A time series recommendation method and system based on double-channel collaborative distillation

Abstract

The application provides a time sequence recommendation method and system based on double-channel collaborative distillation, relates to the technical field of recommendation systems, and acquires a user set, an item set and an interaction triple; constructs a time sequence interaction graph and a folding interaction graph; inputs the time sequence interaction graph into a time sequence dynamic channel, sequentially performs interval division, timestamp-aware edge coding, local snapshot modeling and global sequence modeling on the time sequence interaction graph, obtains user time sequence representation and item time sequence representation; utilizes a semantic representation channel, performs edge reliability estimation and weighted graph propagation on the semantic representation channel, obtains user semantic preference and item semantic preference representation; performs collaborative distillation and gate fusion on the time sequence representation and the semantic preference representation, and obtains final user representation and item representation; and calculates a prediction score to generate a Top-K unseen item recommendation list. The application can retain local collaborative structure and preference evolution information, and reduce irrelevant semantic propagation and heterogeneous channel interference.

Description

Technical Field

[0001] This application belongs to the field of recommendation system technology, and in particular relates to a time-series recommendation method and system based on dual-channel collaborative distillation. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Recommender systems typically generate personalized candidate ranking lists based on historical interactions between users and items, and have become a core technology for businesses such as e-commerce, local services, and video platforms. With the continuous expansion of user scale and item variety, and the increasing dynamism and diversity of user interests, recommender systems not only need to efficiently model user preferences from massive amounts of data, but also need the ability to capture the temporal evolution of interests, utilize multi-source auxiliary information, and integrate heterogeneous knowledge.

[0004] Currently, while graph neural network-based methods can aggregate high-order collaborative relationships on user-item bipartite graphs, when all historical interactions are folded into a time-independent static graph, local behavioral patterns at different stages are easily mixed. Static graph neural network recommendations suffer from the problem of losing temporal information and find it difficult to simultaneously preserve the collaborative structure within a time interval and the preference evolution between time intervals.

[0005] However, while traditional time-series or sequential recommendation systems incorporate interaction order or continuous time information, their modeling focus is often on the transfer patterns of adjacent interactions, lacking explicit preservation of higher-order collaborative structures within local time intervals. Furthermore, fixing a single time scale makes it difficult to simultaneously account for both short-term, phased preferences and long-term, intertemporal changes, limiting the model's ability to capture the multi-scale dynamics of user interests and leading to a significant decline in recommendation accuracy.

[0006] While existing semantic enhancement recommendation methods based on large language models can generate semantic vectors from text such as user interaction history, item titles, categories, and descriptions to supplement information in sparse interaction scenarios, offline semantic vectors only represent the general semantics of the text and cannot fully match users' personalized preferences. If semantic features are propagated indiscriminately along all user-item interaction edges, irrelevant and distorted semantic information will spread along the graph structure, polluting node representations. In other words, the semantic priors introduced by large language models have the risk of diffusion and preference mismatch problems.

[0007] Furthermore, the behavioral representations obtained from time sequence diagrams and the semantic representations obtained from texts are heterogeneous in terms of information sources and representation spaces. Direct splicing, averaging, or forced global alignment can easily lead to problems such as representation redundancy, representation distortion, channel interference, or loss of channel-specific information, making it difficult to achieve effective information complementarity. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a time-series recommendation method and system based on dual-channel collaborative distillation. The method obtains time-series behavior representation and semantic preference representation through a time-series dynamic channel and a semantic representation channel, respectively. While maintaining the specialization of the two types of representations, it performs prediction-level and feature-level bidirectional distillation and uses instance-level dimensional gating for fusion, thereby outputting a Top-K unseen item recommendation list for the target user.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a time-series recommendation method based on dual-channel synergistic distillation, comprising: Retrieve the user set, item set, and interaction triplet; Construct a temporal interaction graph and a folding interaction graph based on interaction triples; The temporal interaction graph is input into the temporal dynamic channel. The temporal interaction graph is then divided into intervals and timestamp-aware edge encoding is performed sequentially to obtain an encoded temporal subgraph. Local snapshot modeling and global sequence modeling are then performed sequentially on the encoded temporal subgraph to obtain user temporal representation and item temporal representation. By utilizing semantic representation channels, edge reliability estimation and weighted graph propagation are performed on the semantic representation channels to obtain representations of user semantic preferences and item semantic preferences; Cooperative distillation and gated fusion are performed on temporal representations and semantic preference representations to obtain end-user representations and item representations; Calculate the predicted score based on the end-user representation and item representation, and generate a Top-K list of unseen items.

[0010] As one implementation method, the temporal interaction graph is sequentially divided into intervals and timestamp-aware edge encoded to obtain an encoded temporal subgraph. Specifically, the temporal interaction graph is divided into several temporal subgraphs according to time intervals; the interaction edges of the temporal subgraphs are timestamp-aware edge encoded to obtain the encoded temporal subgraph.

[0011] As one implementation method, local snapshot modeling is performed on the coding time sequence subgraph. The specific process is as follows: The time series subgraph is divided into several time windows based on the timestamp, resulting in several local snapshot images; By using the dimension-wise maximum aggregation operation, the time-encoded neighbor information in the local snapshot image is aggregated to obtain the time-aware neighbor information of users and items respectively. Based on the time-aware neighbor information of users and items, a multilayer perceptron is used to stitch together and update the local time snapshot map to obtain the local representation of users and the local representation of items.

[0012] As one implementation method, global sequence modeling is performed on the user local representation and the item local representation to obtain the user temporal representation and the item temporal representation, respectively. The specific process is as follows: Based on the local representations of users and items, a sequence of interval-level representations of users and items is formed; The interval-level representation sequences of users and items are input into the gated loop unit for encoding in chronological order to obtain the dependency relationships between adjacent snapshots of users and items. A multi-head self-attention mechanism is used to obtain the attention weights of the time snapshot, and the time states after attention are aggregated to obtain the temporal behavior representations of users and items respectively.

[0013] As one implementation method, edge reliability estimation and weighted graph propagation are performed on the semantic representation channels to obtain user semantic preference representations and item semantic preference representations. The specific process is as follows: Based on user sets and item sets, semantic prior initialization is performed on the offline large language model to obtain the initial semantic state of users and items; Based on the semantic state of users and items, multilayer perceptron is used to calculate the layer-by-layer edge compatibility score and convert it into edge reliability weights; Perform normalized bidirectional semantic propagation of users and items based on reliability weights to obtain the semantic states of users and items at each propagation layer. The semantic states of users and items at each propagation layer are aggregated across propagation layers to obtain a semantic preference representation.

[0014] As one implementation method, collaborative distillation and gated fusion are performed on temporal representations and semantic preference representations to obtain end-user representations and item representations. The specific process is as follows: Construct a unique representation loss and impose feature loss constraints on user temporal representation, item temporal representation, user semantic preference representation, and item semantic preference representation; Based on temporal and semantic representations respectively, temporal and semantic prediction score distributions are generated, and prediction-level bidirectional distillation is performed using symmetric KL divergence. Based on temporal and semantic representations respectively, a bidirectional projection bridge is constructed to form a bidirectional projection, and feature-level bidirectional distillation is performed using the mean square error between each projection representation and the corresponding target channel representation. By using user-adaptive gating and item-adaptive gating, the distilled dual-channel representations are weighted and fused to obtain the final user representation and the final item representation.

