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 .
Priority
Acknowledgment is made of applicant’s claim for foreign priority under 35 U.S.C. 119 (a)-(d). The certified copy has been filed in parent Application No. KR10-2022-0014638 and KR10-2022-0074989, filed on 02/04/2022 and 06/20/2022 respectively.
Response to Amendments
Applicant’s submission filed on 02/12/2026 has been entered. The claim amendments have overcome each and every Claim Rejection under 112(f) and 112(b) as previously set forth in the Office Action mailed 11/21/2025. The status of claims is as follows:
Claims 1-2, 5-16 remain pending in the application.
Claims 3-4 are cancelled.
Claim 16 is newly added.
Claims 1, 2, 5-7, 10, 12-15 are amended.
Response to Arguments
In reference to the Claim Rejections under 35 USC 101:
Applicant’s arguments, see Remarks pg. 9-12, filed 02/12/2026, with respect to Claim Rejections under 35 USC 101 have been fully considered and are persuasive. The rejections under 35 USC 101 have been withdrawn.
In reference to the Claim Rejections under 35 USC 103:
Applicant’s arguments, see Remarks pg. 13-17, filed 02/12/2026, with respect to the rejection(s) of claim(s) under 35 USC 103 have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Lee et al. (“l-Injection: Toward Effective Collaborative Filtering Using Uninteresting Items”) and Lee (2) (“Collaborative Distillation for Top-N
Recommendation”)
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.
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(s) 1-2, 5-6 are rejected under 35 U.S.C. 103 as being unpatentable over Kang et al. (“DE-RRD: A Knowledge Distillation Framework for Recommender System”) in view of GL et al. (“Efficient knowledge distillation of teacher model to multiple student models”) and further in view of Lee et al. (“l-Injection: Toward Effective Collaborative Filtering Using Uninteresting Items”)
Regarding claim 1, Kang explicitly discloses:
A recommendation method, performed by a recommendation system including at
least one processor in a multiple-class collaborative filtering environment, the recommendation method comprising: (Kang, Pg. 1, Abstract: “In this paper, we propose a novel knowledge distillation framework for recommender system, called DE-RRD, which enables the student model to learn from the latent knowledge encoded in the teacher model as well as from the teacher’s predictions.”, Pg. 3, Col. 2, Section 3, ¶[1]: “In this work, we focus on top-𝑁 recommendations for implicit feedback. Let U and I denote the set of users and items, respectively. Given collaborative filtering (CF) information (i.e., implicit interactions between users and items), we build a binary matrix”)
recommending, by an item recommendation unit implemented by the at least one processor, an item having a high pre-use preference and a high post-use preference of a user by using the student model, which has performed learning by using the transferred knowledge information, (Kang, Pg. 3, Col. 2, ¶[2]: “the distillation loss of the existing methods makes the student model follow the teacher’s predictions on unobserved items with particular emphasis on the high-ranked items. In RS, only high-ranked items in the recommendation list are matter. Also, such high-ranked items reveal hidden patterns among entities (i.e., users and items); the high-ranked items in the recommendation list would have strong correlations to the user [25]. By using such additional supervisions from the teacher, they have achieved the comparable performance to the teacher with faster inference time.”, Pg. 3, Col. 2, ¶[3]: “First, the student can be further improved by directly distilling the latent knowledge stored in the teacher model. Latent knowledge refers to all information of users, items, and relationships among them that is discovered and stored in the teacher model. Such knowledge is valuable for the student because it provides detailed explanations on the final prediction of the teacher. Second, they transfer the knowledge from the teacher’s predictions with a point-wise approach that considers a single item at a time.”)
wherein a recommendation model used by the recommendation system comprises the plurality of teacher models and the student model and is configured to recommend the item based on a knowledge distillation technique. (Kang, Pg. 2, Figure 1:
PNG
media_image1.png
402
576
media_image1.png
Greyscale
, Pg. 2, Col. 1, ¶[2]: “In this paper, we propose a novel knowledge distillation framework for RS, named DE-RRD, which distills both the latent knowledge stored in the teacher model (Fig. 1a) and the knowledge revealed from teacher’s predictions (Fig. 1b). By learning both the teacher’s final predictions and the detailed knowledge that provides the bases for such predictions, the student model can be further improved.”)
Kang fails to disclose:
transferring to a student model, by a knowledge transfer unit implemented by at the at least one processor, knowledge information about an item deduced by using collaboration of a plurality of teacher models; and
wherein a first teacher model among the plurality of teacher models has learned a
pre-use preference of a user for an item, which indicates whether the item has been evaluated by the user, and is configured to predict a pre-use preference for items unobserved by the user based on the learned pre-use preference, and
a second teacher model among the plurality of teacher models has learned a post- use preference of the user for an item, which indicates a rating score assigned by the user, and is configured to predict a post-use preference for the items unobserved by the user;
based on predictions of the first teacher model and the second teacher model
However, GL explicitly discloses:
transferring to a student model, by a knowledge transfer unit implemented by at the at least one processor, knowledge information about an item deduced by using collaboration of a plurality of teacher models; and (GL, Pg. 174, Col. 1, Section II, ¶[3]: “While all the aforementioned approaches essentially focus on reducing the complexity of the same network, KD [14] focuses on distilling knowledge from a complex teacher model to smaller student model as shown in Figure 1. The idea here is to train the student model such that it matches the softmax outputs of teacher model to student model. One of the foremost work in this area by Caruana et al. [15] stated that it is possible to compress knowledge in an ensemble of teacher models into a single student model.”, Fig. 1:
PNG
media_image2.png
272
578
media_image2.png
Greyscale
, Pg. 174, Col. 2, Section III, ¶[1]: “In KD we transfer the knowledge of a large model which will act as a Teacher model T to a smaller student model S.”)
based on predictions of the first teacher model and the second teacher model (GL, Pg. 175, Col. 1, ¶[1]: “where yS is the predictions of student model, yT is the predictions of the teacher model, yS, t is the softmax predictions of student model obtained at temperature t, yT, t is softmax predictions of student model obtained at temperature t and y is the true labels which we used to train the teacher model T.”)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Kang and GL. Kang teaches a knowledge distillation framework for recommender system. GL teaches efficient knowledge distillation of teacher model to multiple student models. One of ordinary skill would have motivation to combine Kang and GL to enhance the effectiveness of knowledge transfer to a student model by utilizing collaborative inference from multiple teacher models. Instead of relying on a single teacher’s perspective, knowledge information about an item is deduced through the combined reasoning of several teacher models and organized into a unified knowledge unit. This enables the student model to receive richer, more balanced, and contextually accurate knowledge, leading to improved generalization and learning efficiency.
