DETAILED ACTION
This final office action is responsive to application 17/993,832 with applicant’s remarks and request for reconsideration with amendments filed 12 Mar 2026.
Claim status is currently pending for claims 1-20 with amendment to claims 10, 12 and 18.
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Remarks
Applicant’s responsive remarks filed 03/12/26 are fully considered on the remaining issues as follows
Objection to claims is withdrawn as necessitated by applicant’s amendments.
Rejection of claims under 35 U.S.C. 103 as obvious over a combination of prior art is maintained. Applicant’s traversal has been considered but is not persuasive to sufficiently distinguish from prior art of record. Particularly:
Applicant argues that Wu does not teach or suggest limitation performing local iterative training an item embedding matrix and a user item embedding matrix corresponding to target scoring matrix. However, the examiner respectfully disagrees. Wu introduces [Abst] “In our method, we locally train GNN model in each user client based on the user-item” e.g. [P.2 ¶2] “in our method each user device locally learns a GNN model and the embeddings of the user and items” emphasis locally train/learn. As a whole, Fig 2 plainly illustrates at top-half where users & items (left) are input for embeddings fed into GNN graph neural network for rating (scoring) predictor and computes gradients of loss function, within local user i. Algorithm 1 details implementation with local gradient calculation for local update by computing gradients on user client, e.g. [Sect3.2 Last¶] “compute gradients locally… iteratively executed until the model converges.” The problem formulation sets up notation [Sect3.1] “rating matrix between users and items as Y ∈ RPxQ” and denotes the user and item embeddings [Sect3.2 ¶2]. An example of locally updated embeddings is disclosed [Sect2.2 Last¶]. While it is noted that the user and item embeddings may further be locally tuned, this occurs during training which points to the local training already discussed being that local training is a core aspect of Wu’s method.
Secondly, applicant notes limitations pertaining to noise. The first limitation of sending was rejected over Wu (1st reference) and the second limitation in response was met by Baek (2nd reference) in the same order as presented in the claim. That is to say, 1st art for 1st noise, and 2nd art for 2nd noise. The first noise is added with gradient of loss function and sent to server is a single line limitation. Wu goes above and beyond such breadth of limitation by specifically parameterizing noise [Sect 4.4 ¶1], the noise is added to gradient at Equation 1 [Sect3.3 Last¶] and “uploaded to the server” shown Fig 3. What Wu does not explicitly recite is that the noise is reduced, nor detail a survival notification which is why Baek was brought in. Baek expressly discloses [P.148095] “reduction in the noise” Eq. describes dropout among users, the user set being [P.148097 ¶4] “users who survived” and implemented Alg.2 Lines17-20 notification is n’ of broadcast function communicated responsively for server-client federated learning. While applicant points out that the notification is of a weight sum parameter, it is critical to recognize that the sum is summation is over the set of alive/surviving users, this is the subscript of sigma ∑i∈Uwi where symbol ∈ denotes set membership. Thus, alive users are notified by broadcasting and the recalibration of noise noted in remarks would be appropriate. Replacing noise with reduced noise is indistinguishable from the remarked partial canceling of first noise or smaller in magnitude, all of which simply indicates reduction in noise. No specific equation is required nor pointed to in support from the instant specification as to how noise must be reduced in a technical manner to distinguish from Baek.
Lastly, applicant traverses combination as incompatibility of Wu and Baek. When combining references, MPEP 2144 states that the strongest rationale for combining is the expectation of some advantage, a rationale may be in a reference and supports obviousness if the facts of the case are sufficiently similar. In this case, both Wu and Baek are drawn to the same problem of federated learning for privacy protection thus being sufficiently similar, and the reference Baek provides an advantage as motivated by noise playing a role in reduced privacy loss [P.14097 ¶2] and robustness to user dropout where users generate distributed differentially private noise [P.14096 ¶1]. The test for obviousness is not bodily incorporation but rather what the teachings suggest to a person of ordinary skill in the art whom exhibits at least some reasonable amount of experimentation. Such artisan skilled in relevant machine learning techniques is not limited to only a single technique for noise optimization and the remarks do not distill a technical rationale as to why the principled teachings render inoperable or unsuitable for intended purpose as alleged. Without speculating over specification details, the claims broadly convey techniques that are taught or suggested by the references in combination with a clear motivation for combination. Accordingly, the combination is sufficient to render obvious under §103.