[0015] As one implementation method, temporal and semantic prediction score distributions are generated based on temporal and semantic representations, respectively, and bidirectional distillation at the prediction level is performed using symmetric KL divergence. The specific process is as follows: Based on user time-series representation and item time-series representation, calculate the relevant scores of the time-series dynamic channel for positive and negative items; Based on user semantic representation and item semantic representation, calculate the relevant scores of semantic representation channels for positive and negative items; Based on the correlation scores of the temporal dynamic channel and the semantic representation channel, the positive and negative candidate prediction distributions of the two channels are obtained using temperature softmax. Predictive-level bidirectional distillation is performed to minimize the symmetric KL divergence between the two prediction distributions.

[0016] As one implementation, a bidirectional projection bridge is constructed to form a bidirectional projection, wherein the bidirectional projection bridge includes a temporal-to-semantic projection network and a semantic-to-temporal projection network, and a bidirectional projection representation is formed through two dedicated projection networks in two directions.

[0017] As one implementation method, during training, a total loss function is constructed to optimize the model. The formula for the total loss function is: ; in, Represents the total loss function of the model; This represents the BPR recommendation loss; This represents the total loss from co-distillation; This represents the total contrastive learning loss; This represents all trainable parameters; This represents the contrastive learning and L2 regularization weights; This represents the squared L2 norm of the parameter.

[0018] A second aspect of the present invention provides a time-series recommendation system based on dual-channel synergistic distillation, comprising: The data acquisition module is used to acquire user sets, item sets, and interaction triples; The interaction graph construction module is used to construct temporal interaction graphs and folded interaction graphs based on interaction triples; The temporal dynamic channel module is used to input the temporal interaction graph into the temporal dynamic channel, and to perform interval division and timestamp-aware edge encoding on the temporal interaction graph in sequence to obtain the encoded temporal subgraph; and to perform local snapshot modeling and global sequence modeling on the encoded temporal subgraph in sequence to obtain the user temporal representation and the item temporal representation. The semantic representation channel module is used to perform edge reliability estimation and weighted graph propagation on the semantic representation channel to obtain user semantic preferences and item semantic preferences representations. The collaborative distillation and gating fusion module is used to perform collaborative distillation and gating fusion on temporal representations and semantic preference representations to obtain the final user representation and item representation; The prediction module is used to calculate prediction scores based on end-user and item representations, and generate a Top-K list of unseen items.

[0019] A third aspect of the present invention provides a computer device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to perform the steps of the method described in the first aspect of the present invention.

[0020] A fourth aspect of the present invention is to provide a computer-readable storage medium having a program stored thereon that, when executed by a processor, performs the steps described in the first aspect of the present invention.

[0021] The above one or more technical solutions have the following beneficial effects: In this embodiment, a temporal dynamic channel and a semantic representation channel are constructed. Time intervals are divided and timestamp-aware edge encoding is performed in the temporal dynamic channel to obtain an encoded temporal subgraph. Local snapshot modeling and global sequence modeling are performed on the encoded temporal subgraph in sequence. Using the semantic representation channel, edge reliability estimation and weighted graph propagation are performed on the semantic representation channel. Combined with large model semantic priorization, this solves the problem that existing recommendation systems have difficulty in simultaneously maintaining high-order collaborative structures within local time intervals and characterizing preference evolution across time intervals. At the same time, it overcomes the defect of noise propagation caused by the indiscriminate diffusion of offline semantic priors along the interaction graph.

[0022] In this embodiment, by performing collaborative distillation and adaptive gating fusion on temporal and semantic representations, the problem of effective fusion of heterogeneous representations is solved, a balance between the preservation of proprietary information and the controllable exchange of knowledge is achieved, and a personalized and enhanced final representation is generated to output accurate Top-K recommendation results, thereby improving recommendation accuracy.

[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 This is a schematic diagram of the overall framework of a time-series recommendation method based on dual-channel synergistic distillation in this embodiment. Figure 2 This is a graph showing the comparison of recommendation performance of different algorithms in this embodiment 1 under multiple datasets as the number of retained training interactions changes; Figure (a) shows the comparison curves of recall@20 and normalized depreciation cumulative gain@20 of each algorithm on the Yelp dataset as a function of the number of retained training interactions; Figure (b) shows the comparison curves of recall@20 and normalized depreciation cumulative gain@20 of each algorithm on the Amazon-Electronics dataset as a function of the number of retained training interactions; and Figure (c) shows the comparison curves of recall@20 and normalized depreciation cumulative gain@20 of each algorithm on the ML-10M dataset as a function of the number of retained training interactions. Figure 3 This is a comparison curve of the recommendation robustness of different algorithms in this embodiment 1 under multiple datasets, as the proportion of semantic prior perturbation changes. Figure (a) shows the comparison curves of recall retention rate and normalized loss cumulative gain retention rate of various algorithms on the Yelp dataset as a function of the semantic prior perturbation ratio; Figure (b) shows the comparison curves of recall retention rate and normalized loss cumulative gain retention rate of various algorithms on the Amazon-Electronics dataset as a function of the semantic prior perturbation ratio; Figure (c) shows the comparison curves of recall retention rate and normalized loss cumulative gain retention rate of various algorithms on the ML-10M dataset as a function of the semantic prior perturbation ratio. Figure 4 This is a bar chart comparing the performance degradation of users with stable preferences and users with drifting preferences after removing different channels in this embodiment. Figure (a) is a bar chart comparing the decrease in recall rate between users with stable preferences and users with drifting preferences after removing the time-series dynamic channel and the semantic representation channel; Figure (b) is a bar chart comparing the decrease in normalized depreciation cumulative gain between users with stable preferences and users with drifting preferences after removing the time-series dynamic channel and the semantic representation channel. Figure 5 This is a heatmap showing the distribution of temporal attention weights for various preference-drifting user groups under different datasets in this embodiment. Figure (a) shows the heatmap of the time attention weight distribution for different preference-drifting user groups on the Yelp dataset; Figure (b) shows the heatmap of the time attention weight distribution for different preference-drifting user groups on the Amazon-Electronics dataset; and Figure (c) shows the heatmap of the time attention weight distribution for different preference-drifting user groups on the ML-10M dataset. Figure 6 This is a comparison chart of the semantic reliability score distribution of clean semantic edges and disturbed semantic edges under different datasets in this embodiment 1; Figure (a) shows a comparison of the semantic reliability score distribution of clean semantic edges and disturbed semantic edges in the Yelp dataset; Figure (b) shows a comparison of the semantic reliability score distribution of clean semantic edges and disturbed semantic edges in the Amazon-Electronics dataset; and Figure (c) shows a comparison of the semantic reliability score distribution of clean semantic edges and disturbed semantic edges in the ML-10M dataset. Figure 7 This is a heatmap of cross-channel similarity representing each module under different datasets in this embodiment 1; Figure (a) shows the cross-channel similarity heatmap of each module in the Yelp dataset; Figure (b) shows the cross-channel similarity heatmap of each module in the Amazon-Electronics dataset; and Figure (c) shows the cross-channel similarity heatmap of each module in the ML-10M dataset. Figure 8 This is a graph showing the cosine similarity evolution of different item themes over time snapshots in this embodiment 1. Figure 9 This is an adaptive channel allocation and dimensional gating graph for a representative user in this embodiment. Figure (a) shows the user-level dual-channel weight allocation diagram for users 1565 and 3564; Figure (b) shows the dimension-by-dimensional time-series dynamic channel gating distribution diagram for users 1565 and 3564. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] Example 1 This embodiment discloses a time-series recommendation method based on dual-channel synergistic distillation.