However, Lee explicitly discloses:
wherein a first teacher model among the plurality of teacher models has learned a
pre-use preference of a user for an item, which indicates whether the item has been evaluated by the user, and is configured to predict a pre-use preference for items unobserved by the user based on the learned pre-use preference, and (Lee, Pg. 5, Col. 2, ¶[2]: “First, we build a pre-use preference matrix
PNG
media_image3.png
25
110
media_image3.png
Greyscale
by examining a rating matrix
PNG
media_image4.png
24
105
media_image4.png
Greyscale
It is set as one if rui ϵ R has been already rated (i.e., u should have liked i if she bought i) (Step 1). It is the highest because pui is set as a real value in [0, 1]. Next, we infer pre-use preference scores on “unrated” user-item pairs (u, i) (i.e., puj = null) based on other observed pre-use preferences (i.e., pui = 1) and add them in P, which becomes ^P (Step 2).”, Pg. 5, Col. 2, Section 3.1, ¶[1]: “It is straightforward to determine a pre-use preference pui if a user u has already rated an item i (i.e., rui # null). This is because i may have been interesting to u at first consideration, i.e.,
PNG
media_image5.png
26
75
media_image5.png
Greyscale
. As such, we set the pre-use preference pui as one. However, when u has not rated i (i.e., rui = null), it is non-trivial to determine pui. Therefore, it is essential to infer pre-use preferences pui if rui is unrated.”, Lee, Pg. 6, Col. 1, ¶[1]: “That is, known pre-use preferences for rated items have positive values (i.e., pui = 1) and missing pre-use preferences for unrated items are ambiguous… We thus employ the OCCF method [8] to infer pre-use preferences.”)
a second teacher model among the plurality of teacher models has learned a post- use preference of the user for an item, which indicates a rating score assigned by the user, and is configured to predict a post-use preference for the items unobserved by the user; (Lee, Pg. 5, Col. 2, ¶[2]: “First, we build a pre-use preference matrix
PNG
media_image3.png
25
110
media_image3.png
Greyscale
by examining a rating matrix
PNG
media_image4.png
24
105
media_image4.png
Greyscale
…The augmented matrix L thus includes both the original ratings for rated items and the imputed ratings for uninteresting items. Lastly, existing CF algorithms are applied to the augmented matrix L. We recommend top-N items by predicting the post-use preferences of empty entries (dotted circles).”) [Examiner’s note: the “empty entries” in Lee is being interpreted as the unobserved or unrated user-item pairs.]
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Kang, GL and Lee. Kang teaches a knowledge distillation framework for recommender system. GL teaches efficient knowledge distillation of teacher model to multiple student models. Lee teaches addressing the sparsity problem of recommender systems. One of ordinary skill would have motivation to combine Kang, GL and Lee because MPEP 2143 sets forth the Supreme Court rationales for obviousness including: (D) Applying a known technique to a known device (method, or product) ready for improvement to yield predictable results; (E): “Obvious to try” choosing from a finite number of identified, predictable solutions, with a reasonable expectation of success; (F) Known work in one field of endeavor may prompt variations of it for use in either the same field or a different one based on design incentives or other market forces if the variations are predictable to one of the ordinary skill in the art.
Regarding Claim 2, the combination of Kang, Lee and GL discloses all the limitations of Claim 1 (as shown in the rejections above).
Kang in view of GL and Lee further discloses:
wherein the transferring comprises transferring, to the student model, an item having a high pre-use preference and an item a having low pre-use preference, predicted by the first teacher model among the plurality of teacher models included in the recommendation model, among items unobserved by the user. (Kang, Pg. 2, Fig. 1:
PNG
media_image6.png
400
575
media_image6.png
Greyscale
, Kang, Pg. 5, Collection 4.2, ¶[2]: “The recommendation list from the teacher model contains information about a user’s potential preference on each unobserved item; A few items that the user would be interested in (i.e., interesting items) are located near the top of the list, whereas the majority of items that the user would not be interested in (i.e., uninteresting items) are located far from the top.”)
Regarding Claim 5, the combination of Kang, Lee and GL discloses all the limitations of Claim 1 (as shown in the rejections above).
Kang in view of GL and Lee further discloses:
wherein the first teacher model is configured to select a first item based on a high pre-use preference predicted by the first teacher model and a second item based on a low pre-use preference predicted by the first teacher model, (Kang, Pg. 4, Figure 2:
PNG
media_image7.png
439
1164
media_image7.png
Greyscale
, Pg. 4, Col. 1, Section 4.1, ¶[1]: “We first introduce “expert” to distill the summarized knowledge that can restore the detailed teacher’s knowledge of each entity. Then, we introduce a novel expert selection strategy for effectively distilling CF knowledge that contains information of all the entities having diverse preferences and characteristics.”, Pg. 4, Col. 2, Section 4.1.2, ¶[2-3]: “The knowledge transfer of DE is conducted in the two steps: 1) a selection network first computes each expert’s degree of specialization for the knowledge to be distilled. 2) DE selects an expert based on the computed distribution, then distills the knowledge through the selected expert… Then, DE selects an expert based on the computed distribution. We represent the selection variable s𝑢 that determines which expert to be selected for distilling ℎ𝑡 (𝑢). s𝑢 is a 𝑀-dimensional one hot vector where an element is set to 1 if the corresponding expert is selected for distillation”) [Examiner’s note: “a high pre-use preference” is being interpreted as an “interesting items K”, and “a low pre-use preference” is being interpreted as an “uninteresting items L”]
wherein the transferring comprises transferring, to the student model, a post-use preference predicted by a second teacher model for the first item, and a post-use preference predicted by the second teacher model for the second item. (Kang, Pg. 5, Col. 2, Section 4.2, ¶[2]: “Because a user is interested in only a few items among the numerous total items [10], learning the detailed ranking orders of all the unobserved items is not only daunting but also ineffective. The recommendation list from the teacher model contains information about a user’s potential preference on each unobserved item; A few items that the user would be interested in (i.e., interesting items) are located near the top of the list, whereas the majority of items that the user would not be interested in (i.e., uninteresting items) are located far from the top”, Pg. 3, Col. 2, Section 4: “DE-RRD consists of two methods: 1) Distillation Experts (DE) that directly transfers the latent knowledge from the teacher, 2) Relaxed Ranking Distillation (RRD) that transfers the knowledge revealed from the teacher’s predictions with direct consideration of ranking orders among items.”, Pg. 4, Figure 2:
PNG
media_image7.png
439
1164
media_image7.png
Greyscale
)
Regarding Claim 6, the combination of Kang, Lee and GL discloses all the limitations of Claim 5 (as shown in the rejections above).