In consideration of the foregoing points when viewing the balance of evidence, the arguments are not found persuasive to sufficiently distinguish the claimed invention from the prior art of record. The arguments presented above support the rejections to independent claims 1, 12, 16 and related dependent claims. Dependent claims are traversed by virtue of dependency without further adding to the substance of merit. Thus, the rejection stands under 35 U.S.C. 103.
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.
Claims 1-3, 5-8, 16-18 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over: Wu et al., “FedGNN: Federated Graph Neural Network for Privacy-Preserving Recommendation” hereinafter Wu (arXiv: 2102.04925v2) in view of Baek et al., “Enhancing Differential Privacy for Federated Learning at Scale” hereinafter Baek, in view of Sun et al., US PG Pub No 2021/0248449A1 hereinafter Sun.
The response to remarks above are incorporated herein.
With respect to claim 1, Wu teaches:
A method performed by a target terminal device corresponding to a target user, the method comprising: {Wu [P.2 ¶2] “we propose a federated framework named FedGNN for privacy-preserving GNN-based recommendation… in our method each user device… user devices compute the gradients of models and user/item embeddings and upload them to a central server” Figs 2-3, Alg.1}
acquiring a target scoring matrix corresponding to the target user, the target scoring matrix describing a scoring situation of the target user on a plurality of items {Wu [Sect 3.1-3.2] “rating matrix between users and items as Y ∈ RPxQ, …ratings that given to these items by user ui are denoted by [yi,1, yi,2, …, yi,K]” so as to “predict the ratings given by user ui to her interacted items [y’i,1, y’i,2, …, y’i,K]” Fig 2 shows rating predictor and y-rating proximate to loss, the ratings matrix corresponds to scoring matrix. User can be from selected (i.e. target) subset [Alg.1 Line6], or simply indexed by subscript}; and
performing, during at least one iteration of locally iteratively training an item embedding matrix and a user embedding matrix corresponding to the target scoring matrix, operations including: {Wu [Alg.1] Iteration count, repeat until convergence FedGNN Fig 2 loop, LocalGradCal is local training, the user and item embeddings are introduced [Sect3.2 ¶2,4] “[eti,1, eti,2, …, eti,K] and [eui,1, eui,2, …, eui,N]” as “embeddings of the user ui and the item embeddings are can be locally tuned during model training …compute gradients locally” Alg.1 LocalGradCal so as to [Abst] “locally train GNN model”}
determining a loss function gradient corresponding to the item embedding matrix in a current iteration {Wu [Sect3.2 ¶3] “loss function Li …We use the loss Li to derive the gradients of the models and embeddings, which are denoted by gmi and gei, respectively” shown Fig 2 and implemented [Alg.1]};
sending to a server the loss function gradient added with a first noise {Wu discloses [Sect3.2 Last¶] “compute gradients locally and send them to the server” notably with [Sect3.3 Last¶] “apply a local differential policy (LDP) module with zero-mean Laplacian noise to the unified gradients… λ is the strength of Laplacian noise. The protected gradients gi are uploaded to the server” The noise is moderated by hyperparameter λ so as to balance privacy and accuracy [Sect4.4 ¶1]};
performing a next iteration according to the item embedding matrix that is updated by the server and the user embedding matrix in the current iteration {Wu discloses [Alg.1] Iteration count, repeat until convergence where subgraphs Gi include user and item shown Fig 3 in a server-client/user loop similar at Fig 2 FedGNN with embeddings for the users and items [Sect3.2], for example [Sect2.2 ¶2] “updated item embeddings”};
predicting a degree of preference of the target user to the plurality of items according to the user embedding matrix and the item embedding matrix obtained at an end of training {Wu [Sect3.3 ¶] “GNN model gradients encode the preference of users on items” Alg.1 Line26 and Fig 2 shows GNN, model gradients and embeddings from user-item data. The GNN is further detailed [Sect3.2 ¶3] where hidden representations are used for prediction of ratings, and Fig 1 “training of GNN based recommendation”}; and
Wu also suggests [Sect3.2 ¶4] “server awakes a certain number of user clients”
However, Wu does not fairly teach the following limitation which is met by Baek:
in response to receiving a survival notification sent by the server, sending to the server a second noise for reducing the first noise added to the loss function gradient {Baek [P.148097 ¶4] “users who survived in the local update process are alive” survived/alive user set under server function for broadcast at Alg.2 Lines17-20 [P.148096] where n’ is notification and σ’ is second noise calculated [P.148095] Eq. “reduction in the noise” describes user dropout, and employs [P.148093 ¶10] “gradient of the loss function L”};
Baek is directed to federated learning with privacy protection thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to employ Baek’s survival/alive user notification and noise in distributed server-client communication in combination for a motivation “we aim to provide 1) a distributed DP-noise generation method from users and 2) robustness to user dropouts” [P.14096 ¶1] and because “noise calibration process plays a role in reducing such increased privacy loss” [P.14097 ¶2].