[0030] To more clearly illustrate this embodiment, a timing recommendation implementation process based on dual-channel synergistic distillation can be specifically described as follows: A time-series recommendation method based on dual-channel synergistic distillation includes: S1. Obtain the user set, the item set, and the interaction triplet consisting of user ID, item ID, and interaction timestamp; S2. Construct a temporal interaction diagram and a folded interaction diagram based on interaction triples; S3. Input the time-series interaction graph into the time-series dynamic channel, and perform interval division and timestamp-aware edge encoding on the time-series interaction graph in sequence to obtain the encoded time-series subgraph; perform local snapshot modeling and global sequence modeling on the encoded time-series subgraph in sequence to obtain the user time-series representation and the item time-series representation; S4. Using the semantic representation channel, perform edge reliability estimation and weighted graph propagation on the semantic representation channel to obtain the user semantic preference representation and the item semantic preference representation; S5. Perform co-distillation and gated fusion on the temporal representation and semantic preference representation to obtain the end-user representation and item representation; S6. Calculate the predicted score based on the end-user representation and item representation, and generate a Top-K unseen item recommendation list.

[0031] This embodiment employs a dual-graph, dual-channel, and specialization-then-coordination overall technical concept to construct a time-series recommendation model based on dual-channel collaborative distillation. This model mainly includes a time-series dynamic channel, a semantic representation channel, a collaborative distillation and fusion module, a prediction module, and a training and optimization module. For the same batch of timestamped user-item interactions, on the one hand, a time-series interaction graph retaining time information is constructed for the time-series dynamic channel to extract collaborative relationships within intervals and evolution between intervals; on the other hand, a folded interaction graph ignoring time but retaining all observed connections is constructed for the semantic representation channel to refine offline semantic priors based on edge reliability. The two channels first form their respective specialized representations, then coordinate them through prediction-level and feature-level bidirectional distillation, and finally generate a fused representation through instance-level dimensional gating.

[0032] In step S1, the user set, the item set, and the interaction triplet consisting of user ID, item ID, and interaction timestamp are obtained.

[0033] Specifically, user-item historical interaction records are obtained from Yelp, Amazon-Electronics, and ML-10M datasets. Each interaction is represented as a triple (u,i,t). The user set (U), item set (I), and interaction triple set (D) are determined by all interaction records.

[0034] In step S2, a temporal interaction graph and a folded interaction graph are constructed based on the interaction triples.

[0035] Specifically, the timing interaction graph is defined as ,in, For any interactive triple The timestamp of this user-item interaction Recorded as ,Right now This allows us to retain the time information corresponding to each interaction.

[0036] Foldable interaction graph is defined as ,in, Only user-item connections are retained.

[0037] like Figure 1 As shown, in step S3, the time-series interaction graph is input into the time-series dynamic channel, and the time-series interaction graph is sequentially divided into intervals and timestamp-aware edge encoding to obtain an encoded time-series subgraph; the encoded time-series subgraph is sequentially modeled by local snapshots and global sequences to obtain user time-series representation and item time-series representation.

[0038] S301. The temporal interaction graph is divided into intervals and timestamp-aware edge encoding is performed sequentially to obtain the encoded temporal subgraph.

[0039] (1) Divide the time series interaction graph into several time series subgraphs according to the time interval.

[0040] (2) Encode the interaction edges of the time series subgraph with time stamp awareness to obtain the encoded time series subgraph.

[0041] The interaction timestamp encoding as a learnable periodic time vector can be represented as: (1) in, A d-dimensional encoding representing the interaction timestamp; This represents the timestamp of the interaction between user u and item i; This represents a learnable d-by-1 mapping matrix; This represents a learnable d-dimensional bias. This formula is used to encode interaction timestamps into learnable periodic time vectors.

[0042] S302. Perform local snapshot modeling and global sequence modeling on the encoded temporal subgraph in sequence to obtain the user temporal representation and the item temporal representation.

[0043] S3021. Perform local snapshot modeling on each time series subgraph to obtain local representations of users and items under each time window.

[0044] (1) Divide the time series subgraph into several time windows according to the timestamp to obtain several local snapshot images.

[0045] The time series subgraph is divided into several non-overlapping time windows of equal size based on the timestamp, and a snapshot graph is constructed for each time window.

[0046] Divide the range from the earliest timestamp to the latest timestamp into T non-overlapping time windows, and construct a snapshot graph for each time window.

[0047] (2) Using the dimension-wise maximum aggregation operation, the time-encoded neighbor information in the local snapshot map is aggregated to obtain time-aware neighbor information.

[0048] By using a learnable matrix and a cosine transform of the bias to encode the timestamp of each interaction within the snapshot, the neighbor item representation is added to the corresponding time code and a dimension-wise maximum aggregation is performed.

[0049] By aggregating time-coded neighbor messages within a local time snapshot, time-aware neighbor information is obtained.

[0050] The formula for user node time-aware neighbor information is: (2) in, This represents the neighbor messages of user u at layer l in the t-th snapshot; Let represent the set of neighboring items of user u within the t-th snapshot; This represents the upper-level representation of item i; This represents the dimension-wise maximum aggregation operation; This represents the zero vector returned when there are no neighbors. This formula is used to aggregate time-coded neighbor messages within a local time snapshot.

[0051] The formula for time-aware neighbor messages corresponding to an item node is: (3) in, This represents the neighbor messages of user i at layer l in the t-th snapshot; Let represent the set of neighboring items of user i within the t-th snapshot; This represents the upper-level representation of user u; This represents the dimension-wise maximum aggregation operation; This represents the zero vector returned when there are no neighbors.

[0052] (3) Based on time-aware neighbor information, the local time snapshot map is stitched and updated using a multilayer perceptron to obtain the local representation of the user and the local representation of the item.

[0053] 1) Concatenate the representation of the central node with the time-aware neighbor information, and update it through a multilayer perceptron. That is, concatenate the current node state with the time-aware neighbor messages and update the node to obtain the updated node representations for users and items, respectively, using the following formulas: , (4) in, This represents the l-th layer representation of user u in the t-th snapshot; This represents a multilayer perceptron; This represents the vector concatenation operation; This indicates that the user is aware of neighbor messages over time. This represents the l-th layer representation of item i in the t-th snapshot; This represents the time-aware neighbor message for an item. This expression is used to concatenate the current node state with the time-aware neighbor message and update the node representation.

[0054] 2) Based on the updated node representation, obtain the user local representation and the item local representation.

