Deep Learning-Based Approach to Enhance Interest Graph Modeling for User Profiles

Abstract

This invention discloses a deep learning-based method for enhancing the modeling of user profile interest graphs. The method comprises the following steps: S1. Generating a sequence of snapshots of user profile interest graphs; S2. Obtaining a set of performance evaluation results; S3. Selecting the optimal pooling hyperparameter configuration based on the fitness evaluation results; S4. Outputting a set of hierarchical vectors representing time windows; S5. Concatenating the hierarchical vectors of time windows according to the temporal order of the snapshot sequence to obtain an interest fusion representation vector, and constructing an offline teacher model; S6. Training the simplified lightweight student model using knowledge distillation, deploying the lightweight student model on an online real-time recommendation system, and outputting inference results under real-time traffic scenarios; S7. When a key performance indicator is detected to be below a preset threshold or a change in user interest distribution is detected to be greater than a preset threshold, steps S2 to S6 are re-executed. This invention effectively improves the ability to identify complex behavioral patterns of cold-start users and users with shifting interests.

Description

Technical Field

[0001] This invention relates to the field of interest graph technology, and in particular to a method for enhancing the modeling of interest graphs for user profiles based on deep learning. Background Technology

[0002] With the development of big data and artificial intelligence technologies, user profiling and interest modeling have gradually become key components of intelligent recommendation systems. In particular, in user profiling, extracting user interest tags and behavioral patterns from multi-dimensional behavioral data has been widely applied in e-commerce recommendations, content push, and advertising targeting scenarios. Current mainstream methods usually rely on static graph or sequence modeling techniques, using user behavior logs to construct graph structures or temporal feature vectors, and using graph neural networks or temporal neural networks to predict and represent user interests.

[0003] However, existing technologies still have significant shortcomings in handling the dynamics and structural heterogeneity of user interests. First, most graph modeling methods only construct static graphs based on the binary relationship between users and interest tags, lacking the integration of contextual information, resulting in a lack of context-awareness in interest relationship modeling and failing to effectively reflect the differences in user behavior in different scenarios. Second, current graph neural networks are easily affected by node degree imbalance and redundant edges when processing large-scale graph data, lacking efficient and selective graph pruning mechanisms, leading to low training efficiency and poor model generalization ability. In addition, when dealing with the problem of graph structure evolution, existing methods usually simply splice or independently model graph structures under different time windows, ignoring the evolutionary laws of graph structures in the time dimension, making it difficult to capture the continuity and abrupt changes in interest evolution.

[0004] In addition, most existing graph neural network models are offline models with complex structures and large parameters. Although they have strong representation capabilities, they are difficult to deploy in online real-time recommendation systems, resulting in large response latency and poor model adaptability during online inference. Although some attempts to introduce model compression and knowledge distillation have made breakthroughs, they lack high-quality teacher model construction mechanisms and dynamic adjustment strategies. The distilled model still has a significant lag when dealing with sudden changes in user interests.

[0005] Therefore, there is an urgent need for a method to enhance the modeling of user interest graphs that can integrate heterogeneous structural information, have dynamic pruning capabilities, and take into account both temporal evolution and real-time performance, so as to improve the overall performance of recommendation systems. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method for enhancing the modeling of user profile interest graphs. This invention effectively improves the ability to identify complex behavioral patterns of cold-start users and users with shifting interests.

[0007] A method for enhancing the modeling of user profile interest graphs based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Construct a ternary heterogeneous graph and generate a snapshot sequence of the interest map of the population profile based on the ternary heterogeneous graph using a multi-granularity time window;

[0009] S2. For each snapshot in the interest graph snapshot sequence of the user profile, perform heuristic pruning based on node degree threshold, semantic similarity threshold and time decay factor to generate corresponding candidate subgraphs. Gather the candidate subgraphs of all graph snapshots to form a candidate subgraph set. Perform offline task model training on the candidate subgraph set on the training dataset, calculate the click-through rate evaluation result, transaction conversion evaluation result and normalized depreciation evaluation result to obtain the performance evaluation result set;

[0010] S3. Taking the candidate subgraph set and the performance evaluation result set as input, define the search space and fitness function of the artificial bee colony optimization algorithm, initialize the artificial bee colony optimization algorithm, iteratively search the pooling hyperparameter candidate solution set through the discovery bee stage, follower bee stage and scout bee stage of the artificial bee colony optimization algorithm, and evaluate the fitness of the pooling hyperparameter candidate solution set based on the fitness function, and select the optimal pooling hyperparameter configuration according to the fitness evaluation result;

[0011] S4. Under the optimal pooling hyperparameter configuration, for each snapshot in the interest graph snapshot sequence of the population profile and its corresponding candidate subgraph, execute a hierarchical graph neural network model composed of the DiffPool module and the graph convolutional network module. The hierarchical graph neural network model is used to extract the structural embedding vector of each snapshot and output a set of hierarchical structural vectors for the time window.

[0012] S5. The hierarchical structure vector set of time windows is concatenated according to the time order of the graph snapshot sequence and fused using a sparse attention mechanism to obtain the interest fusion representation vector for downstream supervised training. The interest fusion representation vector is used as input to train and construct an offline teacher model for offline supervised learning. The offline teacher model is a high-capacity neural network structure.

[0013] S6. Based on the output results and parameters of the offline teacher model, knowledge distillation training is performed on the simplified lightweight student model. The lightweight student model is used to approximate the behavior of the teacher model without relying on the complete graph structure, forming a lightweight student model that can be deployed online. The lightweight student model is then deployed in an online real-time recommendation system and outputs inference results in real-time traffic scenarios.

[0014] S7. Continuously monitor the key performance indicators of the online real-time recommendation system. When the key performance indicators are detected to be lower than the preset threshold or the user interest distribution changes to a value greater than the preset threshold, collect the latest user data and regenerate the snapshot sequence of the user profile interest graph. Then, repeat steps S2 to S6 to achieve enhanced modeling of the user profile interest graph structure vector.

[0015] Optionally, step S1 includes the following steps:

[0016] S11. Collect raw behavior log data to form a raw behavior log set. Each user behavior record in the raw behavior log set includes four fields: user identifier, interest action tag, behavior timestamp, and behavior context scene attribute. Each user behavior record in the raw behavior log set is used to describe a user's interest-related behavior triggered at a specific time and in a specific scene.

[0017] S12. Based on the original behavior log set, extract the user identifier set, interest tag set, and context scene set respectively. Jointly construct a ternary heterogeneous graph using the user identifier set, interest tag set, and context scene set. The nodes of the ternary heterogeneous graph consist of user nodes, interest tag nodes, and context scene nodes. In the ternary heterogeneous graph, the connection edge between the user node and the interest tag node represents the relationship between the user's behavior and the interest tag, and the connection edge between the user node and the context scene node represents the context scene information when the user's behavior occurs. The interest tag node V... (k) E, the connection edge between the context scene node (k) This is used to represent the co-occurrence relationship of interest behaviors in a specific context. Each connection edge is associated with a connection strength parameter. The connection strength parameter between user nodes and interest tag nodes is used to characterize the weighted result of the frequency and freshness of user behavior on a certain interest tag. The connection strength parameter between user nodes and context scene nodes is used to characterize the expression strength of user activity in different context scenes. The connection strength parameter between interest tag nodes and context scene nodes is used to characterize the frequency of co-occurrence of interest tags in different context scenes.