However, the combination Wu and Baek does not appear to expressly disclose the following limitation which is disclosed by Sun:
recommending one or more items to the target user according to the degree of preference {Sun Figs 6-bottom “User Specific Item Recommendations” similar Fig 7-bottom described [0069-79] “capture the hidden preferences of users” using graph networks that employ embedding representations of user-item interactions for ratings, and which may entail an adjacency matrix}.
Sun is directed to recommender system and methods for user-item matrix factorization or collaborative filtering thus being analogous. A person having ordinary skill in the art would have considered it obvious to recommend items per Sun for the recommendation of Wu in combination as obvious to try when recommending from user-item interactions and/or for a motivation “in order to recommend additional items with similar properties” from a “target or active user’s past behavior” [0006].
With respect to claim 2, the combination of Wu, Baek and Sun teaches the method according to claim 1, wherein:
the target terminal device is among a plurality of user terminal devices configured to perform federated learning {Wu [Sect.3 ¶1] “our federated GNN-based recommendation framework (FedGNN)” shown Fig 2 with plurality of users i-j, [Sect3.2 ¶4] “server aims to coordinate all user devices”}; and
a scoring matrix collected by each of the plurality of user terminal devices has a same item set {Wu [Sect3.1] “rating matrix between users and items” is scoring matrix, and particularly [Sect3.4 ¶2] “server finds the users who interacted with the same items” is same item set}.
With respect to claim 3, the combination of Wu, Baek and Sun teaches the method according to claim 2, further comprising:
determining, by the server, a surviving user terminal device based on the loss function gradient added with the first noise and sent by each of the plurality of user terminal devices {Baek [P.148096] Alg.2 Lines17-20 server function for alive/surviving users, determined by calculation [P.148095 Sect IV.B] in a federated learning technique that is based on loss gradient [P.148093 Sect.C]}.
With respect to claim 5, the combination of Wu, Baek and Sun teaches the method according to claim 2. Baek teaches wherein:
the first noise is zero-mean Gaussian noise and has a first variance coefficient {Baek discloses [P.148095 ¶4] “Gaussian noise is sampled from N(0,σ2I)” zero-mean is first term in parentheses and variance is σ2 consistent with instant specification [0097], a coefficient of variance is the factor n/n’ [P.148095 Sect.B Last2¶]};
the first variance coefficient is greater than or equal to a reciprocal of a set survival number threshold {Baek P.148095 Sect.B Last2¶] “variance of the noise of the final output is n/n’σ2” factor n/n’ is variance coefficient where denominator conveys reciprocal n set of alive users and noise is based on identity matrix I, a threshold is where [P.148097¶3] “minimum weight sum of the users… n’≥n0” similar [P.148099] operand ‘ ≥ ’ is greater than or equal to}; and
the set survival number threshold is determined according to a set drop-out rate and a total number of the plurality of user terminal devices {Baek [P.148095] Tbl.2 “Dropout rate” e.g. 10% or 30% and summation Eq. totals number of users, threshold may include [P.148097 ¶3] “minimum weight sum of the users” and/or [P.148102 ¶2] “If more than k users are alive, the masking of dropped users can be recovered and removed”}.
A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to employ noise, variance and dropout per Baek in combination to arrive at the invention as claimed for a motivation that “user dropouts may occur even during the noise calibration” [P.148097 ¶4] and “The idea behind the construction is to develop a noise calibration step to ensure DP-robustness towards user dropouts in terms of privacy” [P.148100 Last¶], effects are detailed [P.148098].