Kang in view of GL and Lee further discloses:
wherein the transferring comprises determining a soft label for the first item as a post-use preference predicted by the second teacher model, and (Kang, Pg. 1, Col. 2, Section 1, ¶[2-3]: “First, the teacher model is trained with the user-item interactions in the training set which has binary labels – ‘1’ for observed interactions, and ‘0’ for unobserved interactions. Then, the student model is trained with the “soft” labels generated by the teacher model (i.e., teacher’s predictions) along with the available binary labels… The core idea behind this process is that the soft labels predicted by the teacher model reveal hidden relations among entities (i.e., users and items) not explicitly included in the training set, so that they accelerate and improve the learning of the student model. Specifically, the items ranked near the top of a user’s recommendation list would have strong correlations to the items that the user has interacted before [25]. Also, the soft labels provide guidance for distinguishing the items that each user would like and the items that each user would not be interested in among numerous unobserved items only labeled as ‘0’ [13].”)
determining a soft label for the second item as a rating score equal to or less than a preset reference, (Kang, Pg. 1, Col. 2, Section 1, ¶[2]: “The knowledge transfer is conducted as follows: First, the teacher model is trained with the user-item interactions in the training set which has binary labels – ‘1’ for observed interactions, and ‘0’ for unobserved interactions. Then, the student model is trained with the “soft” labels generated by the teacher model (i.e., teacher’s predictions) along with the available binary labels.”)
instead of a post-use preference predicted by the second teacher model for the second item. (Lee, Pg. 5, Col. 2, ¶[2]: “We recommend top-N items by predicting the post-use preferences of empty entries (dotted circles).”)
Claim(s) 7-15 are rejected under 35 U.S.C. 103 as being unpatentable over Kang et al. (“DE-RRD: A Knowledge Distillation Framework for Recommender System”) in view of GL et al. (“Efficient knowledge distillation of teacher model to multiple student models”), Lee et al. (“l-Injection: Toward Effective Collaborative Filtering Using Uninteresting Items”) and further in view of Kim et al. (US 2016/0171036 A1)
Regarding Claim 7, Kang explicitly discloses:
A recommendation method, performed by a recommendation system including a knowledge transfer unit and item recommendation unit, [implemented by at least one processor], in a multiple-class collaborative filtering environment, the knowledge transfer unit including a first teacher and a second teacher, the recommendation method comprising: : (Kang, Pg. 1, Abstract: “In this paper, we propose a novel knowledge distillation framework for recommender system, called DE-RRD, which enables the student model to learn from the latent knowledge encoded in the teacher model as well as from the teacher’s predictions.”, Pg. 3, Col. 2, Section 3, ¶[1]: “In this work, we focus on top-𝑁 recommendations for implicit feedback. Let U and I denote the set of users and items, respectively. Given collaborative filtering (CF) information (i.e., implicit interactions between users and items), we build a binary matrix”)
learning, by the first teacher, a pre-use preference among pieces of multiple-class feedback received from a user, and (Kang, Pg. 1, Col. 2, Section 1, ¶[2]: “First, the teacher model is trained with the user-item interactions in the training set which has binary labels – ‘1’ for observed interactions, and ‘0’ for unobserved interactions.”, Pg. 5, Section 4.2, ¶[2]: “The recommendation list from the teacher model contains information about a user’s potential preference on each unobserved item; A few items that the user would be interested in (i.e., interesting items) are located near the top of the list, whereas the majority of items that the user would not be interested in (i.e., uninteresting items) are located far from the top.”) [Examiner’s note: “a pre-use preference of the user for an item” is being interpreted as “unobserved interactions” item]
learning, by the second teacher, a post-use preference among the pieces of multiple-class feedback; (Kang, Pg. 1, Col. 2, Section 1, ¶[2]: “First, the teacher model is trained with the user-item interactions in the training set which has binary labels – ‘1’ for observed interactions, and ‘0’ for unobserved interactions.”, Pg. 5, Section 4.2, ¶[2]: “The recommendation list from the teacher model contains information about a user’s potential preference on each unobserved item; A few items that the user would be interested in (i.e., interesting items) are located near the top of the list, whereas the majority of items that the user would not be interested in (i.e., uninteresting items) are located far from the top.”) [Examiner’s note: “a post-use preference of the user for an item” is being interpreted as “observed interactions” item]
selecting items, to be transferred to a student model based on the predicted pre-use preference; (Kang, Pg. 4, Figure 2:
PNG
media_image7.png
439
1164
media_image7.png
Greyscale
, Pg. 4, Col. 1, Section 4.1, ¶[1]: “We first introduce “expert” to distill the summarized knowledge that can restore the detailed teacher’s knowledge of each entity. Then, we introduce a novel expert selection strategy for effectively distilling CF knowledge that contains information of all the entities having diverse preferences and characteristics.”, Pg. 4, Col. 2, Section 4.1.2, ¶[2-3]: “The knowledge transfer of DE is conducted in the two steps: 1) a selection network first computes each expert’s degree of specialization for the knowledge to be distilled. 2) DE selects an expert based on the computed distribution, then distills the knowledge through the selected expert… Then, DE selects an expert based on the computed distribution. We represent the selection variable s𝑢 that determines which expert to be selected for distilling ℎ𝑡 (𝑢). s𝑢 is a 𝑀-dimensional one hot vector where an element is set to 1 if the corresponding expert is selected for distillation”) [Examiner’s note: “a high pre-use preference” is being interpreted as an “interesting items K”, and “a low pre-use preference” is being interpreted as an “uninteresting items L”]
determining, by the second teacher, a soft label based on a post-use preference, which is predicted for items selected by the first teacher based on the learned post-use preference, and (Kang, Pg. 1, Col. 2, Section 1, ¶[2-3]: “First, the teacher model is trained with the user-item interactions in the training set which has binary labels – ‘1’ for observed interactions, and ‘0’ for unobserved interactions. Then, the student model is trained with the “soft” labels generated by the teacher model (i.e., teacher’s predictions) along with the available binary labels… The core idea behind this process is that the soft labels predicted by the teacher model reveal hidden relations among entities (i.e., users and items) not explicitly included in the training set, so that they accelerate and improve the learning of the student model. Specifically, the items ranked near the top of a user’s recommendation list would have strong correlations to the items that the user has interacted before [25]. Also, the soft labels provide guidance for distinguishing the items that each user would like and the items that each user would not be interested in among numerous unobserved items only labeled as ‘0’ [13].”)