[0018] S13. Set a multi-granularity time window set to divide the time dimension of user behavior. For each time window, extract the structural subgraph within the time period from the ternary heterogeneous graph to generate a graph snapshot sequence. Each graph snapshot in the graph snapshot sequence consists of the set of valid connection edges within the corresponding time period and the set of nodes involved in the connection edges. Each graph snapshot contains only user nodes, interest tag nodes, and context scene nodes that have performed actions within that time period, as well as all connection edges formed between nodes within the time period.

[0019] S14. For each graph snapshot in the graph snapshot sequence, record its time window label, number of active nodes, and edge density. The number of active nodes in the graph snapshot is used to represent the total number of nodes that participated in at least one action within that time period. The edge density is used to measure the density of the structure in the graph snapshot. The edge density is obtained by dividing the number of connecting edges in the graph snapshot by the product of the number of nodes in the graph snapshot and the number of nodes minus one. The time window label, number of active nodes, and edge density of the graph snapshot are defined as the graph snapshot sequence of the interest graph of the population profile.

[0020] Optionally, step S2 includes the following steps:

[0021] S21. For each snapshot G in the sequence of snapshots of the interest graph of the user profile. (k) =(V (k) E (k) ), calculate the node degree value of each node and set the node degree threshold θ. d Remove all sets of nodes whose degree value is lower than the node degree threshold. Define node degree value

[0022]

[0023] in, Represents node v i In the snapshot G (k) The weighted degree value in the middle, Represents node v i With node v j In the time window Δt k Connection strength parameter of inner connecting edge, 1 {·} This is an indicator function; it is 1 if the edge exists, and 0 otherwise.

[0024] S22. Based on graph snapshot G (k) Interest semantic vector representation of each node Construct the node semantic similarity matrix S (k) A semantic similarity threshold is set, and edges corresponding to all node pairs with semantic similarity below the threshold are filtered out, resulting in node v. i With node v j In the snapshot G (k) semantic similarity Semantic similarity is used to characterize the degree of similarity between two nodes in their interest semantic vector representations. Semantic similarity is obtained by taking the dot product of the interest semantic vector representations of the two nodes and then dividing by the product of the magnitudes of the two vectors.

[0025] S23. For each graph snapshot G (k) Combined with behavior timestamp t iWith time window Δt k Calculate the node activity decay weight factor The node activity decay weight factor is used to reflect the timeliness of node behavior relative to the current system time. It is obtained by dividing the difference between the current system time and the time of the node's most recent behavior by the duration of the time window corresponding to the graph snapshot, and then multiplying by the time decay control coefficient and taking the negative exponent.

[0026] S24. Overall node degree value semantic similarity Weighting factor for node activity decay Constructing a candidate subgraph pruning function:

[0027]

[0028] in, Represents node v i Satisfy the clipping retention condition, θ d θ s θ γ This indicates the corresponding threshold.

[0029] S25. Snapshot G (k) All candidate subgraphs that satisfy the clipping function Node set and its associated edge set Constructing candidate subgraphs A candidate subgraph set is formed by combining candidate subgraphs from all time windows:

[0030]

[0031] S26. On the training dataset, select the candidate subgraph set G. sub The objective function is trained using the click-through rate (AUC) metric. (k) Transaction conversion rate (CVR) (k) With structural fidelity index Dist (k) For each candidate subgraph Perform performance evaluation and construct a performance evaluation result set M. perf :

[0032]

[0033] Among them, AUC (k) Represents a snapshot of the graph G (k) The area under the curve (AUC) of the corresponding candidate subimage in the click prediction task, CVR (k) Represents a snapshot of the graph G (k) Conversion rate performance in transaction conversion prediction tasks, Dist (k)This represents the fidelity measure between the candidate subgraph structure and the original snapshot structure.

[0034] Optionally, step S3 includes the following steps:

[0035] S31. Using the candidate subgraph set G sub With the performance evaluation result set M perf As input, construct the hyperparameter search space for the artificial bee colony optimization algorithm, where each solution vector x represents a set of pooling hyperparameter configurations:

[0036] x = (p, m, λ, α);

[0037] Where p represents the graph pooling ratio hyperparameter in the DiffPool module, m represents the number of cluster centers after each graph pooling layer, λ represents the graph structure preservation regularization coefficient, which controls the structural constraint strength between the original graph and the pooled graph, and α represents the graph attention gating coefficient, which is used to adjust the information fusion weights.

[0038] S32. Define the fitness function f(x) of the artificial bee colony optimization algorithm. The fitness function is used to measure the overall effect of each pooling hyperparameter configuration on the candidate subgraph set. Construct the objective function based on the evaluation index in S26:

[0039]

[0040] Among them, AUC (k) (x), CVR (k) (x) and Dist (k) (x) represents the k-th candidate subgraph. Above, the area under the click-through rate curve score, transaction conversion rate score, and structure fidelity metric obtained after applying pooling hyperparameter configuration x, with w1, w2, and w3 as weighting coefficients;

[0041] S33. Initialize the set of individual bee colonies X = {x1, x2, ..., x...} for the artificial bee colony optimization algorithm. n}, where n is the initial bee colony size, and x is the number of individual bees in the colony. i Let f(x) represent a set of candidate pooling hyperparameter configurations and compute the fitness function f(x) for each individual in the bee colony. i );

[0042] S34. In the bee discovery phase of the artificial bee colony optimization algorithm, for each individual bee x i Execute parameter perturbation to generate neighborhood solution x′ i And evaluate the fitness of the neighborhood solutions. If the fitness function of the neighborhood solutions is f(x′) i )>f(x i If x′ = 0, then the neighborhood solution x′ is used. i Replace individual bee in the colony xi ;

[0043] S35. In the follow-bee stage of the artificial bee colony optimization algorithm, a probabilistic selection strategy is constructed based on the fitness distribution of the current bee colony individual set. Some individuals are selected for local search and perturbation in a probabilistic manner, the solution space is updated and the fitness function is recalculated.

[0044] S36. In the scout bee phase of the artificial bee colony optimization algorithm, bee individuals whose fitness function has not improved within a preset time are randomly reinitialized to form a new solution vector x. r Add them to the bee colony individual set X, and perform fitness function calculation;

[0045] S37. Set the maximum number of iterations or the optimal solution convergence condition, execute the artificial bee colony optimization iterative process, and output the optimal pooling hyperparameter configuration of the fitness function when the termination condition is met.