With respect to claim 6, the combination of Wu, Baek and Sun teaches the method according to claim 1, wherein
the survival notification comprises a number of currently surviving user terminal devices {Baek [P.148095 ¶5-6] “users drop out, and n’(< n) is the weight sum of the alive users. Here, the changed weight sum, n’, is also notified to the alive users” Eq. and Alg.1 Lines 17-20 where ∑w is the weighted sum over set of alive users}.
With respect to claim 7, the combination of Wu, Baek and Sun teaches the method according to claim 6, wherein:
the second noise is a difference between a third noise and the first noise {Baek [P.148095 ¶4] “noise vector… noise update vector wΔ+ei” where Δ is difference for updating noise and the vectorial form conveys 2nd, 3rd or Nth noises};
the third noise is zero-mean Gaussian noise and has a second variance coefficient {Baek [P.148095 ¶4,3] noise vector is sampled from N(0,nwiσ2I) introduced from “Gaussian noise” with zero-mean corresponding to first term in parentheses, a second variance coefficient may correspond to nwi of second term in parentheses preceding variance σ2}; and
a value of the second variance coefficient is between a reciprocal of the number of the currently surviving user terminal devices and a reciprocal of set survival number threshold {Baek [P.148096] Alg.2 Line 15 employs nwi interpreted as a second variance coefficient for the noise vector Line 16 from alive/surviving users, between reciprocals is interpreted to entail identity matrix function denoted I second term of parentheses N(0,nwiσ2I)}.
With respect to claim 8, the combination of Wu, Baek and Sun teaches the method according to claim 1, wherein sending to the server the loss function gradient added with the first noise comprises:
performing encryption processing on the loss function gradient added with the first noise by using a security aggregation algorithm {Wu [Sect3.3] “encryption techniques are applied to gradients to protect private ratings” Fig 3 shows homomorphic encryption between user and server, Alg.2}; and
sending an encryption result to the server {Wu Fig 3 arrows indicate sending to 3rd party server from the encryption (homomorphic) in communication, [Sect3.3-3.4]}.
With respect to claim 16, the rejection of claim 1 is incorporated. The difference in scope being a system comprising processor and memory storing instructions executed by processor to perform limitations similar to method claim 1. Wu shows server-client federated learning Fig 2 and Algorithm 1 instructions to be “executed” [Sect3.2 Last¶] and [Sect3.3 ¶1] “memorize the embedding” to convey necessary computer implementation as would be appreciated by the skilled artisan per instant spec [0179]. A processor and non-transitory verbiage is made express by Sun Fig 8, [0087] as an obvious combination of computer hardware and software.
With respect to claim 17-18, which recite the limitations of claim 2, the combination of Wu, Baek and Sun teaches the system according to claim 16 and further teaches the limitations of claim 2. Therefore, the rejection of claim 2 is applied to claims 17-18.
With respect to claim 20, which recites the limitation of claim 6, the combination of Wu, Baek and Sun teaches the system according to claim 16 and further teaches the limitation of claim 6. Therefore, the rejection of claim 6 is applied to claim 20.
Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Wu, Baek and Sun in view of Zhang et al., “Graph Embedding for Recommendation against Attribute Inference Attacks” hereinafter Zhang, in view of Kondrashkin et al., US PG Pub No 2022/0107872A1 hereinafter Kondrashkin.
With respect to claim 4, the combination of Wu, Baek and Sun teaches the method according to claim 3. Zhang teaches further comprising:
updating, by the server, the item embedding matrix based on an aggregation result of the loss function gradient added with the first noise and the second noise {Zhang discloses [P.3004 Last¶] Eq.4 zv=EVxv (subscript ‘V’, not ‘U’) where Ev ∈ Rd x d1 is described as the item embedding matrix, and zv plugs into Update Eq.9 Aggregate function, illustratively Fig 1 ∑-sum for concatenation with graph convolution having ReLU and MLP (neural net). Further, gradient of loss is implemented Alg.3 final for-loop with Eq.17 and noise is disclosed [P.3007-08 Sect4.2] “injecting Laplace noise to enforce ϵ-DP” differential privacy with “scale factor in determining noise intensity” }; and
Zhang is directed to recommender systems for user-item interactions and rating thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to update item embedding matrix based on aggregate per Zhang in combination for a motivation “we aim to learn a privacy-preserving recommender system that can recommend K products of interest” so as for “better recommendation effectiveness” [P.3004 ¶7-8,1] e.g. “Motivated by the effectiveness of attention mechanism in graph representation learning, we quantify the varied contributions of each element… trainable parameters are shared in the computation of both user and item embeddings” [P.3005 ¶1-2].