performing, by the student model, learning based on the received distilled knowledge, and (Kang, Pg. 3, Col. 2, Section 4, ¶[1]: “We propose DE-RRD framework which enables the student model to learn both from the teacher’s predictions and from the latent knowledge encoded in the teacher model. DE-RRD consists of two methods: 1) Distillation Experts (DE) that directly transfers the latent knowledge from the teacher, 2) Relaxed Ranking Distillation (RRD) that transfers the knowledge revealed from the teacher’s predictions with direct consideration of ranking orders among items.”, Pg. 4, Col. 1, Section 4.1.1, ¶[2]: “The student model has smaller capacity compared to the teacher (𝑑𝑠 << 𝑑𝑡 ). By minimizing the equation 4, the student learns compressed information on the user’s preference that can restore more detailed knowledge in the teacher as accurate as possible. This approach provides a kind of filtering effect and improves the learning of the student model.”)
Kang fails to disclose:
implemented by at least one processor
the pre-use preference indicating whether the item has been evaluated by the user,
the post-use preference indicating a rating score assigned by the user;
predicting, by the first teacher, a pre-use preference for items unobserved by the user based on the learned pre-use preference, and
among the items unobserved by the user
transferring the determined soft label as distilled knowledge to the student model; and
recommending items having a high pre-use preference and a high post- use preference by using the item recommendation unit.
However, GL explicitly discloses
transferring the determined soft label as distilled knowledge to the student model; and (GL, Pg. 174, Col. 1, Section II, ¶[3]: “While all the aforementioned approaches essentially focus on reducing the complexity of the same network, KD [14] focuses on distilling knowledge from a complex teacher model to smaller student model as shown in Figure 1. The idea here is to train the student model such that it matches the softmax outputs of teacher model to student model. One of the foremost work in this area by Caruana et al. [15] stated that it is possible to compress knowledge in an ensemble of teacher models into a single student model.”, Fig. 1:
PNG
media_image2.png
272
578
media_image2.png
Greyscale
, Pg. 174, Col. 2, Section III, ¶[1]: “In KD we transfer the knowledge of a large model which will act as a Teacher model T to a smaller student model S.”)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Kang and GL. Kang teaches a knowledge distillation framework for recommender system. GL teaches efficient knowledge distillation of teacher model to multiple student models. One of ordinary skill would have motivation to combine Kang and GL to enhance the effectiveness of knowledge transfer to a student model by utilizing collaborative inference from multiple teacher models. Instead of relying on a single teacher’s perspective, knowledge information about an item is deduced through the combined reasoning of several teacher models and organized into a unified knowledge unit. This enables the student model to receive richer, more balanced, and contextually accurate knowledge, leading to improved generalization and learning efficiency.
However, Lee explicitly discloses:
the pre-use preference indicating whether the item has been evaluated by the user, (Lee, Pg. 5, Col. 2, ¶[2]: “First, we build a pre-use preference matrix
PNG
media_image3.png
25
110
media_image3.png
Greyscale
by examining a rating matrix
PNG
media_image4.png
24
105
media_image4.png
Greyscale
It is set as one if rui ϵ R has been already rated (i.e., u should have liked i if she bought i) (Step 1). It is the highest because pui is set as a real value in [0, 1]. Next, we infer pre-use preference scores on “unrated” user-item pairs (u, i) (i.e., puj = null) based on other observed pre-use preferences (i.e., pui = 1) and add them in P, which becomes ^P (Step 2).”, Pg. 5, Col. 2, Section 3.1, ¶[1]: “It is straightforward to determine a pre-use preference pui if a user u has already rated an item i (i.e., rui # null). This is because i may have been interesting to u at first consideration, i.e.,
PNG
media_image5.png
26
75
media_image5.png
Greyscale
. As such, we set the pre-use preference pui as one. However, when u has not rated i (i.e., rui = null), it is non-trivial to determine pui. Therefore, it is essential to infer pre-use preferences pui if rui is unrated.”, Lee, Pg. 6, Col. 1, ¶[1]: “That is, known pre-use preferences for rated items have positive values (i.e., pui = 1) and missing pre-use preferences for unrated items are ambiguous… We thus employ the OCCF method [8] to infer pre-use preferences.”)
the post-use preference indicating a rating score assigned by the user; (Lee, Pg. 5, Col. 2, ¶[2]: “First, we build a pre-use preference matrix
PNG
media_image3.png
25
110
media_image3.png
Greyscale
by examining a rating matrix
PNG
media_image4.png
24
105
media_image4.png
Greyscale
…The augmented matrix L thus includes both the original ratings for rated items and the imputed ratings for uninteresting items. Lastly, existing CF algorithms are applied to the augmented matrix L. We recommend top-N items by predicting the post-use preferences of empty entries (dotted circles).”)
among the items unobserved by the user (Lee, Pg. 5, Col. 2, ¶[2]: “We recommend top-N items by predicting the post-use preferences of empty entries (dotted circles).”) [Examiner’s note: the “empty entries” in Lee is being interpreted as the unobserved or unrated user-item pairs.]
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Kang and Lee. Kang teaches a knowledge distillation framework for recommender system. Lee teaches addressing the sparsity problem of recommender systems. One of ordinary skill would have motivation to combine Kang and Lee because MPEP 2143 sets forth the Supreme Court rationales for obviousness including: (D) Applying a known technique to a known device (method, or product) ready for improvement to yield predictable results; (E): “Obvious to try” choosing from a finite number of identified, predictable solutions, with a reasonable expectation of success; (F) Known work in one field of endeavor may prompt variations of it for use in either the same field or a different one based on design incentives or other market forces if the variations are predictable to one of the ordinary skill in the art.