[0046] Optionally, step S4 includes the following steps:

[0047] S41. Using the optimal pooling hyperparameter configuration as input configuration parameters, for each graph snapshot G in the sequence of graph snapshots of the interest graph of the user profile, ... (k) and its corresponding candidate subgraph Construct a structured adaptive graph convolutional network module that incorporates a multi-head attention mechanism, and use node neighborhood interest to aggregate weights. Update the interest embedding vector for each node

[0048]

[0049] in, The node v in the k-th graph snapshot after convolution of the l-th layer graph represents the node v. i Interest embedding vector, N(v) i ) represents node v i The set of first-order interest neighborhood nodes, W (l) Let σ(·) be the learnable feature transformation weight matrix of the l-th layer, where σ(·) represents the nonlinear activation function and σ(·) represents the node neighborhood interest aggregation weight. Represents node v j For node v i The neighborhood contribution of interest representation is obtained by jointly calculating the semantic similarity of node interest and the edge connection strength parameter;

[0050] S42. Concatenate the DiffPool module with the structure-adaptive graph convolutional network module to form a hierarchical graph neural network model. In the l-th layer of DiffPool pooling, construct a soft clustering matrix S that adapts to the criticality of interest nodes in the user group. (k,l)The soft cluster assignment matrix represents the probability that each interest node belongs to each interest cluster. The assignment probability is generated through mapping by a multilayer perceptron network. The graph attention gating coefficient α is introduced into the multilayer perceptron network. * This is used to adjust the influence of each interest node on the clustering decision. The number of interest clusters in each layer is determined by the graph pooling ratio hyperparameter p. * The control is the integer part of the product of the number of interest nodes in the current layer and the graph pooling ratio;

[0051] S43. Based on the soft clustering matrix generated in each layer of the graph pooling module, perform adaptive interest aggregation on the node interest embedding vectors. Adaptive interest aggregation obtains the l-th layer node interest embedding vector X by weighting and summing the interest embedding vectors of all interest nodes according to their corresponding clustering probabilities. (k,l) The adjacency structure relationships of the original graph are mapped according to the assigned probabilities to obtain the graph structure vector A of the l-th layer interest clusters. (k,l) ;

[0052]

[0053] S44. Construct an interest structure fidelity loss function in each graph pooling module. The interest structure fidelity loss function is used to measure the degree of structural deviation between the original graph adjacency structure and the compressed interest cluster graph.

[0054]

[0055] in, This is used to measure the deviation between the interest graph structure after pooling in the l-th layer of DiffPool and the interest graph structure before pooling. The interest structure fidelity loss function guides the DiffPool module to adaptively retain key node and interest cluster information.

[0056] S45. In the top-level interest cluster graph representation of the hierarchical graph neural network model, a cross-layer residual connection mechanism is used to fuse the interest cluster embedding vectors of different layers. The fusion process involves multiplying the node interest embedding vectors output by each layer by the learnable residual connection weights and then weighting and stacking them layer by layer. A global average pooling operation is then used to summarize all node interest embedding vectors to obtain the high-order interest graph embedding vector h of the population profile interest graph corresponding to the graph snapshot. (k) ;

[0057] S46. High-order interest graph embedding vector h that aggregates all time window snapshots (k) Construct a hierarchical structure vector set H of time windows for subsequent interest representation fusion and supervised learning. snap .

[0058] Optionally, step S5 includes the following steps:

[0059] S51. The high-order interest graph embedding vectors corresponding to each snapshot in the time window hierarchical structure vector set are concatenated in chronological order to form a time series interest vector matrix. The time series interest vector matrix is ​​used to preserve the evolution trajectory of the interest graph structure of the population profile under different time windows. Each row represents a high-order interest graph embedding vector of a time snapshot, and each column represents an interest embedding dimension.

[0060] S52. Construct a sparse attention fusion module to process the time series interest vector matrix H. seq Attention weights are assigned by embedding snapshots of each image, resulting in a time-weighted interest fusion representation vector h. agg :

[0061]

[0062] Where, α k This represents the attention weight embedded in the k-th time snapshot;

[0063] S53. Use the interest fusion representation vector generated by the sparse attention fusion module as the training input to construct an offline supervised learning model. The offline supervised learning model is defined as an offline teacher model. The offline teacher model is a multi-layer neural network with a high-capacity structure. Its input is the interest fusion representation vector and its output is the prediction result, which is used to restore the recommendation task results of users in specific interest behavior scenarios.

[0064] S54. Based on the true label value of each sample in the training dataset and the prediction result of the offline teacher model, a supervised loss function is constructed. The supervised loss function is used to measure the degree of difference between the predicted value and the true label. The supervised loss function uses the form of binary cross-entropy to calculate the logarithmic error between the predicted value and the label value at the sample level, and then averages it over all samples. The smaller the supervised loss function, the stronger the prediction ability of the offline teacher model.

[0065] S55. The gradient descent optimization algorithm is used to minimize the supervised loss function and iteratively optimize the parameter set of the offline teacher model. The training process ends when the supervised loss function reaches its minimum value or the termination condition is met. The obtained model parameter set is used to fix the state of the offline teacher model.

[0066] Optionally, step S6 includes the following steps:

[0067] S61. Construct a lightweight student model with a simplified structure. The lightweight student model is a deep compressed graph neural network structure. Its feature expression dimension and network depth are both smaller than those of the offline teacher model. The input is a set of sparse node-level features of candidate subgraphs or a set of low-dimensional interest representations. It does not depend on the complete graph structure information.

[0068] S62. Extract the output distribution of each training sample from the offline teacher model as a soft label. Simultaneously, combining the true labels of each sample Constructing a knowledge distillation loss function The knowledge distillation loss function consists of two parts: the Kullback-Leibler divergence loss between the soft labels and the student model's predictions, and the cross-entropy loss between the true labels and the student model's predictions.

[0069]

[0070] in, Let represent the predicted output of the i-th sample calculated by the lightweight student model, β be the weight hyperparameter of the distillation loss, KL(·||·) represent the Kullback-Leibler divergence, and CE(·,·) represent the cross-entropy loss function.

[0071] S63. Utilizing the knowledge distillation loss function Optimize the parameter set Θ of the lightweight student model student Iteratively minimize using backpropagation and gradient descent Obtain the lightweight student model structure after distillation;

[0072] S64. Deploy the lightweight student model trained by distillation to the online real-time recommendation system, build an inference interface to accept user request data from real-time traffic scenarios, and input data includes user current context scene features, user historical behavior interest tag sequence and timestamp compression features. After calculation by the lightweight student model, the recommendation task results are output.

[0073] S65. In real-time traffic scenarios, the recommendation score is based on the output of the lightweight student model. Based on preset classification decision rules, determine the recommended status label for the user's current behavior:

[0074] If recommended score If the user behavior context is a cold start environment, then the current user state is marked as having unknown interests;

[0075] If recommended score Furthermore, if there are sparse jumps in time in the historical interest tag sequence, then the user's state is marked as an interest change;

[0076] If recommended score If the similarity between the current interest tag and the historical tags exceeds the threshold, then the user's status is marked as stable in terms of interests.

[0077] The beneficial effects of this invention are:

[0078] (1) This invention introduces an artificial bee colony optimization algorithm to dynamically select pooling hyperparameters in a graph neural network, thereby achieving joint optimization of graph pooling ratio, number of interest clusters, structure preservation coefficient and attention gating coefficient in the DiffPool module. By constructing a fitness function with prediction accuracy and structure fidelity as its core, the optimal combination of hyperparameters is accurately identified by utilizing the global search capability of the artificial bee colony in the solution space.