However, the combination Wu, Baek and Zhang does not disclose the following limitations which are met by Kondrashkin:
sending, by the server, the item embedding matrix that is updated to the surviving user terminal device {Kondrashkin [0204] “transmit data indicative of the second item embeddings” as [0018] “item embeddings (second version of digital item matrix)” is item embedding matrix to transmit over distributed system Fig 1, updated as per[0204] “second (updated) item embedding” e.g. see Fig 9:916 item embeddings of distributed system for re-trained matrix factorization }.
Kondrashkin is directed to recommendation systems for user-item interactions and matrix factorization thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to transmit item embedding matrix for a motivation of providing recommendation to user smart phone [0077,100] for content feed, and where recommendation is communicatively coupled to a server [0130].
Claims 9-10 are rejected under 35 U.S.C. 103 as being unpatentable over Wu, Baek and Sun in view of Li et al., “Time Interval Aware Self-Attention for Sequential Recommendation” hereinafter Li.
With respect to claim 9, the combination of Wu, Baek and Sun teaches the method according to claim 1, wherein prior to the performing the next iteration according to the item embedding matrix that is updated by the server and the user embedding matrix in the current iteration {Claim 1}. Li teaches the operations further include:
performing clipping processing on the item embedding matrix that is updated and the user embedding matrix that is updated according to an upper limit of a scoring value in the target scoring matrix collected by each of a plurality of user terminal devices, so as to limit a value of a vector therein {Li [P.324 Last¶] “clipped matrix is Muclipped = clip(Mu), where the clip operation of the matrix applies to every element rui j= min(k, rui)” Fig 3 shows items for r with embeddings from Mu, k is thresh [P.325 ¶3] “After retrieving the clipped relation matrix Muclipped, we get the embedding matrix” the embedding layer includes vector representation, and discloses [P.326 ¶1] “scores into a range (0,1)” where range is both an upper and lower limit}.
Li is directed to recommendation techniques for user-item interaction and ratings thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to perform clipping on matrix per Li in combination to arrive at the invention as claimed as applying known techniques to known methods ready for improvement to yield predictable results and/or for a motivation “Clipping the maximum intervals also avoids sparse relation encodings and enables the model to generalize” [P.324 Last¶].
With respect to claim 10, the combination of Wu, Baek, Sun and Li teaches the method according to claim 9, wherein the clipping processing comprises:
converting a negative value therein into zero {Li [P.325 Sect3.4 ¶5] “ReLU activation” known to convert negative values to zero. See also [P.325 Sect 3.3 ¶1] zero padding};
for each vector therein: determining a ratio of a norm of the vector to a square root of the upper limit of the scoring value {Li [P.325] Eq.11 ratio of LayerNorm where “x is a vector”, square root is shown in denominator}; and
performing normalization processing on an element in the vector according to the ratio {Li [P.325] Eq.11 “Layer normalization is used to normalize the inputs across features (i.e. zero-mean and unit-variance)… x is a vector containing all features” and Eq.10 uses LayerNorm for a Dropout function}.
A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to employ ReLU and LayerNorm per Li as applying known techniques to known methods ready for improvement to yield predictable results and/or for a motivation to address problems of “overfitting, unstable training processes (e.g. vanishing gradients), and requiring more training time” [P.325 Sect3.4 ¶6].
Claim 11 is rejected under 35 U.S.C. 103 as being unpatentable over Wu, Baek and Sun in view of Sundaresan et al., US PG Pub No 2022/0245702A1 hereinafter Sundaresan, in view of Krishnan et al., US PG Pub No 2021/0110306A1, and in view of Li.