However, Kim explicitly discloses:
implemented by at least one processor (Kim, ¶[0031]: “The item recommendation apparatus 102 may be, for example, a processor included in an application server to provide a user with an application associated with content, or in a content production server to produce and distribute content.”)
predicting, by the first teacher, a pre-use preference for items unobserved by the user based on the learned pre-use preference, and (Kim, ¶[0039]: “The users' pre-use preferences for the unrated items are determined based on post-use preference for rated items that is used and evaluated by the user. The user's pre-use preferences for unrated items that have not yet used by the user may be predicted based on the observed pre-use preference matrix implying pre-use preferences of rated items.)
recommending items having a high pre-use preference and a high post- use preference by using the item recommendation unit. (Kim, ¶[0071-0072]: “In addition, a set of items that are likely to get high ratings are preferred items, denoted by Iupre , which is another subset of interesting items as shown in FIG. 4. Based on this viewpoint, the item recommendation apparatus identifies top-N preferred items to user u by using both pre-use and post-use preferences.”)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Kang and Kim. Kang teaches a knowledge distillation framework for recommender system. Kim teaches item recommendation method and apparatus which may recognize items preferred before a user uses items, based on items rated by the user, and may recommend an item to the user based on preferences for the recognized items. One of ordinary skill would have motivation to combine Kang and Kim because MPEP 2143 sets forth the Supreme Court rationales for obviousness including: (D) Applying a known technique to a known device (method, or product) ready for improvement to yield predictable results; (E): “Obvious to try” choosing from a finite number of identified, predictable solutions, with a reasonable expectation of success; (F) Known work in one field of endeavor may prompt variations of it for use in either the same field or a different one based on design incentives or other market forces if the variations are predictable to one of the ordinary skill in the art.
Regarding Claim 8, the combination of Kang, GL, Lee and Kim discloses all the limitation of Claim 7 (as shown in the rejections above).
Kang in view of GL, Lee and Kim further discloses:
wherein the knowledge transfer unit is configured to train the first teacher by generating a pre-use preference matrix based on items having a record of being evaluated by the user, and (Kang, Pg. 3, Col. 2, Section 3: “Given collaborative filtering (CF) information (i.e., implicit interactions between users and items), we build a binary matrix 𝑹 ∈ {0, 1} |U|×|I| . Each element of 𝑹 has a binary value indicating whether a user has interacted with an item (1) or not (0). Note that an unobserved interaction does not necessarily mean a user’s negative preference on an item, it can be that the user is not aware of the item. For each user, a recommender model ranks all items that have not interacted with the user (i.e., unobserved items) and provides a ranked list of top-𝑁 unobserved items.”)
train the second teacher by generating a post-use preference matrix based on a rating score actually evaluated by the user for the items having the record of being evaluated by the user. (Kang, Pg. 3, Col. 2, Section 3: “Given collaborative filtering (CF) information (i.e., implicit interactions between users and items), we build a binary matrix 𝑹 ∈ {0, 1} |U|×|I| . Each element of 𝑹 has a binary value indicating whether a user has interacted with an item (1) or not (0). Note that an unobserved interaction does not necessarily mean a user’s negative preference on an item, it can be that the user is not aware of the item. For each user, a recommender model ranks all items that have not interacted with the user (i.e., unobserved items) and provides a ranked list of top-𝑁 unobserved items.”)
Regarding Claim 9, the combination of Kang, GL, Lee and Kim discloses all the limitation of Claim 7 (as shown in the rejections above).
Kang in view of GL, Lee and Kim further discloses:
wherein the student model is configured to receive the distilled knowledge that is distilled twice by using collaboration of the first teacher and the second teacher. (Kang, Pg. 2, Col. 1, ¶[2-3]: “we propose a novel knowledge distillation framework for RS, named DE-RRD, which distills both the latent knowledge stored in the teacher model (Fig. 1a) and the knowledge revealed from teacher’s predictions (Fig. 1b). By learning both the teacher’s final predictions and the detailed knowledge that provides the bases for such predictions, the student model can be further improved… DE adopts the multiple experts and a novel expert selection strategy that clearly distinguishes the knowledge that each expert distills based on the correlations among the entities in the teacher representation space”)
Regarding Claim 10, the combination of Kang, Lee, GL and Kim discloses all the limitation of Claim 7 (as shown in the rejections above).
Kang in view of GL, Lee and Kim further discloses:
wherein the knowledge transfer unit is configured to: use, as the soft label, the post-use preference predicted by the second teacher only for an item having a high pre-use preference among the items, selected by the first teacher among the items unobserved by the user, and (Kang, Pg. 1, Col. 2, Section 1, ¶[2-3]: “First, the teacher model is trained with the user-item interactions in the training set which has binary labels – ‘1’ for observed interactions, and ‘0’ for unobserved interactions. Then, the student model is trained with the “soft” labels generated by the teacher model (i.e., teacher’s predictions) along with the available binary labels… The core idea behind this process is that the soft labels predicted by the teacher model reveal hidden relations among entities (i.e., users and items) not explicitly included in the training set, so that they accelerate and improve the learning of the student model. Specifically, the items ranked near the top of a user’s recommendation list would have strong correlations to the items that the user has interacted before [25]. Also, the soft labels provide guidance for distinguishing the items that each user would like and the items that each user would not be interested in among numerous unobserved items only labeled as ‘0’ [13].”)
determine a rating score equal to or less than a preset reference for an item having a low pre-use preference among the items selected by the first teacher as the soft label, and (Kim, ¶0064-0065]: “hen, user u assigns a high rating to Movie #1 and a low rating to Movie #2. In contrast, user u does not watch Movie #3, so post-use preference of user u for the Movie #3 remains unknown (i.e., empty in a rating matrix). Because user u is less interested in Movie #3, Movie #3 is likely to get a low rating score from user u. The rating score for Movie #3 can be regarded 0 because user u is less interested in Movie #3 than Movie #2. That is, user u implicitly gives an extremely low rating to Movie #3, (i.e., uninteresting item), based on "peruse preferences" of the user u. Note that some unrated items are implicitly rated as low ratings based on a user's pre-use preferences without post-use preferences.”)