[0079] (2) In the graph structure modeling stage, this invention introduces a heuristic multi-factor pruning mechanism, which integrates the three-dimensional features of node degree, semantic similarity and temporal activity to extract candidate subgraphs from the graph snapshot. This avoids the computational redundancy and information dilution problems caused by full graph computation in traditional graph neural networks. Through the node retention function, it realizes the automatic selection of a subset of nodes with high information density, active behavior and strong semantic concentration in the graph snapshot to form a set of candidate subgraphs, effectively compressing the graph size while retaining the core semantic structure.

[0080] (3) This invention designs an interest graph vector integration mechanism that integrates hierarchical graph structure and time evolution expression. Combined with the sparse attention fusion module, the high-order interest embedding vector sequence generated under multi-granularity time windows is weighted and integrated to construct an interest fusion representation vector for offline supervised learning teacher model training. Furthermore, it is transferred to a lightweight student model through knowledge distillation for deployment in real-time recommendation system. It can significantly capture the nonlinear change characteristics of user interests in the time dimension and effectively improve the ability to identify complex behavioral patterns of cold start users and interest-switching users. Attached Figure Description

[0081] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0082] Figure 1 This is a flowchart of a deep learning-based method for enhancing the modeling of interest graphs in user profiles, as proposed in this invention. Detailed Implementation

[0083] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0084] refer to Figure 1 A deep learning-based method for enhancing the modeling of user profile interest graphs includes the following steps:

[0085] S1. Construct a ternary heterogeneous graph and generate a snapshot sequence of the interest map of the population profile based on the ternary heterogeneous graph using a multi-granularity time window;

[0086] S2. For each snapshot in the interest graph snapshot sequence of the user profile, perform heuristic pruning based on node degree threshold, semantic similarity threshold and time decay factor to generate corresponding candidate subgraphs. Gather the candidate subgraphs of all graph snapshots to form a candidate subgraph set. Perform offline task model training on the candidate subgraph set on the training dataset, calculate the click-through rate evaluation result, transaction conversion evaluation result and normalized depreciation evaluation result to obtain the performance evaluation result set;

[0087] S3. Taking the candidate subgraph set and the performance evaluation result set as input, define the search space and fitness function of the artificial bee colony optimization algorithm, initialize the artificial bee colony optimization algorithm, iteratively search the pooling hyperparameter candidate solution set through the discovery bee stage, follower bee stage and scout bee stage of the artificial bee colony optimization algorithm, and evaluate the fitness of the pooling hyperparameter candidate solution set based on the fitness function, and select the optimal pooling hyperparameter configuration according to the fitness evaluation result;

[0088] S4. Under the optimal pooling hyperparameter configuration, for each snapshot in the interest graph snapshot sequence of the population profile and its corresponding candidate subgraph, execute a hierarchical graph neural network model composed of the DiffPool module and the graph convolutional network module. The hierarchical graph neural network model is used to extract the structural embedding vector of each snapshot and output a set of hierarchical structural vectors for the time window.

[0089] S5. The hierarchical structure vector set of time windows is concatenated according to the time order of the graph snapshot sequence and fused using a sparse attention mechanism to obtain the interest fusion representation vector for downstream supervised training. The interest fusion representation vector is used as input to train and construct an offline teacher model for offline supervised learning. The offline teacher model is a high-capacity neural network structure.

[0090] S6. Based on the output results and parameters of the offline teacher model, knowledge distillation training is performed on the simplified lightweight student model. The lightweight student model is used to approximate the behavior of the teacher model without relying on the complete graph structure, forming a lightweight student model that can be deployed online. The lightweight student model is then deployed in an online real-time recommendation system and outputs inference results in real-time traffic scenarios.

[0091] S7. Continuously monitor the key performance indicators of the online real-time recommendation system. When the key performance indicators are detected to be lower than the preset threshold or the user interest distribution changes to a value greater than the preset threshold, collect the latest user data and regenerate the snapshot sequence of the user profile interest graph. Then, repeat steps S2 to S6 to achieve enhanced modeling of the user profile interest graph structure vector.

[0092] In this embodiment, step S1 includes the following steps:

[0093] S11. Collect raw behavior log data to form a raw behavior log set. Each user behavior record in the raw behavior log set includes four fields: user identifier, interest action tag, behavior timestamp, and behavior context scene attribute. Each user behavior record in the raw behavior log set is used to describe a user's interest-related behavior triggered at a specific time and in a specific scene.

[0094] S12. Based on the original behavior log set, extract the user identifier set, interest tag set, and context scene set respectively. Jointly construct a ternary heterogeneous graph using the user identifier set, interest tag set, and context scene set. The nodes of the ternary heterogeneous graph consist of user nodes, interest tag nodes, and context scene nodes. In the ternary heterogeneous graph, the connection edge between the user node and the interest tag node represents the relationship between the user's behavior and the interest tag, and the connection edge between the user node and the context scene node represents the context scene information when the user's behavior occurs. The interest tag node V... (k) E, the connection edge between the context scene node (k) This is used to represent the co-occurrence relationship of interest behaviors in a specific context. Each connection edge is associated with a connection strength parameter. The connection strength parameter between user nodes and interest tag nodes is used to characterize the weighted result of the frequency and freshness of user behavior on a certain interest tag. The connection strength parameter between user nodes and context scene nodes is used to characterize the expression strength of user activity in different context scenes. The connection strength parameter between interest tag nodes and context scene nodes is used to characterize the frequency of co-occurrence of interest tags in different context scenes.

[0095] S13. Set a multi-granularity time window set to divide the time dimension of user behavior. For each time window, extract the structural subgraph within the time period from the ternary heterogeneous graph to generate a graph snapshot sequence. Each graph snapshot in the graph snapshot sequence consists of the set of valid connection edges within the corresponding time period and the set of nodes involved in the connection edges. Each graph snapshot contains only user nodes, interest tag nodes, and context scene nodes that have performed actions within that time period, as well as all connection edges formed between nodes within the time period.

[0096] S14. For each graph snapshot in the graph snapshot sequence, record its time window label, number of active nodes, and edge density. The number of active nodes in the graph snapshot is used to represent the total number of nodes that participated in at least one action within that time period. The edge density is used to measure the density of the structure in the graph snapshot. The edge density is obtained by dividing the number of connecting edges in the graph snapshot by the product of the number of nodes in the graph snapshot and the number of nodes minus one. The time window label, number of active nodes, and edge density of the graph snapshot are defined as the graph snapshot sequence of the interest graph of the population profile.

[0097] In this embodiment, step S2 includes the following steps:

[0098] S21. For each snapshot G in the sequence of snapshots of the interest graph of the user profile. (k) =(V (k) E (k) ), calculate the node degree value of each node and set the node degree threshold θ. d Remove all sets of nodes whose degree value is lower than the node degree threshold. Define node degree value

[0099]

[0100] in, Represents node v i In the snapshot G (k) The weighted degree value in the middle, Represents node v i With node v j In the time window Δt k Connection strength parameter of inner connecting edge, 1 {·} This is an indicator function; it is 1 if the edge exists, and 0 otherwise.

