reached 8-6.
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, Matt Ell can be reached on (571) 270-3264. 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.
/JONATHAN J KIM/Examiner, Art Unit 2141
/MATTHEW ELL/Supervisory Patent Examiner, Art Unit 2141
Prosecution Insights