instead of the post-use preference predicted by the second teacher (Lee, Pg. 5, Col. 2, ¶[2]: “First, we build a pre-use preference matrix
PNG
media_image3.png
25
110
media_image3.png
Greyscale
by examining a rating matrix
PNG
media_image4.png
24
105
media_image4.png
Greyscale
…The augmented matrix L thus includes both the original ratings for rated items and the imputed ratings for uninteresting items. Lastly, existing CF algorithms are applied to the augmented matrix L. We recommend top-N items by predicting the post-use preferences of empty entries (dotted circles).”)
transfers the soft label to the student model as the distilled knowledge. (GL, Pg. 174, Col. 1, Section II, ¶[3]: “While all the aforementioned approaches essentially focus on reducing the complexity of the same network, KD [14] focuses on distilling knowledge from a complex teacher model to smaller student model as shown in Figure 1. The idea here is to train the student model such that it matches the softmax outputs of teacher model to student model. One of the foremost work in this area by Caruana et al. [15] stated that it is possible to compress knowledge in an ensemble of teacher models into a single student model.”, Fig. 1:
PNG
media_image2.png
272
578
media_image2.png
Greyscale
, Pg. 174, Col. 2, Section III, ¶[1]: “In KD we transfer the knowledge of a large model which will act as a Teacher model T to a smaller student model S.”)
Regarding Claim 11, the combination of Kang, GL, Lee and Kim discloses all the limitation of Claim 1 (as shown in the rejections above).
Kang in view of GL, Lee and Kim further discloses:
A non-transitory computer-readable medium storing program executable by at least one processor to perform the recommendation method of claim 1 (Kim, ¶[0199]: “The method according to the above-described embodiments of the present invention may be recorded in non-transitory computer-readable media including program instructions to implement various operations embodied by a computer.”)
Regarding Claim 12, Kang explicitly discloses:
at least one processor to implement: at least one memory configured to store computer program executable by the at least one processor, wherein the computer program, when executed by the at least one processor, causes the at least one processor to: (Kang, Pg. 8, Col. 1, ¶[2]: “All inferences are made using PyTorch with CUDA from Tesla P40 GPU and Xeon on Gold 6148 CPU.”)
recommend an item having a high pre- use preference and a high post-use preference of the user by using the student model having learned by using the transferred knowledge information, (Kang, Pg. 3, Col. 2, ¶[2]: “the distillation loss of the existing methods makes the student model follow the teacher’s predictions on unobserved items with particular emphasis on the high-ranked items. In RS, only high-ranked items in the recommendation list are matter. Also, such high-ranked items reveal hidden patterns among entities (i.e., users and items); the high-ranked items in the recommendation list would have strong correlations to the user [25]. By using such additional supervisions from the teacher, they have achieved the comparable performance to the teacher with faster inference time.”, Pg. 3, Col. 2, ¶[3]: “First, the student can be further improved by directly distilling the latent knowledge stored in the teacher model. Latent knowledge refers to all information of users, items, and relationships among them that is discovered and stored in the teacher model. Such knowledge is valuable for the student because it provides detailed explanations on the final prediction of the teacher. Second, they transfer the knowledge from the teacher’s predictions with a point-wise approach that considers a single item at a time.”)
wherein the at least one processor includes a knowledge distillation technique framework using a recommendation model, the recommendation model comprising the plurality of teacher models and the student model (Kang, Pg. 4, Figure 2:
PNG
media_image7.png
439
1164
media_image7.png
Greyscale
, Pg. 4, Col. 1, Section 4.1, ¶[1]: “We first introduce “expert” to distill the summarized knowledge that can restore the detailed teacher’s knowledge of each entity. Then, we introduce a novel expert selection strategy for effectively distilling CF knowledge that contains information of all the entities having diverse preferences and characteristics.”, Pg. 4, Col. 2, Section 4.1.2, ¶[2-3]: “The knowledge transfer of DE is conducted in the two steps: 1) a selection network first computes each expert’s degree of specialization for the knowledge to be distilled. 2) DE selects an expert based on the computed distribution, then distills the knowledge through the selected expert… Then, DE selects an expert based on the computed distribution. We represent the selection variable s𝑢 that determines which expert to be selected for distilling ℎ𝑡 (𝑢). s𝑢 is a 𝑀-dimensional one hot vector where an element is set to 1 if the corresponding expert is selected for distillation)
Kang fails to disclose:
transfer, to a student model, knowledge information related with an item deduced by using collaboration of a plurality of teacher models; and
wherein a first teacher model among the plurality of teacher models has learned a
pre-use preference of a user for an item, which indicates whether the item has been evaluated by the user, and is configured to predict a pre-use preference for items unobserved by the user based on the learned pre-use preference, and
a second teacher model among the plurality of teacher models has learned a post- use preference of the user for an item, which indicates a rating score assigned by the user, and is configured to predict a post-use preference for the items unobserved by the user;
based on predictions of the first teacher model and the second teacher model
However, GL explicitly discloses:
a knowledge transfer unit configured to transfer, to a student model, knowledge information related with an item deduced by using collaboration of a plurality of teacher models; and (GL, Pg. 174, Col. 1, Section II, ¶[3]: “While all the aforementioned approaches essentially focus on reducing the complexity of the same network, KD [14] focuses on distilling knowledge from a complex teacher model to smaller student model as shown in Figure 1. The idea here is to train the student model such that it matches the softmax outputs of teacher model to student model. One of the foremost work in this area by Caruana et al. [15] stated that it is possible to compress knowledge in an ensemble of teacher models into a single student model.”, Fig. 1:
PNG
media_image2.png
272
578
media_image2.png
Greyscale
, Pg. 174, Col. 2, Section III, ¶[1]: “In KD we transfer the knowledge of a large model which will act as a Teacher model T to a smaller student model S.”)
based on predictions of the first teacher model and the second teacher model (GL, Pg. 175, Col. 1, ¶[1]: “where yS is the predictions of student model, yT is the predictions of the teacher model, yS, t is the softmax predictions of student model obtained at temperature t, yT, t is softmax predictions of student model obtained at temperature t and y is the true labels which we used to train the teacher model T.”)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Kang and GL. Kang teaches a knowledge distillation framework for recommender system. GL teaches efficient knowledge distillation of teacher model to multiple student models. One of ordinary skill would have motivation to combine Kang and GL to enhance the effectiveness of knowledge transfer to a student model by utilizing collaborative inference from multiple teacher models. Instead of relying on a single teacher’s perspective, knowledge information about an item is deduced through the combined reasoning of several teacher models and organized into a unified knowledge unit. This enables the student model to receive richer, more balanced, and contextually accurate knowledge, leading to improved generalization and learning efficiency.