[0101] S22. Based on graph snapshot G (k) Interest semantic vector representation of each node Construct the node semantic similarity matrix S (k) A semantic similarity threshold is set, and edges corresponding to all node pairs with semantic similarity below the threshold are filtered out, resulting in node v. i With node v j In the snapshot G (k) semantic similarity Semantic similarity is used to characterize the degree of similarity between two nodes in their interest semantic vector representations. Semantic similarity is obtained by taking the dot product of the interest semantic vector representations of the two nodes and then dividing by the product of the magnitudes of the two vectors.

[0102] S23. For each graph snapshot G (k) Combined with behavior timestamp t i With time window Δt k Calculate the node activity decay weight factor The node activity decay weight factor is used to reflect the timeliness of node behavior relative to the current system time. It is obtained by dividing the difference between the current system time and the time of the node's most recent behavior by the duration of the time window corresponding to the graph snapshot, and then multiplying by the time decay control coefficient and taking the negative exponent.

[0103] S24. Overall node degree value semantic similarity Weighting factor for node activity decay Constructing a candidate subgraph pruning function:

[0104]

[0105] in, Represents node v i Satisfy the clipping retention condition, θ d θ s θ γ This indicates the corresponding threshold.

[0106] S25. Snapshot G (k) All candidate subgraphs that satisfy the clipping function Node set and its associated edge set Constructing candidate subgraphs A candidate subgraph set is formed by combining candidate subgraphs from all time windows:

[0107]

[0108] S26. On the training dataset, select the candidate subgraph set G. sub The objective function is trained using the click-through rate (AUC) metric. (k) Transaction conversion rate (CVR) (k) With structural fidelity index Dist (k) For each candidate subgraph Perform performance evaluation and construct a performance evaluation result set M. perf :

[0109]

[0110] Among them, AUC (k) Represents a snapshot of the graph G (k) The area under the curve (AUC) of the corresponding candidate subimage in the click prediction task, CVR (k) Represents a snapshot of the graph G (k) Conversion rate performance in transaction conversion prediction tasks, Dist (k) This represents the fidelity measure between the candidate subgraph structure and the original snapshot structure.

[0111] Click-through rate (CTR) is used to measure the classification ability of candidate subgraphs in user click behavior prediction tasks, that is, the model's ability to distinguish between positive and negative samples (clicks / non-clicks). The metric is constructed as follows:

[0112] For each candidate subgraph, the user node in the graph is taken as the target node. The node embedding vector extracted by the graph neural network is input, and the downstream click-through rate prediction task is performed. The user click probability is output. Combined with the real click row labels in the training set, the area under the ROC curve is used as the indicator to measure the prediction ranking performance. The closer the AUC value is to 1, the more accurate the model is in predicting click behavior on the candidate subgraph.

[0113] The transaction conversion metric is used to evaluate the effectiveness of candidate subgraphs in predicting user transaction behavior, specifically the ability to model the conversion rate from click to transaction. The metric is constructed as follows:

[0114] Given a graph structure input similar to click prediction, the output is the predicted transaction probability, i.e., the probability that a user completes a purchase or transaction. This is compared with the actual transaction behavior label and evaluated using the average conversion rate, defined as the proportion of samples that were predicted as positive but were actually positive.

[0115]

[0116] Where τ is the conversion probability threshold, which is usually set to 0.5. This is an indicator function.

[0117] The structure fidelity metric measures whether, after graph pruning and graph pooling, the candidate subgraph structure faithfully preserves the key topological structure of the original graph snapshot. It is constructed as follows:

[0118] The adjacency matrix of the original graph snapshot is compared with the adjacency matrix of the candidate subgraph. The Frobenius norm is used to characterize the structural difference between the two. The smaller Dist is, the more complete the structure is preserved, avoiding the loss of structural information due to pruning or pooling operations.

[0119] In this embodiment, step S3 includes the following steps:

[0120] S31. Using the candidate subgraph set G sub With the performance evaluation result set M perf As input, construct the hyperparameter search space for the artificial bee colony optimization algorithm, where each solution vector x represents a set of pooling hyperparameter configurations:

[0121] x = (p, m, λ, α);

[0122] Where p represents the graph pooling ratio hyperparameter in the DiffPool module, m represents the number of cluster centers after each graph pooling layer, λ represents the graph structure preservation regularization coefficient, which controls the structural constraint strength between the original graph and the pooled graph, and α represents the graph attention gating coefficient, which is used to adjust the information fusion weights.

[0123] S32. Define the fitness function f(x) of the artificial bee colony optimization algorithm. The fitness function is used to measure the overall effect of each pooling hyperparameter configuration on the candidate subgraph set. Construct the objective function based on the evaluation index in S26:

[0124]

[0125] Among them, AUC (k) (x), CVR (k) (x) and Dist (k) (x) represents the k-th candidate subgraph. Above, the area under the click-through rate curve score, transaction conversion rate score, and structure fidelity metric obtained after applying pooling hyperparameter configuration x, with w1, w2, and w3 as weighting coefficients;

[0126] S33. Initialize the set of individual bee colonies X = {x1, x2, ..., x...} for the artificial bee colony optimization algorithm. n}, where n is the initial bee colony size, and x is the number of individual bees in the colony. i Let f(x) represent a set of candidate pooling hyperparameter configurations and compute the fitness function f(x) for each individual in the bee colony. i );

[0127] S34. In the bee discovery phase of the artificial bee colony optimization algorithm, for each individual bee x i Execute parameter perturbation to generate neighborhood solution x′ i And evaluate the fitness of the neighborhood solutions. If the fitness function of the neighborhood solutions is f(x′) i )>f(x i If x′ = 0, then the neighborhood solution x′ is used. i Replace individual bee in the colony x i ;

[0128] S35. In the follow-bee stage of the artificial bee colony optimization algorithm, a probabilistic selection strategy is constructed based on the fitness distribution of the current bee colony individual set. Some individuals are selected for local search and perturbation in a probabilistic manner, the solution space is updated and the fitness function is recalculated.

[0129] S36. In the scout bee phase of the artificial bee colony optimization algorithm, bee individuals whose fitness function has not improved within a preset time are randomly reinitialized to form a new solution vector x. r Add them to the bee colony individual set X, and perform fitness function calculation;

[0130] S37. Set the maximum number of iterations or the optimal solution convergence condition, execute the artificial bee colony optimization iterative process, and output the optimal pooling hyperparameter configuration of the fitness function when the termination condition is met.