[0055] Perform dual processing on the item nodes and sum the representations of each propagation layer to obtain the time-ordered sequence of user and item interval-level representations, namely the user local representation and the item local representation, with the following formulas: , (5) in, This represents the aggregate representation of user u in the t-th snapshot; Indicates the number of propagation layers in the graph; This represents the user representation at layer l; This represents the aggregated representation of item i in the t-th snapshot. This formula is used to aggregate snapshot-level node representations across graph propagation layers.

[0056] S3022. Perform global sequence modeling on the local representations of users and items to obtain the temporal representations of users and items, respectively.

[0057] (1) Based on the local representation of users and the local representation of items, form a sequence of interval-level representations of users and items.

[0058] The interval-level representation sequence that makes up users and items is given by the following formula: (6) in, Represents the interval-level representation sequence of user u and item i; Indicates the number of time snapshots; Let represent the user and item representations of the t-th snapshot. This formula is used to construct the interval-level representation sequence of users and items.

[0059] (2) Input the interval-level representation sequence of users and items into the gated loop unit for encoding in chronological order to obtain the dependency relationship between adjacent snapshots.

[0060] The interval-level representation is input into the gated loop unit in chronological order, so that each time state contains both the current snapshot information and the dependency of the previous snapshot.

[0061] The dependency relationship between adjacent time snapshots is encoded using GRU, and the formula is as follows: (7) in, Represents the contextualized temporal state of a user or item in the t-th snapshot; Indicates a gated loop unit; This indicates the state of the previous snapshot. This formula is used to encode the dependency between adjacent snapshots using GRU.

[0062] (3) Use a multi-head self-attention mechanism to obtain the attention weights of the time snapshot, and aggregate the time states after attention to obtain the temporal behavior representation.

[0063] Multi-head self-attention is performed on the obtained state sequence. Attention weights are generated based on the contribution of different historical time intervals to the prediction of future preferences. The states after attention are aggregated to obtain the temporal behavior representation of users and items.

[0064] 1) Perform multi-head self-attention on the obtained state sequence, and generate attention weights based on the contribution of different historical time intervals to future preference prediction, using the following formula: (8) in, and These represent the user state sequence and the item state sequence after GRU encoding, respectively; and These represent the attention weights of the user state sequence and the item state sequence at different time snapshots, respectively. These represent the user time state sequence and the item time state sequence after multi-head self-attention processing, respectively. and Representing users respectively and items In the Attention to the post-time state on a snapshot.

[0065] 2) The temporal state after aggregation attention is represented by the temporal behavior, as shown in the formula: (9) in, This represents the user and item representations output by the time-series dynamic channel; Indicates the number of time snapshots; This represents the t-th time state after attention. This formula is used to aggregate the time states after attention to obtain a temporal behavior representation.

[0066] Through the above steps, by dividing the complete interaction history into time snapshots, performing time-aware graph propagation within snapshots, and performing GRU and attention aggregation between snapshots, this invention can simultaneously preserve high-order collaborative structures within local time intervals and cross-interval preference transfers, avoiding the mixing of stage-based behavioral patterns by static folded graphs.

[0067] like Figure 1 As shown, in step S4, the semantic representation channel is used to perform edge reliability estimation and weighted graph propagation to obtain the user semantic preference representation and the item semantic preference representation.

[0068] S401. Based on the user set and the item set, perform semantic prior initialization on the offline large language model to obtain the initial semantic state of users and items.

[0069] (1) Generate user semantic priors and item semantic priors based on user set and item set.

[0070] The system organizes user history text chronologically and retrieves item titles, categories, or descriptions. It then generates and caches user semantic priors and item semantic priors offline. User semantic priors are generated offline from the user's chronologically ordered interaction history text, while item semantic priors are generated offline from the item description text.

[0071] Specifically, user history text and item description text are input into the large language model embedding interface via preset prompt templates to generate user semantic priors and item semantic priors, with the following formulas: (10) in, User interaction history text sorted by time. The semantic priors for the item title, category, and description text are cached offline.

[0072] (2) Through the two-layer adapter of user branch and item branch, the semantic prior of the offline large language model is mapped to the model representation space to obtain the initial semantic state of users and items.

[0073] A two-layer adapter is used to map the LLM semantic priors of users and items to the model space, as shown in the formula: (11) in, Represents the initial semantic state of users and items; This represents the semantic prior of an offline large language model; This represents the projection matrix of the two-layer adapter; This represents the bias of the two-layer adapter. This formula is used to map the user and item LLM semantic priors to the model space using a two-layer adapter.

[0074] After the above steps, the semantic prior is generated and cached offline before training and remains fixed during the training and inference phases. It does not require online invocation of large language models. When the embedding dimension, propagation depth and number of snapshots are fixed, the main computational cost increases approximately linearly with the number of nodes and the number of observed interactions. This solves the problem of the possible diffusion of preference-irrelevant or disturbed semantics.

[0075] S402. Based on the initial semantic state of users and items, perform edge reliability estimation on the semantic representation channel.

[0076] Edge reliability estimation is achieved through multilayer perceptron and Gumbel (binary concrete relaxation)-Softmax.

[0077] (1) Based on the semantic state of users and items, use a multilayer perceptron to calculate the edge compatibility score layer by layer and convert it into edge reliability weight.

[0078] During training, differentiable soft reliability weights are generated by combining logical noise, temperature parameters, and the Sigmoid function.

[0079] 1) For each user-item edge in the folded interaction graph, concatenate the user semantic state and the item semantic state from the previous propagation layer and input them into the multilayer perceptron to obtain the edge compatibility score, as shown in the formula: (12) in, Let $\frac{ ... Indicates the semantic state of the user and item at the next higher level; Represents a multilayer perceptron for edge reliability estimation; This represents a vector concatenation operation. This formula is used to calculate the layer-by-layer edge compatibility score based on the semantic state of the user and item.

[0080] 2) Use binary Concrete relaxation to convert the compatibility score into a reliable weight, using the following formula: (13) in, This represents the soft reliability weight of the user-item edge in layer l; Indicates edge-compatible fractions; This represents a random variable sampled uniformly between 0 and 1; Indicates the relaxation temperature of the binary Concrete; This represents the Sigmoid function. This formula is used to convert compatibility scores into compatibility reliability weights using binary Concrete relaxation.

[0081] (2) Perform normalized bidirectional semantic propagation of users and items according to the reliability weight to obtain the semantic state of users and items at each propagation layer.

[0082] The reliability weights of the software nodes are normalized by multiplying the square root of the degree of the user node by the degree of the item node. The normalized reliability weights are then used to aggregate the semantic states of neighboring items and neighboring users, respectively, using the following formulas: (14) in, Indicates the first Semantic states of layered users and items; Indicates edge soft reliability weights; Let represent the set of neighboring items for user u and the set of neighboring users for item i; This represents the cardinality of the corresponding neighbor set. This formula is used to perform normalized user-item bidirectional semantic propagation based on reliability weights.

[0083] S403. Aggregate the semantic states of each propagation layer across propagation layers to obtain a semantic preference representation.

[0084] The semantic states are aggregated across the propagation layer to obtain semantic preference representations, resulting in user semantic preference representations and item semantic preference representations, as shown in the formula: (15) in, This represents the user and item representations output by the semantic representation channel; Indicates the number of propagation layers in the semantic graph; Indicates the first Layer semantic states. This formula is used to aggregate semantic states across propagation layers to obtain a semantic preference representation.