However, Lee explicitly discloses:
wherein a first teacher model among the plurality of teacher models has learned a
pre-use preference of a user for an item, which indicates whether the item has been evaluated by the user, and is configured to predict a pre-use preference for items unobserved by the user based on the learned pre-use preference, and (Lee, Pg. 5, Col. 2, ¶[2]: “First, we build a pre-use preference matrix
PNG
media_image3.png
25
110
media_image3.png
Greyscale
by examining a rating matrix
PNG
media_image4.png
24
105
media_image4.png
Greyscale
It is set as one if rui ϵ R has been already rated (i.e., u should have liked i if she bought i) (Step 1). It is the highest because pui is set as a real value in [0, 1]. Next, we infer pre-use preference scores on “unrated” user-item pairs (u, i) (i.e., puj = null) based on other observed pre-use preferences (i.e., pui = 1) and add them in P, which becomes ^P (Step 2).”, Pg. 5, Col. 2, Section 3.1, ¶[1]: “It is straightforward to determine a pre-use preference pui if a user u has already rated an item i (i.e., rui # null). This is because i may have been interesting to u at first consideration, i.e.,
PNG
media_image5.png
26
75
media_image5.png
Greyscale
. As such, we set the pre-use preference pui as one. However, when u has not rated i (i.e., rui = null), it is non-trivial to determine pui. Therefore, it is essential to infer pre-use preferences pui if rui is unrated.”, Lee, Pg. 6, Col. 1, ¶[1]: “That is, known pre-use preferences for rated items have positive values (i.e., pui = 1) and missing pre-use preferences for unrated items are ambiguous… We thus employ the OCCF method [8] to infer pre-use preferences.”)
a second teacher model among the plurality of teacher models has learned a post- use preference of the user for an item, which indicates a rating score assigned by the user, and is configured to predict a post-use preference for the items unobserved by the user; (Lee, Pg. 5, Col. 2, ¶[2]: “First, we build a pre-use preference matrix
PNG
media_image3.png
25
110
media_image3.png
Greyscale
by examining a rating matrix
PNG
media_image4.png
24
105
media_image4.png
Greyscale
…The augmented matrix L thus includes both the original ratings for rated items and the imputed ratings for uninteresting items. Lastly, existing CF algorithms are applied to the augmented matrix L. We recommend top-N items by predicting the post-use preferences of empty entries (dotted circles).”) [Examiner’s note: the “empty entries” in Lee is being interpreted as the unobserved or unrated user-item pairs.]
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Kang, GL and Lee. Kang teaches a knowledge distillation framework for recommender system. GL teaches efficient knowledge distillation of teacher model to multiple student models. Lee teaches addressing the sparsity problem of recommender systems. One of ordinary skill would have motivation to combine Kang, GL and Lee because MPEP 2143 sets forth the Supreme Court rationales for obviousness including: (D) Applying a known technique to a known device (method, or product) ready for improvement to yield predictable results; (E): “Obvious to try” choosing from a finite number of identified, predictable solutions, with a reasonable expectation of success; (F) Known work in one field of endeavor may prompt variations of it for use in either the same field or a different one based on design incentives or other market forces if the variations are predictable to one of the ordinary skill in the art.
Regarding Claim 13, the combination of Kang, GL, Lee and Kim discloses all the limitation of Claim 12 (as shown in the rejections above).
Kang in view of GL and Lee further discloses:
wherein computer program, when executed by the at least one processor, further causes the at least one processor (Kang, Pg. 8, Col. 1, ¶[2]: “All inferences are made using PyTorch with CUDA from Tesla P40 GPU and Xeon on Gold 6148 CPU.”)
transfer unit is further configured to transfer, to the student model, an item having a high pre-use preference and an item having a low pre-use preference predicted by a teacher model among the plurality of teacher models included in the recommendation model among items unevaluated by the user, and (Kang, Pg. 2, Fig. 1:
PNG
media_image6.png
400
575
media_image6.png
Greyscale
, Kang, Pg. 5, Collection 4.2, ¶[2]: “The recommendation list from the teacher model contains information about a user’s potential preference on each unobserved item; A few items that the user would be interested in (i.e., interesting items) are located near the top of the list, whereas the majority of items that the user would not be interested in (i.e., uninteresting items) are located far from the top.”)
the teacher model has learned a pre-use preference of the user for the item. (Kang, Pg. 1, Col. 2, Section 1, ¶[2]: “First, the teacher model is trained with the user-item interactions in the training set which has binary labels – ‘1’ for observed interactions, and ‘0’ for unobserved interactions.”, Pg. 5, Section 4.2, ¶[2]: “The recommendation list from the teacher model contains information about a user’s potential preference on each unobserved item; A few items that the user would be interested in (i.e., interesting items) are located near the top of the list, whereas the majority of items that the user would not be interested in (i.e., uninteresting items) are located far from the top.”) [Examiner’s note: “a pre-use preference of the user for an item” is being interpreted as “unobserved interactions” item]
Regarding Claim 14, the combination of Kang, GL and Kim discloses all the limitation of Claim 12 (as shown in the rejections above).