With respect to claim 11, the combination of Wu, Baek and Sun teaches method according to claim 1. Sundaresan teaches wherein the operations further comprise:
locally randomly initializing the user embedding matrix, and acquiring an initialized item embedding matrix shared by the server {Sundaresan [0083] “randomly initialized… user and item embedding matrices” Fig 1 shows distributed server and clients in communication described [0035-36]}; and
Sundaresan is directed to recommendation from user-item interactions for matrix factorization thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to randomly initialize per Sundaresan in combination for a motivation “matrices become well-defined and can be used… to arrive at the final user-item scores” [0083] and/or to address the issue of “cold start” items for new users without history [0003].
However, Sundaresan in combination does not disclose the following limitation which is met by Krishnan:
fixing the initialized item embedding matrix, and performing pre-training for a set number of iterations on the user embedding matrix {Krishnan discloses [0237] “pre-training epochs” are iterations for [0084] “user and item embedding matrices”, and where fixing is [0113-14] “hold the merchant and user constant.. fixed” e.g. [0076] “fixed user-item pair” see Figs 3A-B and 6C},
Krishnan is directed to recommendation systems for user-item interactions thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to fix and pre-train per Krishnan in combination for a motivation it is “useful to tune the transferred layer to ensure optimal performance in the target domain” [0121].
However, Krishnan in combination does not disclose the following limitation which is met by Li:
wherein clipping processing is performed on the user embedding matrix after each update based on an upper limit of a scoring value in the target scoring matrix collected by each of the plurality of user terminal devices, so as to limit a value of a vector therein {Li discloses [P.324 Last¶] “clipped matrix is Muclipped = clip(Mu), where the clip operation of the matrix applies to every element rui j= min(k, rui)” Fig 3 shows items for r with embeddings from Mu, k is thresh [P.325 ¶3] “After retrieving the clipped relation matrix Muclipped, we get the embedding matrix” the embedding layer includes vector representation, and discloses [P.326 ¶1] “scores into a range (0,1)” range is upper and lower limits}.
Li is directed to recommendation techniques for user-item interaction and ratings thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to perform clipping on matrix per Li in combination to arrive at the invention as claimed as applying known techniques to known methods ready for improvement to yield predictable results and/or for a motivation “Clipping the maximum intervals also avoids sparse relation encodings and enables the model to generalize” [P.324 Last¶].
Claim 12 is rejected under 35 U.S.C. 103 as being unpatentable over: Wu and Baek in view of Zhang and Kondrashkin.
With respect to claim 12, Wu teaches:
A server comprising: {Wu Fig 2 server}
receive a loss function gradient added with a first noise and uploaded by each of a plurality of user terminal devices in a current iteration {Wu [Sect3.2] “receiving the gradients from these users” as “Each client.. uploads the gradients to a central server” shown Fig 2 arrows indicate receiving in a flowchart where gradients of a loss are in a distributed or uploaded in a federated learning between server-client, with [Sect3.3 Last¶] “noise to the unified gradients” and iterations of Alg.1};
However, Wu does not disclose the following limitations which are met by Baek:
determining a currently surviving user terminal device {Baek [P.148095] Equation is determining alive users, thus [P.148097 ¶3] “users who survived in the local update process are alive”};
sending a survival notification to the currently surviving user terminal device {Baek [P.148095 ¶6] “n’, is also notified to the alive users” Alg.2 Lines17-20 server broadcast to alive users};
receiving a second noise sent by the currently surviving user terminal device {Baek [P.148096 ¶4] “sample new noise… each user sends it to the server” Alg.2 server-user communication to send and thus receive by counterparty, noise σ’, i.e. [P.148095 ¶3] “FL so that each user can generate a distributed noise”},
Baek is directed to federated learning with privacy protection thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to employ Baek’s survival/alive user notification and noise in distributed server-client communication in combination for a motivation “we aim to provide 1) a distributed DP-noise generation method from users and 2) robustness to user dropouts” [P.14096 ¶1] and because “noise calibration process plays a role in reducing such increased privacy loss” [P.14097 ¶2].