[0085] After the above steps, weighted graph propagation is performed based on reliability weights to filter the semantic propagation of low-reliability interaction edges, and the semantic representation channels are output using SRC representation and item SRC representation. By estimating the soft reliability of edges based on the semantic states of users and items and performing weighted propagation accordingly, the propagation contribution of semantic mismatch relationships is reduced, making the general text semantics more consistent with personalized interaction evidence.

[0086] like Figure 1 As shown, in step S5, co-distillation and gated fusion are performed on the temporal representation and semantic preference representation to obtain the end-user representation and item representation.

[0087] S501. Construct a unique representation loss and impose feature loss constraints on user temporal representation, item temporal representation, user semantic preference representation, and item semantic preference representation.

[0088] Representation specialization is performed on user TDC representations, item TDC representations, user SRC representations, and item SRC representations, and the dual-channel representations are kept differentiated by specialization loss constraints. Specifically, the cosine similarity squared between the temporal behavior representation and the semantic preference representation is subject to representation specialization constraints, i.e., feature loss constraints, to reduce the directional redundancy of the original channel representations.

[0089] A unique representation loss is constructed to suppress excessive directional overlap between temporal and semantic channels, and the formula is as follows: (16) in, This represents the uniqueness loss. This represents the set of users and the set of deduplicated items in the current mini-batch; Indicates cosine similarity; This represents the semantic channel representation and the temporal channel representation. This formula is used to suppress excessive directional overlap between the temporal and semantic channels by using a uniqueness representation loss.

[0090] S502. Based on temporal and semantic representations respectively, generate temporal and semantic prediction score distributions, and use symmetric KL divergence for prediction-level bidirectional distillation.

[0091] Predictive-level bidirectional distillation is based on forming temperature softmax prediction distributions from the correlation scores of positive and negative candidates through two channels, and minimizing the symmetric Kullback-Leibler divergence between the two prediction distributions. The specific process is as follows: (1) Based on the user time-series representation and the item time-series representation, calculate the relevant scores of the time-series dynamic channel for positive and negative items.

[0092] The original relevance score of the semantic channel to positive and negative candidate items is calculated using the following formula: (17) in, This represents the raw scores of the semantic channel for positive and negative items; This represents positively interacting items and negatively sampled items; Indicates vector transpose; This represents the semantic channel representation. This formula is used to calculate the raw relevance score of the semantic channel to positive and negative candidate items.

[0093] (2) Based on user semantic representation and item semantic representation, calculate the relevant scores of semantic representation channels for positive and negative items.

[0094] The original correlation scores of time-series channels with respect to positive and negative candidate items can be expressed as: (18) in, This represents the raw score of the time-series channel for positive and negative items; This represents positively interacting items and negatively sampled items; Indicates vector transpose; This represents the temporal channel representation. This formula is used to calculate the raw correlation score of the temporal channel with respect to positive and negative candidate items.

[0095] (3) Use temperature Softmax to form a prediction distribution and perform prediction-level bidirectional distillation with symmetric Kullback-Leibler divergence.

[0096] 1) The positive and negative candidate prediction distributions for the two channels are obtained using temperature softmax, with the following formula: (19) in, Represents the positive and negative candidate prediction distributions for semantic and temporal channels; Indicates the predicted distillation temperature; This represents the original scores of the positive and negative candidates for the corresponding channel; This represents the normalized exponential function. This formula is used to obtain the positive and negative candidate prediction distributions for two channels using temperature softmax.

[0097] 2) By minimizing the symmetric KL divergence of the two prediction distributions, a prediction-level distillation loss is constructed to achieve prediction-level bidirectional distillation. The formula is as follows: (20) in, Indicates the predicted distillation loss; Indicates the Kullback-Leibler divergence; This represents the predicted distribution of the two channels; This represents the separator between the two divergence distributions. This formula is used to achieve predictive-level bidirectional distillation with symmetric KL divergence.

[0098] S503. Based on temporal representation and semantic representation respectively, construct a bidirectional projection bridge to form a bidirectional projection, and perform feature-level bidirectional distillation based on the mean square error between each projection representation and the corresponding target channel representation.

[0099] (1) Based on temporal representation and semantic representation respectively, a bidirectional projection bridge is constructed to form a bidirectional projection.

[0100] A bidirectional projection bridge is constructed based on temporal representation and semantic representation, respectively.

[0101] 1) Construct a bidirectional projection bridge, including a temporal-to-semantic projection network and a semantic-to-temporal projection network.

[0102] The bidirectional projection bridge comprises a temporal-to-semantic projection network and a semantic-to-temporal projection network, which respectively map the representation of one channel to the supervision space of the other channel. A bidirectional projection bridge is established using the temporal-to-semantic projection network and the semantic-to-temporal projection network, and feature-level bidirectional distillation is achieved using the mean square error between each projected representation and the corresponding target channel representation.

[0103] 2) A two-way projection representation is formed using dedicated projection networks in two directions, as shown in the formula: (twenty one) in, A projection representation from temporal to semantic direction; A projection representation representing semantics in the temporal direction; Represents a temporal-to-semantic projection network; This represents a semantic-to-temporal projection network. This formula is used to form a bidirectional projection representation through dedicated projection networks in two directions.

[0104] (2) By minimizing the mean square error between the two directional projection representations and the corresponding target channel representations, a feature-level distillation loss is constructed, and feature-level bidirectional distillation is performed.

[0105] Characteristic-level distillation of users and items is achieved by minimizing the bidirectional MSE constraint, as shown in the following formula: (twenty two) in, Indicates characteristic stage distillation loss; Indicates mean square error; This represents the set of users in a small batch and the set of deduplicated items; Representation of projections in two directions; This represents the original representation of the target channel. This formula is used to implement characteristic-level distillation of users and items with bidirectional MSE constraints.

[0106] By applying the above steps, first applying the representation specialization constraint, and then performing prediction-level symmetric KL distillation and feature-level bidirectional projective distillation, it is possible to transfer task-related knowledge while preserving the different modeling focuses of the two channels, and reduce redundancy and cross-channel interference caused by direct splicing, fixed averaging or forced global alignment.

[0107] S504. Through user-adaptive gating and item-adaptive gating, the distilled dual-channel representation is weighted and fused to obtain the final user representation and the final item representation.

[0108] The two-direction projection representations of the user and the two-direction projection representations of the item are concatenated and input into the gating layer to obtain the user gating vector and the item gating vector with values ​​between 0 and 1; the gating vector and its complement vector are used to perform element-wise weighting on the two projection representations.

[0109] (1) Generate the dimension-wise gating vectors of users and items based on the dual-channel projection representation.

[0110] The formula is: (twenty three) in, A d-dimensional gating vector representing a user or item; This represents the weight matrix of the user and item gating layers; Indicates the gate layer bias; This represents the element-wise Sigmoid function; This expression represents a vector concatenation operation. It is used to generate one-dimensional gating vectors for users and items based on a two-channel projection representation.

[0111] (2) Based on the dimension-wise gated vector and its complement vector, the two projection representations are weighted element-wise to obtain the final user representation and the final item representation.