Kang in view of GL further discloses:
wherein the first teacher model is configured to select a first item based on a high pre-use preference predicted by the first teacher model and a second item based on a low pre- use preference predicted by the first teacher model, and (Kang, Pg. 4, Figure 2:
PNG
media_image7.png
439
1164
media_image7.png
Greyscale
, Pg. 4, Col. 1, Section 4.1, ¶[1]: “We first introduce “expert” to distill the summarized knowledge that can restore the detailed teacher’s knowledge of each entity. Then, we introduce a novel expert selection strategy for effectively distilling CF knowledge that contains information of all the entities having diverse preferences and characteristics.”, Pg. 4, Col. 2, Section 4.1.2, ¶[2-3]: “The knowledge transfer of DE is conducted in the two steps: 1) a selection network first computes each expert’s degree of specialization for the knowledge to be distilled. 2) DE selects an expert based on the computed distribution, then distills the knowledge through the selected expert… Then, DE selects an expert based on the computed distribution. We represent the selection variable s𝑢 that determines which expert to be selected for distilling ℎ𝑡 (𝑢). s𝑢 is a 𝑀-dimensional one hot vector where an element is set to 1 if the corresponding expert is selected for distillation”) [Examiner’s note: “a high pre-use preference” is being interpreted as an “interesting items K”, and “a low pre-use preference” is being interpreted as an “uninteresting items L”)
the computer program, when executed by the at least one processor, further causes the at least one processor (Kang, Pg. 8, Col. 1, ¶[2]: “All inferences are made using PyTorch with CUDA from Tesla P40 GPU and Xeon on Gold 6148 CPU.”)
to transfer, to the student model, a post-use preference for the first item, predicted by a second teacher model, and a post- use preference for the second item, predicted by the second teacher model. (Kang, Pg. 5, Col. 2, Section 4.2, ¶[2]: “Because a user is interested in only a few items among the numerous total items [10], learning the detailed ranking orders of all the unobserved items is not only daunting but also ineffective. The recommendation list from the teacher model contains information about a user’s potential preference on each unobserved item; A few items that the user would be interested in (i.e., interesting items) are located near the top of the list, whereas the majority of items that the user would not be interested in (i.e., uninteresting items) are located far from the top”, Pg. 3, Col. 2, Section 4: “DE-RRD consists of two methods: 1) Distillation Experts (DE) that directly transfers the latent knowledge from the teacher, 2) Relaxed Ranking Distillation (RRD) that transfers the knowledge revealed from the teacher’s predictions with direct consideration of ranking orders among items.”, Pg. 4, Figure 2:
PNG
media_image7.png
439
1164
media_image7.png
Greyscale
)
Regarding Claim 15, the combination of Kang, GL, Lee and Kim discloses all the limitation of Claim 12 (as shown in the rejections above).
Kang in view of GL and Lee further discloses:
wherein computer program, when executed by the at least one processor, further causes the at least one processor (Kang, Pg. 8, Col. 1, ¶[2]: “All inferences are made using PyTorch with CUDA from Tesla P40 GPU and Xeon on Gold 6148 CPU.”)
learn, by using the student model, a pre-use preference and a post-use preference for the item transferred as the knowledge information by using the collaboration of the plurality of teacher models, and (Kang, Pg. 1, Col. 2, Section 1, ¶[2]: “First, the teacher model is trained with the user-item interactions in the training set which has binary labels – ‘1’ for observed interactions, and ‘0’ for unobserved interactions.”, Pg. 5, Section 4.2, ¶[2]: “The recommendation list from the teacher model contains information about a user’s potential preference on each unobserved item; A few items that the user would be interested in (i.e., interesting items) are located near the top of the list, whereas the majority of items that the user would not be interested in (i.e., uninteresting items) are located far from the top.”) [Examiner’s note: “a pre-use preference of the user for an item” is being interpreted as “unobserved interactions” item]
recommend an item equal to or greater than a preset reference for each user by using the learned student model. (Kim, ¶[0071-0072]: “In addition, a set of items that are likely to get high ratings are preferred items, denoted by Iupre , which is another subset of interesting items as shown in FIG. 4. Based on this viewpoint, the item recommendation apparatus identifies top-N preferred items to user u by using both pre-use and post-use preferences.”)
Claim(s) 16 is rejected under 35 U.S.C. 103 as being unpatentable over Kang et al. (“DE-RRD: A Knowledge Distillation Framework for Recommender System”) in view of GL et al. (“Efficient knowledge distillation of teacher model to multiple student models”) and further in view of Lee et al. (“l-Injection: Toward Effective Collaborative Filtering Using Uninteresting Items”) and Lee(2) (“Collaborative Distillation for Top-N Recommendation”)
Regarding claim 16, the combination of Kang, Lee and GL discloses all the limitations of Claim 1 (as shown in the rejections above).
Kang in view of GL and Lee fails to disclose:
wherein, in the transferring, only items of which a high pre-use preference predicted by the first teacher model is at a top percentage or higher and items of which a low pre-use preference predicted by the first teacher model is at a bottom percentage or lower are transferred to the student model.
However, Lee(2) explicitly discloses:
wherein, in the transferring, only items of which a high pre-use preference predicted by the first teacher model is at a top percentage or higher and items of which a low pre-use preference predicted by the first teacher model is at a bottom percentage or lower are transferred to the student model. (Lee(2), Pg. 2, ¶[4]: “Our model is influenced by the idea of RD, treating items differently based on their rankings. In the ranking problem, the higher-ranked items are more important because they can be potential inclusions in top-N recommendation. Therefore, we sample items in the soft target according to their rankings; the higher the ranking, the more the items are sampled. Because we sample both high and low-ranked items in a probabilistic manner, our model can learn both positive/negative correlations among items.”)
It would have obvious to one of ordinary skill in the art before the effective filing date of the present application to combine Kang, GL, Lee and Lee(2). Kang teaches a knowledge distillation framework for recommender system. GL teaches efficient knowledge distillation of teacher model to multiple student models. Lee teaches addressing the sparsity problem of recommender systems. Lee (2) teaches Collaborative Distillation for Top-N Recommendation. One of ordinary skill would have motivation to combine Kang, GL, Lee and Lee (2) because MPEP 2143 sets forth the Supreme Court rationales for obviousness including: (D) Applying a known technique to a known device (method, or product) ready for improvement to yield predictable results; (E): “Obvious to try” choosing from a finite number of identified, predictable solutions, with a reasonable expectation of success; (F) Known work in one field of endeavor may prompt variations of it for use in either the same field or a different one based on design incentives or other market forces if the variations are predictable to one of the ordinary skill in the art.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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 AMY TRAN whose telephone number is (571)270-0693. The examiner can normally be reached Monday - Friday 7:30 am - 5:00 pm EST.
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, David Yi can be reached at (571) 270-7519. 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.
/AMY TRAN/Examiner, Art Unit 2126
/DAVID YI/Supervisory Patent Examiner, Art Unit 2126
Prosecution Insights