[0131] In this embodiment, step S4 includes the following steps:

[0132] S41. Using the optimal pooling hyperparameter configuration as input configuration parameters, for each graph snapshot G in the sequence of graph snapshots of the interest graph of the user profile, ... (k) and its corresponding candidate subgraph Construct a structured adaptive graph convolutional network module that incorporates a multi-head attention mechanism, and use node neighborhood interest to aggregate weights. Update the interest embedding vector for each node

[0133]

[0134] in, The node v in the k-th graph snapshot after convolution of the l-th layer graph represents the node v. i Interest embedding vector, N(v) i ) represents node v i The set of first-order interest neighborhood nodes, W (l) Let σ(·) be the learnable feature transformation weight matrix of the l-th layer, where σ(·) represents the nonlinear activation function and σ(·) represents the node neighborhood interest aggregation weight. Represents node v j For node v i The neighborhood contribution of interest representation is obtained by jointly calculating the semantic similarity of node interest and the edge connection strength parameter;

[0135] S42. Concatenate the DiffPool module with the structure-adaptive graph convolutional network module to form a hierarchical graph neural network model. In the l-th layer of DiffPool pooling, construct a soft clustering matrix S that adapts to the criticality of interest nodes in the user group. (k,l) The soft cluster assignment matrix represents the probability that each interest node belongs to each interest cluster. The assignment probability is generated through mapping by a multilayer perceptron network. The graph attention gating coefficient α is introduced into the multilayer perceptron network. * This is used to adjust the influence of each interest node on the clustering decision. The number of interest clusters in each layer is determined by the graph pooling ratio hyperparameter p. * The control is the integer part of the product of the number of interest nodes in the current layer and the graph pooling ratio;

[0136] S43. Based on the soft clustering matrix generated in each layer of the graph pooling module, perform adaptive interest aggregation on the node interest embedding vectors. Adaptive interest aggregation obtains the l-th layer node interest embedding vector X by weighting and summing the interest embedding vectors of all interest nodes according to their corresponding clustering probabilities. (k,l) The adjacency structure relationships of the original graph are mapped according to the assigned probabilities to obtain the graph structure vector A of the l-th layer interest clusters. (k,l) ;

[0137]

[0138] S44. Construct an interest structure fidelity loss function in each graph pooling module. The interest structure fidelity loss function is used to measure the degree of structural deviation between the original graph adjacency structure and the compressed interest cluster graph.

[0139]

[0140] in, This is used to measure the deviation between the interest graph structure after pooling in the l-th layer of DiffPool and the interest graph structure before pooling. The interest structure fidelity loss function guides the DiffPool module to adaptively retain key node and interest cluster information.

[0141] S45. In the top-level interest cluster graph representation of the hierarchical graph neural network model, a cross-layer residual connection mechanism is used to fuse the interest cluster embedding vectors of different layers. The fusion process involves multiplying the node interest embedding vectors output by each layer by the learnable residual connection weights and then weighting and stacking them layer by layer. A global average pooling operation is then used to summarize all node interest embedding vectors to obtain the high-order interest graph embedding vector h of the population profile interest graph corresponding to the graph snapshot. (k) ;

[0142] S46. High-order interest graph embedding vector h that aggregates all time window snapshots (k) Construct a hierarchical structure vector set H of time windows for subsequent interest representation fusion and supervised learning. snap .

[0143] In this embodiment, step S5 includes the following steps:

[0144] S51. The high-order interest graph embedding vectors corresponding to each snapshot in the time window hierarchical structure vector set are concatenated in chronological order to form a time series interest vector matrix. The time series interest vector matrix is ​​used to preserve the evolution trajectory of the interest graph structure of the population profile under different time windows. Each row represents a high-order interest graph embedding vector of a time snapshot, and each column represents an interest embedding dimension.

[0145] S52. Construct a sparse attention fusion module to process the time series interest vector matrix H. seq Attention weights are assigned by embedding snapshots of each image, resulting in a time-weighted interest fusion representation vector h. agg :

[0146]

[0147] Where, α k This represents the attention weight embedded in the k-th time snapshot;

[0148] S53. Use the interest fusion representation vector generated by the sparse attention fusion module as the training input to construct an offline supervised learning model. The offline supervised learning model is defined as an offline teacher model. The offline teacher model is a multi-layer neural network with a high-capacity structure. Its input is the interest fusion representation vector and its output is the prediction result, which is used to restore the recommendation task results of users in specific interest behavior scenarios.

[0149] S54. Based on the true label value of each sample in the training dataset and the prediction result of the offline teacher model, a supervised loss function is constructed. The supervised loss function is used to measure the degree of difference between the predicted value and the true label. The supervised loss function uses the form of binary cross-entropy to calculate the logarithmic error between the predicted value and the label value at the sample level, and then averages it over all samples. The smaller the supervised loss function, the stronger the prediction ability of the offline teacher model.

[0150] S55. The gradient descent optimization algorithm is used to minimize the supervised loss function and iteratively optimize the parameter set of the offline teacher model. The training process ends when the supervised loss function reaches its minimum value or the termination condition is met. The obtained model parameter set is used to fix the state of the offline teacher model.

[0151] In this embodiment, step S6 includes the following steps:

[0152] S61. Construct a lightweight student model with a simplified structure. The lightweight student model is a deep compressed graph neural network structure. Its feature expression dimension and network depth are both smaller than those of the offline teacher model. The input is a set of sparse node-level features of candidate subgraphs or a set of low-dimensional interest representations. It does not depend on the complete graph structure information.

[0153] S62. Extract the output distribution of each training sample from the offline teacher model as a soft label. Simultaneously, combining the true labels of each sample Constructing a knowledge distillation loss function The knowledge distillation loss function consists of two parts: the Kullback-Leibler divergence loss between the soft labels and the student model's predictions, and the cross-entropy loss between the true labels and the student model's predictions.

[0154]

[0155] in, Let represent the predicted output of the i-th sample calculated by the lightweight student model, β be the weight hyperparameter of the distillation loss, KL(·||·) represent the Kullback-Leibler divergence, and CE(·,·) represent the cross-entropy loss function.

[0156] S63. Utilizing the knowledge distillation loss function Optimize the parameter set Θ of the lightweight student model student Iteratively minimize using backpropagation and gradient descent Obtain the lightweight student model structure after distillation;

[0157] S64. Deploy the lightweight student model trained by distillation to the online real-time recommendation system, build an inference interface to accept user request data from real-time traffic scenarios, and input data includes user current context scene features, user historical behavior interest tag sequence and timestamp compression features. After calculation by the lightweight student model, the recommendation task results are output.

[0158] S65. In real-time traffic scenarios, the recommendation score is based on the output of the lightweight student model. Based on preset classification decision rules, determine the recommended status label for the user's current behavior:

[0159] If recommended score If the user behavior context is a cold start environment, then the current user state is marked as having unknown interests;

[0160] If recommended score Furthermore, if there are sparse jumps in time in the historical interest tag sequence, then the user's state is marked as an interest change;

[0161] If recommended score If the similarity between the current interest tag and the historical tags exceeds the threshold, then the user's status is marked as stable in terms of interests.

[0162] Example 1:

[0163] A digital science lab officially launched a special test in City A, aiming to evaluate the performance of different structural modeling algorithms in the scenario of predicting user interests. In the previous user profiling system, the R&D team mainly used a static binary graph structure in conjunction with a standard GCN for interest extraction. However, when dealing with the large-scale user traffic during the Double Eleven promotion, the system frequently experienced problems such as inaccurate user profiling, decreased cold start recommendation accuracy, and fluctuating online response latency. Therefore, the team decided to introduce this invention for comparative experiments to verify its comprehensive advantages in terms of structural adaptability, recommendation effect, and system deployment efficiency.