[0112] Adaptive fusion of temporal and semantic projection representations based on instance and dimension, using the following formula: (twenty four) in, This represents the merged user and item representation; Represents the user and item gating vectors; This represents element-wise multiplication; This represents the complementary weights from semantic to temporal projection representations. This formula is used for adaptive fusion of temporal and semantic projection representations based on instance and dimension.

[0113] Through the above steps, and by implementing user-level, item-level, and dimension-wise adaptive gating, the recommendation system can adjust the contribution of temporal behavioral evidence and semantic evidence based on specific instances.

[0114] In this example, the average time-series channel threshold for users with limited historical data is approximately 0.04, while that for users with rich behavioral data is approximately 0.80, indicating that the gating is not a fixed global ratio.

[0115] like Figure 1 As shown, in step S6, the predicted score is calculated based on the end-user representation and the item representation to generate a Top-K unseen item recommendation list.

[0116] S601. Calculate the preference score based on the fused user representation and candidate item representation. After excluding items that have been interacted with during the training phase, sort them in descending order of score and output the Top-K unseen item recommendation list for the target user.

[0117] The preference score is calculated based on the merged user representation and candidate item representation, using the following formula: (25) in, This represents the predicted preference score for user u for item i.

[0118] Sort the items in descending order of their scores and extract the top K items to form a Top-K list of unseen items recommended for the target user.

[0119] S602. Construct the total loss function and train and optimize the model.

[0120] (1) Construct the total loss function.

[0121] 1) Construct the total loss of synergistic distillation based on the unique characterization loss, the predictive distillation loss, and the characteristic distillation loss.

[0122] (26) in, This represents the total loss from co-distillation; This indicates a unique loss representation; Indicates the predicted distillation loss; Indicates characteristic stage distillation loss; This represents the corresponding loss weight. This formula is used to combine specialization, prediction-level, and feature-level constraints to form a synergistic distillation loss.

[0123] 2) Construct BPR recommendation loss using the difference between positive and negative item preferences.

[0124] The formula is: (27) in, This represents the Bayesian personalized ranking loss; Represents the set of training triples for users, positive items, and negative items; This represents the prediction preference score for positive and negative items; This represents the Sigmoid function. This formula is used to construct the BPR recommendation loss using the difference between positive and negative item preferences.

[0125] 3) Construct a contrastive learning loss between users and positive and negative items within each channel.

[0126] The formula is: (28) in, This represents the contrastive learning loss for channel c; This represents the temporal dynamic channel (tdc) or the semantic representation channel (src). Indicates cosine similarity; Indicates the temperature for comparative learning; Represents the set of negatively sampled items for user u; This represents a positive interaction item. This formula is used to construct the contrastive learning loss between user-positive and negative items within each channel.

[0127] 4) The contrastive learning loss of the temporal dynamic channel and the semantic representation channel are fused to obtain the total contrastive learning loss.

[0128] The sum of the contrastive learning losses for temporal and semantic channels is given by the following formula: (29) in, This represents the total contrastive learning loss between the two channels; Indicates the time-series dynamic channel contrast loss; This represents the contrastive loss for the semantic representation channels. This formula is used to summarize the contrastive learning losses for the temporal and semantic channels.

[0129] 5) Construct the total loss function based on the total loss of co-distillation, the BPR recommendation loss, the total contrastive learning loss, and L2 regularization.

[0130] The overall training objective, formed by combinatorial recommendation, collaborative distillation, contrastive learning, and L2 regularization, can be expressed as: (30) in, Represents the total loss function of the model; This represents the BPR recommendation loss; This represents the total loss from co-distillation; This represents the total contrastive learning loss; This represents all trainable parameters; This represents the contrastive learning and L2 regularization weights; This represents the squared L2 norm of the parameter. This formula is used for combinatorial recommendation, collaborative distillation, contrastive learning, and L2 regularization to form the overall training objective.

[0131] (2) Train and optimize the model.

[0132] The time-series recommendation model based on dual-channel collaborative distillation is trained according to steps S1 to S6. During the training process, the model is iteratively updated using the total loss function.

[0133] (3) Case analysis.

[0134] Taking an implicit feedback scenario as an example, a time-series recommendation model based on dual-channel collaborative distillation is used for recommendation. This includes training, validation, and performance analysis.

[0135] 1) Each interaction record includes a user identifier, an item identifier, and a timestamp. For each user, the latest interaction is used as a test instance, the second newest interaction as a verification instance, and the remaining earlier interactions are used for training. During evaluation, all candidate items except those observed in the training set are sorted. This splitting and full sorting protocol is the specific method used in the experiment of this embodiment. The production system can generate training, verification, and inference data according to business rules, but must not change the core dual-graph dual-channel data stream of this embodiment.

[0136] Offline semantic priors can be generated from user text input arranged chronologically by user interaction history, and from item text input by item title, category, theme, or description. This embodiment uses text-embedding-ada-002 to generate vectors; the generated results are cached before training and remain constant during training and inference.

[0137] The model is built according to steps S1 to S6. The time range is divided into T non-overlapping windows; the initial representation of nodes in the snapshot adopts trainable user and item embeddings; for nodes without neighbors, formula (2) sets their neighbor messages as zero vectors; user nodes and item nodes adopt dual graph propagation operations.

[0138] After completing the joint training of the two channels, the co-distillation module, and the gated fusion module, a fused user representation is obtained for the target user, and a dot product preference score is calculated for unseen candidate items. The scores are sorted in descending order, and the top K items are truncated to form a Top-K unseen item recommendation list for the target user.

[0139] This embodiment uses PyTorch and a single NVIDIA GeForce RTX 3090 graphics card. The model embedding dimension is set to 64, the mini-batch size is set to 2048, and the AdamW optimizer is used. The learning rate is selected on the validation set from 10^-4, 10^-3, and 10^-2. The number of graph propagation layers is selected from 1, 2, and 3. The number of time snapshots is selected from 4, 6, 8, 10, 12, and 14. The early stopping patience value is 10 training epochs. The above parameters are within the validated preferred implementation range and should not be interpreted as the only limitation on the main scheme.

[0140] This embodiment was validated on three types of data: Yelp, Amazon Electronics, and ML-10M. The number of users were 10226, 15278, and 15491, respectively; the number of items were 25428, 23122, and 7181, respectively; and the number of interactions were 361075, 105844, and 1495411, respectively. The D2Rec Recall@20 were 0.0838, 0.0766, and 0.1442, respectively, and the NDCG@20 were 0.0335, 0.0363, and 0.0575, respectively, as shown in the figure.

[0141] Table 1. Recommendation results on the three datasets

[0142] 2) Overall performance analysis.

[0143] Table 1 summarizes the overall recommendation results of this embodiment on the Yelp, Amazon-Electronics, and ML-10M datasets. Evaluation was conducted using Recall@10, Recall@20, NDCG@10, and NDCG@20, and a full candidate ranking was performed on unseen candidate items while retaining the most recent user interaction as a test instance and the second newest interaction as a validation instance.

[0144] Based on the results listed in Table 1, within the given data split, candidate range, and comparison model range in this embodiment, D2Rec achieved the best listed values ​​in all four metrics across the three datasets; the improvement relative to the strongest comparison results ranged from 2.12% to 6.11%. These results are used to verify the effectiveness of the dual-graph, dual-channel, and synergistic distillation combination under the above conditions and are not intended as a general performance commitment detached from experimental conditions.