However, the combination Wu and Baek does not disclose the following limitation which is met by Zhang:
updating an item embedding matrix based on an aggregation result of the loss function gradient added with the first noise and the second noise {Zhang [P.3004 Last¶] Eq.4 zv=EVxv (subscript ‘V’, not ‘U’) where Ev ∈ Rd x d1 is described as the item embedding matrix, and zv plugs into Update Eq.9 Aggregate function, illustratively Fig 1 ∑-sum for concatenation with graph convolution having ReLU and MLP (neural net). Further, gradient of loss is implemented Alg.3 final for-loop with Eq.17 and noise is disclosed [P.3007-08 Sect4.2] “injecting Laplace noise to enforce ϵ-DP” differential privacy with “scale factor in determining noise intensity”}; and
Zhang is directed to recommender systems for user-item interactions and rating thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to update item embedding matrix based on aggregate per Zhang in combination for a motivation “we aim to learn a privacy-preserving recommender system that can recommend K products of interest” so as for “better recommendation effectiveness” [P.3004 ¶7-8,1] e.g. “Motivated by the effectiveness of attention mechanism in graph representation learning, we quantify the varied contributions of each element… trainable parameters are shared in the computation of both user and item embeddings” [P.3005 ¶1-2].
However, the combination Wu, Baek and Zhang does not disclose the following limitations which are met by Kondrashkin:
sending the updated item embedding matrix to the currently surviving user terminal device {Kondrashkin [0204] “transmit data indicative of the second item embeddings” as [0018] “item embeddings (second version of digital item matrix)” is item embedding matrix to transmit over distributed system Fig 1, updated as per[0204] “second (updated) item embedding” e.g. see Fig 9:916 item embeddings of distributed system for re-trained matrix factorization}.
one or more processors {Kondrashkin [0066-67] “processors” e.g. CPU or GPU}; and
one or more memories storing thereon computer-readable instructions that, when executed by the one or more processors, cause the one or more processors to perform acts comprising: {Kondrashkin [0066-67] RAM/ROM memory described for executing software with hardware of computer environment Fig 1}
Kondrashkin is directed to recommendation systems for user-item interactions and matrix factorization thus being analogous. A person having ordinary skill in the art would have considered it obvious prior to the effective filing date to transmit item embedding matrix for a motivation of providing recommendation to user smart phone [0077,100] for content feed, and where recommendation is communicatively coupled to a server [0130]. Further, specifying the use of processor and memory merely conveys obvious and necessary components of computer to implement the techniques according to known devices as would be appreciated by the skilled artisan.
Claims 13-15 are rejected under 35 U.S.C. 103 as being unpatentable over: Wu, Baek, Zhang, Kondrashkin and Sun.
With respect to claim 13, which recites the limitation of claim 5, Baek teaches the limitation of claim 5 and the combination Wu, Baek, Zhang and Kondrashkin teaches the server according to claim 12, further in combination with Sun by dependency of claim 5 on claim 1. Therefore, the rejection of claim 5 is applied to claim 13 with equal motivation applied.
With respect to claim 14, which recites the limitation of claim 6, Baek teaches the limitation of claim 6 and the combination Wu, Baek, Zhang and Kondrashkin teaches the server according to claim 12, further in combination with Sun by dependency of claim 6 on claim 1. Therefore, the rejection of claim 6 is applied to claim 14 with equal motivation applied.
With respect to claim 15, which recites the limitations of claim 7, Baek teaches the limitations of claim 7 and the combination Wu, Baek, Zhang, Kondrashkin and Sun teaches the server according to claim 14. Therefore, the rejection of claim 7 is applied to claim 15 with equal motivation applied.
Claim 19 is rejected under 35 U.S.C. 103 as being unpatentable over: Wu, Baek, Sun, Zhang, Kondrashkin,
With respect to claim 19, which recites the limitations of claim 12, claim 12 is met by the combination of Wu, Baek, Zhang and Kondrashkin, and further the system of claim 18 is met further in combination with Sun by dependency. Therefore, the rejection of claim 12 is applied to claim 19 with equal motivation applied.
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure:
Li et al., “Federated Matrix Factorization with Privacy Guarantee” Inventor disclosure parallels the instant application. 5 authors are noted for inventorship. See [Sect5.1 Last¶] “local fine-tuning…fine-tunes their local version of user/item embeddings”
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 Chase P Hinckley whose telephone number is (571)272-7935. The examiner can normally be reached M-F 9:00 - 5:00.
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, Miranda M. Huang can be reached at 571-270-7092. 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.
/CHASE P. HINCKLEY/Examiner, Art Unit 2124
/MIRANDA M HUANG/Supervisory Patent Examiner, Art Unit 2124
Prosecution Insights