[0164] The A / B testing environment for the test platform's recommendation system used e-commerce user behavior data from three cities between October 1st and November 1st, 2024. This data covered user browsing, clicking, adding to cart, and ordering behaviors, with a total of 120 million raw behavior logs. After data preprocessing, 3.76 million active users, 2.24 million interest tags (such as "headphones," "down jackets," and "children's picture books"), and approximately 470,000 contextual scenarios (such as "browsing in the morning," "ordering late at night," and "subway network") were extracted. This data was used to construct the original ternary heterogeneous graph (user-interest tag-contextual scenario).

[0165] In traditional methods, engineers use a combination of GCN and Attention to model the graph without distinguishing time windows or pruning the structure. They process the entire graph and extract static embeddings for recommendation. In the experimental group, the method of this invention was applied. First, three time window granularities (last 1 day, last 3 days, and last 7 days) were set to generate corresponding graph snapshot sequences. Then, the team used a triple factor pruning based on node degree, semantic similarity, and time activity to extract the core substructure in the graph. On average, each graph was compressed from the initial 390,000 nodes to approximately 56,000 nodes, with an edge compression rate of 86.7%.

[0166] Subsequently, based on the subgraph set generated from each graph snapshot, CTR and CVR task models were trained on the offline training set, and a performance evaluation matrix was generated according to the output metrics. The artificial bee colony optimization module used a population size of n=30 and a maximum number of iterations of 60 rounds. The four-dimensional hyperparameters of pooling ratio p, number of interest clusters m, graph structure preservation coefficient λ, and graph attention gating coefficient α were searched using the fitness function. Finally, the optimal solution was determined to be (p=0.25, m=128, λ=0.05, α=0.9). The average value of this set of parameters was 0.746 for AUC, 0.412 for CVR, and 0.027 for structure fidelity Dist. Compared with manually configured parameters, the average improvements were +4.8%, +3.9%, and -12.2%, respectively.

[0167] In the structural training phase, a hierarchical graph neural network model is constructed, employing a 5-layer DiffPool module stacked with an adaptive graph convolutional structure. For each time-segment graph snapshot, a high-order interest embedding vector is output. All snapshot outputs are then processed by a sparse attention fusion module to form an interest fusion representation vector, which is input into an offline teacher model composed of a multi-layer fully connected neural network. Model training is performed using a joint loss optimization method of AUC and cross-entropy. The final results are then transferred to a simplified graph neural student model through knowledge distillation. This model has only 23% of the parameters of the teacher model and is deployed in an online system for real-time prediction.

[0168] During the Double Eleven pre-sale period from November 5th to 8th, 2024, the platform conducted a 72-hour online A / B test on user interest modeling. The control group continued to use the traditional GCN model, while the experimental group used the distilled student model deployed by the method of this invention. In online real-time recommendation scenarios (including personalized recommendations on the homepage, "You May Also Like," and special pages), the peak daily online request volume reached 34 million times. The test metrics are as follows:

[0169] Table 1 Comparison of test index data between the method of the present invention and the traditional method.

[0170]

[0171] Among them, the sample size of cold-start users (registered for ≤7 days and with no more than 3 shopping records) was approximately 42,000, and the present invention improved the recommendation accuracy for them by over 12%; while among users with abrupt changes in interest (whose interest tags switched to another product category within 24 hours), the system based on the interest stability index set within the model. The value and context aggregation tag similarity automatically detect interest shift behavior and trigger the model fine-tuning process, making its average delay from detecting interest change to adjusting recommended content only 38 minutes, which is significantly better than the traditional 3-hour batch update strategy.

[0172] In addition, system resource consumption assessment data shows that the online deployment version of the model of this invention has reduced CPU utilization by 27.8% and average memory usage by 19.3%, adapts to CDN edge node deployment in multiple locations, and supports higher frequency update requirements.

[0173] In summary, by employing heuristic pruning of the user interest graph, bee colony optimization search with pooling parameters, and hierarchical representation modeling of the graph structure, this invention effectively alleviates the problems of structural redundancy, singular interest expression, and slow response of traditional methods, improving the accuracy, response speed, and resource adaptability of recommendations, demonstrating good feasibility and engineering feasibility.

[0174] This invention introduces an artificial bee colony optimization algorithm to dynamically select pooling hyperparameters in a graph neural network, achieving joint optimization of the graph pooling ratio, number of interest clusters, structure preservation coefficient, and attention gating coefficient in the DiffPool module. By constructing a fitness function with prediction accuracy and structure fidelity as its core, the optimal hyperparameter combination is accurately identified by utilizing the global search capability of the artificial bee colony in the solution space.

[0175] This invention introduces a heuristic multi-factor pruning mechanism in the graph structure modeling stage. It extracts candidate subgraphs from graph snapshots by comprehensively considering three-dimensional features: node degree, semantic similarity, and temporal activity. This avoids the computational redundancy and information dilution problems caused by full-graph computation in traditional graph neural networks. Through a node retention function, it automatically selects a subset of nodes with high information density, active behavior, and strong semantic concentration in the graph snapshot to form a set of candidate subgraphs, effectively compressing the graph size while preserving the core semantic structure.

[0176] This invention designs an interest graph vector integration mechanism that integrates hierarchical graph structure and temporal evolution expression. Combined with a sparse attention fusion module, it weights and integrates high-order interest embedding vector sequences generated under multi-granularity time windows to construct an interest fusion representation vector for offline supervised learning teacher model training. Furthermore, it transfers the knowledge through knowledge distillation to a lightweight student model for deployment in real-time recommendation systems. This mechanism can significantly capture the nonlinear change characteristics of user interests in the time dimension and effectively improve the ability to identify complex behavioral patterns of cold-start users and interest-switching users.