[0145] 3) Performance analysis under limited interaction history conditions.

[0146] To examine the complementary effect of the two channels when behavioral evidence is insufficient, users with at least 10 training interactions were selected, and 20% of them were randomly selected as target users. For each target user, only the most recent 1, 3, 5, or 10 training interactions were retained, while a Full condition for retaining the complete training history was set; validation and test interactions remained unchanged. User semantic priors were regenerated based on the retained training history, while item semantic priors remained unchanged.

[0147] like Figure 2 As shown, the overall Recall@20 and NDCG@20 of each method improve with the increase in the number of available training interactions. Under the conditions illustrated, the full D2Rec maintains high results across different history lengths; the variant with the semantic representation channel removed shows a more significant decrease under extremely short history conditions, indicating that semantic evidence provides supplementation when behavioral observations are insufficient. With increasing history length, the temporal dynamic channel can utilize richer temporal patterns, thus the two channels complement each other at different history scales.

[0148] 4) Robustness analysis under semantic prior perturbation conditions.

[0149] To examine the processing capability when semantic input is unreliable, a certain proportion of the semantic priors of the items are randomly permuted, with perturbation ratios of 0%, 10%, 20%, 30%, 40%, and 50% respectively. This permutation maintains the overall distribution of semantic vectors but disrupts the correspondence between some items and the original semantic priors; the interaction graph and the training, validation, and testing partitions remain unchanged, and each semantic enhancement method is retrained under the corresponding perturbation ratio.

[0150] like Figure 3 As shown in the figure, the retention rates of Recall@20 and NDCG@20 are reported based on the unperturbed condition. As the perturbation ratio increases, the overall retention rate of methods relying on semantic information decreases; the full D2Rec maintains a high overall retention rate across all three datasets. The variant that removes edge reliability weighting and performs direct static propagation shows a more significant decrease, indicating that soft edge reliability can reduce the propagation contribution of semantic mismatch relationships; preserving independent temporal dynamic channels also provides behavioral evidence unaffected by semantic permutations.

[0151] 5) Channel effect analysis under preference drift conditions.

[0152] To examine the channel effect as user preferences change over time, users were divided into a relatively stable preference group and a preference drift group based on the semantic preference differences between their early and recent training histories on the Yelp dataset. The complete model was then compared with variants that removed the temporal dynamic channel or the semantic representation channel.

[0153] like Figure 4 As shown, when the temporal dynamic channel was removed, the decrease in Recall@20 and NDCG@20 was significantly greater in the preference drift group than in the stable group. This observation corresponds to the processing flow of the temporal dynamic channel: when user interests change significantly, cross-time interval modeling is more important for describing preference shifts. Removing the semantic representation channel also caused a decrease in both groups, indicating that the reliable and refined semantic evidence provides supplementary information for users with different degrees of drift.

[0154] 6) Analysis of time attention mechanism.

[0155] To observe how the time-series dynamic channel aggregates across different time intervals, users were divided into three groups—low drift, medium drift, and high drift—based on their preference drift levels. The snapshot-level attention weights for each group of users and each attention head were then averaged.

[0156] like Figure 5As shown, the temporal attention distribution across the three datasets is not uniform, and the distribution differs among different drift groups. High-drift users in Yelp assign higher weights to several intermediate snapshots, Amazon-Electronics exhibits more concentrated weights within specific time intervals, and ML-10M shows a relatively smooth distribution. These results indicate that GRU and multi-head self-attention are not simply averaging historical snapshots, but rather aggregating cross-interval behavioral evidence based on temporal location and the degree of preference change.

[0157] 7) Analysis of edge reliability learning mechanism.

[0158] In the semantic prior perturbation experiment, user-item edges where the semantic prior of connected items is replaced are denoted as perturbed semantic edges, and the remaining edges are denoted as clean semantic edges. The distribution of soft reliability scores obtained from semantic representation channel learning is also statistically analyzed.

[0159] like Figure 6 As shown, on the three datasets, the reliability scores of clean semantic edges are generally higher than those of perturbed semantic edges. The two distributions still overlap, indicating that the mechanism does not mechanically classify edges into valid or invalid categories, but rather adjusts the proportion of semantic information propagating along the corresponding relations with continuous soft weights. Perturbed semantic edges receive relatively lower weights. Figure 3 The semantic perturbation preservation rate provides a corresponding explanation at the mechanistic level.

[0160] 8) Cross-channel representation coordination analysis.

[0161] To observe the relationship between the two representation spaces before and after co-distillation, linear center kernel alignment is used to measure the structural similarity between the temporal dynamic channel representation, the semantic representation channel representation, the projection representation in two directions, and the final fused representation.

[0162] like Figure 7 As shown, the original temporal representation and the original semantic representation exhibit only low or moderate similarity, consistent with the different objectives of modeling temporal behavioral evidence and semantic preference evidence, respectively, in the two channels. After bidirectional projection and feature-level distillation, the similarity between the projected representation and the corresponding target channel increases, while still maintaining a connection with the source channel; the final fused representation remains associated with both projected views. This result demonstrates that specialization constraints and bi-level distillation can reduce structural differences without completely homogenizing the two types of representations.

[0163] 9) Representative user case analysis.

[0164] To observe the phased preferences within time snapshots, a representative user from Yelp was selected, and the cosine similarity between the temporal dynamic channel representation of each time snapshot and the two category prototypes was calculated. For example... Figure 8As shown, the user's similarity to the two categories changes non-monotonically over time snapshots, indicating that snapshot-level modeling preserves phased behavioral patterns. Subsequent sequence aggregation processes state sequences that change over time, rather than a single static profile.

[0165] To observe how instance-by-instance gating allocates semantic and behavioral evidence, a further comparison is made between a user with limited historical data and a user with rich behavioral data. Figure 9 As shown, the average temporal dynamic channel weight for users with limited historical data is approximately 0.04, primarily relying on the semantic representation channel; the average temporal dynamic channel weight for users with rich behavioral data is approximately 0.80, and the threshold values ​​still vary across different embedding dimensions. This case illustrates that gating fusion targets specific users and allocates the contributions of the two channels according to dimensions, rather than using a fixed global proportion.

[0166] Example 2 The purpose of this embodiment is to provide a time-series recommendation system based on dual-channel synergistic distillation, including: The data acquisition module is used to acquire user sets, item sets, and interaction triples; The interaction graph construction module is used to construct temporal interaction graphs and folded interaction graphs based on interaction triples; The temporal dynamic channel module is used to input the temporal interaction graph into the temporal dynamic channel, and to perform interval division and timestamp-aware edge encoding on the temporal interaction graph in sequence to obtain the encoded temporal subgraph; and to perform local snapshot modeling and global sequence modeling on the encoded temporal subgraph in sequence to obtain the user temporal representation and the item temporal representation. The semantic representation channel module is used to perform edge reliability estimation and weighted graph propagation on the semantic representation channel to obtain user semantic preferences and item semantic preferences representations. The collaborative distillation and gating fusion module is used to perform collaborative distillation and gating fusion on temporal representations and semantic preference representations to obtain the final user representation and item representation; The prediction module is used to calculate prediction scores based on end-user and item representations, and generate a Top-K list of unseen items.

[0167] It also includes a model training optimization module, which trains a time-series recommendation model based on dual-channel collaborative distillation. During the training process, the model is iteratively updated using the constructed total loss function.

[0168] The method steps in Example 1 are implemented by providing a time-series recommendation system based on dual-channel collaborative distillation.

[0169] Example 3 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0170] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium.

[0171] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above-described method.

[0172] Example 5 The purpose of this embodiment is to provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods and functions involved in any of the above embodiments.

[0173] The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0174] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0175] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A time-series recommendation method based on dual-channel synergistic distillation, characterized in that, include: Retrieve the user set, item set, and interaction triplet; Construct a temporal interaction graph and a folding interaction graph based on interaction triples; The timing interaction graph is input into the timing dynamic channel, and the timing interaction graph is divided into intervals and timestamp-aware edge encoding is performed sequentially to obtain the encoded timing subgraph. Local snapshot modeling and global sequence modeling are performed sequentially on the encoded temporal subgraph to obtain user temporal representation and item temporal representation; By utilizing semantic representation channels, edge reliability estimation and weighted graph propagation are performed on the semantic representation channels to obtain representations of user semantic preferences and item semantic preferences; Cooperative distillation and gated fusion are performed on temporal representations and semantic preference representations to obtain end-user representations and item representations; Calculate the predicted score based on the end-user representation and item representation, and generate a Top-K list of unseen items.

2. The time-series recommendation method based on dual-channel synergistic distillation as described in claim 1, characterized in that, The temporal interaction graph is divided into intervals and timestamp-aware edge encoding is performed sequentially to obtain an encoded temporal subgraph. Specifically, the temporal interaction graph is divided into several temporal subgraphs according to time intervals; the interaction edges of the temporal subgraphs are encoded with timestamp-aware edges to obtain the encoded temporal subgraph.

3. The time-series recommendation method based on dual-channel synergistic distillation as described in claim 1, characterized in that, Local snapshot modeling is performed on the coding time-series subgraph. The specific process is as follows: The time series subgraph is divided into several time windows based on the timestamp, resulting in several local snapshot images; By using the dimension-wise maximum aggregation operation, the time-encoded neighbor information in the local snapshot image is aggregated to obtain the time-aware neighbor information of users and items respectively. Based on the time-aware neighbor information of users and items, a multilayer perceptron is used to stitch together and update the local time snapshot map to obtain the local representation of users and the local representation of items.

4. The time-series recommendation method based on dual-channel synergistic distillation as described in claim 1, characterized in that, Global sequence modeling is performed on the local representations of users and items to obtain the temporal representations of users and items, respectively. The specific process is as follows: Based on the local representations of users and items, a sequence of interval-level representations of users and items is formed; The interval-level representation sequences of users and items are input into the gated loop unit for encoding in chronological order to obtain the dependency relationships between adjacent snapshots of users and items. A multi-head self-attention mechanism is used to obtain the attention weights of the time snapshot, and the time states after attention are aggregated to obtain the temporal behavior representations of users and items respectively.

5. The time-series recommendation method based on dual-channel synergistic distillation as described in claim 1, characterized in that, Edge reliability estimation and weighted graph propagation are performed on the semantic representation channels to obtain user semantic preference representations and item semantic preference representations. The specific process is as follows: Based on user sets and item sets, semantic prior initialization is performed on the offline large language model to obtain the initial semantic state of users and items; Based on the initial semantic state of users and items, multilayer perceptron is used to calculate the layer-by-layer edge compatibility score and convert it into edge reliability weights; Perform normalized bidirectional semantic propagation of users and items based on reliability weights to obtain the semantic states of users and items at each propagation layer. The semantic states of users and items at each propagation layer are aggregated across propagation layers to obtain user semantic preference representations and item semantic preference representations.

6. The time-series recommendation method based on dual-channel synergistic distillation as described in claim 1, characterized in that, Cooperative distillation and gated fusion are performed on the temporal representation and semantic preference representation to obtain the final user representation and item representation. The specific process is as follows: Construct a unique representation loss and impose feature loss constraints on user temporal representation, item temporal representation, user semantic preference representation, and item semantic preference representation; Based on temporal and semantic representations respectively, temporal and semantic prediction score distributions are generated, and prediction-level bidirectional distillation is performed using symmetric KL divergence. Based on temporal and semantic representations respectively, a bidirectional projection bridge is constructed to form a bidirectional projection, and feature-level bidirectional distillation is performed using the mean square error between each projection representation and the corresponding target channel representation. By using user-adaptive gating and item-adaptive gating, the distilled dual-channel representations are weighted and fused to obtain the final user representation and the final item representation.

7. The time-series recommendation method based on dual-channel synergistic distillation as described in claim 6, characterized in that, Based on temporal and semantic representations respectively, temporal and semantic prediction score distributions are generated, and bidirectional distillation at the prediction level is performed using symmetric KL divergence. The specific process is as follows: Based on user time-series representation and item time-series representation, calculate the relevant scores of the time-series dynamic channel for positive and negative items; Based on user semantic representation and item semantic representation, calculate the relevant scores of semantic representation channels for positive and negative items; Based on the correlation scores of the temporal dynamic channel and the semantic representation channel, the positive and negative candidate prediction distributions of the two channels are obtained using temperature softmax. Predictive-level bidirectional distillation is performed to minimize the symmetric KL divergence between the two prediction distributions.

8. The time-series recommendation method based on dual-channel synergistic distillation as described in claim 6, characterized in that, A bidirectional projection bridge is constructed to form a bidirectional projection, which includes a temporal-to-semantic projection network and a semantic-to-temporal projection network. The bidirectional projection representation is formed through two dedicated projection networks in two directions.

9. The time-series recommendation method based on dual-channel synergistic distillation as described in claim 1, characterized in that, During training, a total loss function is constructed, and the model is trained and optimized. The formula for the total loss function is: ； in, Represents the total loss function of the model; This represents the BPR recommendation loss; This represents the total loss from co-distillation; This represents the total contrastive learning loss; This represents all trainable parameters; This represents the contrastive learning and L2 regularization weights; This represents the squared L2 norm of the parameter.

10. A time-series recommendation system based on dual-channel collaborative distillation, characterized in that, include: The data acquisition module is used to acquire user sets, item sets, and interaction triples; The interaction graph construction module is used to construct temporal interaction graphs and folded interaction graphs based on interaction triples; The timing dynamic channel module is used to input the timing interaction graph into the timing dynamic channel, and to perform interval division and timestamp-aware edge encoding on the timing interaction graph in sequence to obtain the encoded timing subgraph; Local snapshot modeling and global sequence modeling are performed sequentially on the encoded temporal subgraph to obtain user temporal representation and item temporal representation; The semantic representation channel module is used to perform edge reliability estimation and weighted graph propagation on the semantic representation channel to obtain user semantic preferences and item semantic preferences representations. The collaborative distillation and gating fusion module is used to perform collaborative distillation and gating fusion on temporal representations and semantic preference representations to obtain the final user representation and item representation; The prediction module is used to calculate prediction scores based on end-user and item representations, and generate a Top-K list of unseen items.

Images