[0177] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for enhancing the modeling of user profile interest graphs based on deep learning, characterized in that, Includes the following steps: S1. Construct a ternary heterogeneous graph, and generate a snapshot sequence of the user profile interest graph based on the ternary heterogeneous graph using a multi-granularity time window. The nodes of the ternary heterogeneous graph consist of user nodes, interest tag nodes, and context scene nodes. S2. Perform heuristic cropping on each snapshot in the sequence of snapshots of the interest graph of the user profile to generate corresponding candidate subgraphs and form a set of candidate subgraphs. Perform offline task model training on the set of candidate subgraphs on the training dataset to obtain a set of performance evaluation results. Step S2 includes the following steps: S21. For each snapshot in the sequence of interest graph snapshots for the user profile ,in, For interest tag nodes, For the connection edges between interest tag nodes and context scene nodes, calculate the node degree value of each node and set a node degree threshold. Remove all sets of nodes whose degree value is lower than the node degree threshold. , obtain the node degree value ; S22. Based on graph snapshot Interest semantic vector representation of each node Construct a node semantic similarity matrix And set a semantic similarity threshold. Remove all edges corresponding to node pairs with semantic similarity below the semantic similarity threshold to obtain the node... With nodes In the snapshot semantic similarity ; S23. Snapshot for each image Combined with behavior timestamps With time window Calculate the node activity decay weight factor It is obtained by dividing the difference between the current system time and the time of the node's most recent behavior by the duration of the time window corresponding to the graph snapshot, and then multiplying by the time decay control coefficient by taking the negative exponent. S24. Overall node degree value semantic similarity Weighting factor for node activity decay Construct candidate subgraph pruning functions: ； in, Represents a node Meets the conditions for cropping and retaining. This indicates the corresponding threshold. S25. Snapshot of the image All candidate subgraphs that satisfy the clipping function Node set and its associated edge set Constructing candidate subgraphs A candidate subgraph set is formed by aggregating candidate subgraphs from all time windows. ; S26. On the training dataset, select the candidate subgraph set. The objective function is trained using the click-through rate (CTR) metric. Transaction conversion metrics With structural fidelity index For each candidate subgraph Perform performance evaluation and construct a performance evaluation result set. ; S3. Using the candidate subgraph set and the performance evaluation result set as input, define the search space and fitness function of the artificial bee colony optimization algorithm, and evaluate the fitness of the pooling hyperparameter candidate solution set based on the fitness function. Select the optimal pooling hyperparameter configuration based on the fitness evaluation result. Define the fitness function of the artificial bee colony optimization algorithm. The fitness function measures the overall performance of each pooling hyperparameter configuration on the candidate subgraph set. The objective function is constructed based on the evaluation metric S26. ； in, , and They represent the first time. candidate subgraphs Above, application pooling hyperparameter configuration The resulting area under the click-through rate curve score, conversion rate score, and structure fidelity metric were: , , These are weighting coefficients; S4. Under the optimal pooling hyperparameter configuration, execute a hierarchical graph neural network model for each graph snapshot and its corresponding candidate subgraph, extract the structural embedding vector of each snapshot, and output a set of hierarchical structural vectors for the time window. S5. Concatenate the hierarchical structure vector set of time windows according to the time order of the graph snapshot sequence to obtain the interest fusion representation vector, and construct the offline teacher model; S6. Based on the output results and parameters of the offline teacher model, a simplified lightweight student model is trained by knowledge distillation, and the lightweight student model is deployed on an online real-time recommendation system, and the inference results are output in real-time traffic scenarios. S7. When a key performance indicator is detected to be lower than a preset threshold or a change in user interest distribution is detected to be greater than a preset threshold, repeat steps S2 to S6.

2. The method for enhancing the modeling of user profile interest graphs based on deep learning according to claim 1, characterized in that, Step S1 includes the following steps: S11. Collect raw behavior log data to form a raw behavior log set; S12. Based on the original behavior log set, extract the user identifier set, interest tag set and context scene set respectively, and jointly construct a ternary heterogeneous graph by combining the user identifier set, interest tag set and context scene set; S13. Set a set of multi-granularity time windows to divide the time dimension of user behavior. For each time window, extract the structural subgraph within the time period from the ternary heterogeneous graph to generate a graph snapshot sequence. Each graph snapshot in the graph snapshot sequence consists of the set of valid connection edges and the set of nodes involved in the connection edges within the corresponding time period. S14. For each graph snapshot in the graph snapshot sequence, record its time window label, number of active nodes, and edge density. The edge density is obtained by dividing the number of connected edges in the graph snapshot by the product of the number of nodes in the graph snapshot and the number of nodes minus one. The time window label, number of active nodes, and edge density of the graph snapshot are defined as the graph snapshot sequence of the interest graph of the population profile.

3. The method for enhancing the modeling of user profile interest graphs based on deep learning according to claim 2, characterized in that, In the ternary heterogeneous graph, the connection edges between user nodes and interest tag nodes represent the relationship between user behavior and interest tag behavior, while the connection edges between user nodes and context scene nodes represent the context scene information in which the user behavior occurs. Connection edges between the context scene nodes This is used to represent the co-occurrence relationship of interest behaviors in a specific context. Each connection edge is associated with a connection strength parameter. The connection strength parameter between user nodes and interest tag nodes is used to characterize the weighted result of the frequency and freshness of user behavior on a certain interest tag. The connection strength parameter between user nodes and context scene nodes is used to characterize the expression strength of user activity in different context scenes. The connection strength parameter between interest tag nodes and context scene nodes is used to characterize the frequency of co-occurrence of interest tags in different context scenes.

4. The method for enhancing and modeling user profile interest graphs based on deep learning according to claim 3, characterized in that, Step S3 includes the following steps: S31. Using the candidate subgraph set With performance evaluation results set Using the input, construct the hyperparameter search space for the artificial bee colony optimization algorithm, where each solution vector in the search space... This represents a set of pooling hyperparameter configurations: ； in, This represents the graph pooling ratio hyperparameter in the DiffPool module. This represents the number of cluster centers after pooling at each graph level. The graph structure maintains regularity coefficients, controlling the strength of structural constraints between the original graph and the pooling graph. This represents the graph attention gating coefficient, used to adjust the information fusion weights; S32. Initialize the set of bee colony individuals for the artificial bee colony optimization algorithm. ,in For the initial colony size, each bee colony individual This represents a set of candidate pooling hyperparameter configurations, and computes the fitness function for each individual in the bee colony. ; S33. In the bee discovery phase of the artificial bee colony optimization algorithm, for each individual bee in the colony... Execution parameter perturbation to generate neighborhood solutions And evaluate the fitness of the neighborhood solutions. If the fitness function of the neighborhood solutions is... Then use the neighborhood solution. Replace individual bees in the colony ; S34. In the follow-bee stage of the artificial bee colony optimization algorithm, a probabilistic selection strategy is constructed based on the fitness distribution of the current bee colony individual set. Some individuals are selected for local search and perturbation in a probabilistic manner, the solution space is updated and the fitness function is recalculated. S35. In the scout bee phase of the artificial bee colony optimization algorithm, bee individuals whose fitness function has not improved within a preset time are randomly reinitialized to form new solution vectors. Join the bee colony individual collection In the middle, and perform fitness function calculation; S36. Set the maximum number of iterations or the optimal solution convergence condition, execute the artificial bee colony optimization iterative process, and output the optimal pooling hyperparameter configuration of the fitness function when the termination condition is met. .

5. The method for enhancing the modeling of user profile interest graphs based on deep learning according to claim 4, characterized in that, Step S4 includes the following steps: S41. Using the optimal pooling hyperparameter configuration as input configuration parameters, and targeting the snapshot sequence of the interest graph of the user profile. Each snapshot in the image and its corresponding candidate subgraph We construct a structured adaptive graph convolutional network module that incorporates a multi-head attention mechanism, and use node neighborhood interest to aggregate weights. Update the interest embedding vector for each node ; S42. Concatenate the DiffPool module with the structure-adaptive graph convolutional network module to form a hierarchical graph neural network model. In the layered DiffPool pooling, a soft cluster allocation matrix is ​​constructed that adapts to the criticality of interest nodes in the user group. The soft cluster assignment matrix represents the probability that each interest node belongs to each interest cluster. The assignment probability is generated through mapping by a multilayer perceptron network, in which graph attention gating coefficients are introduced. This is used to adjust the influence of each interest node on the clustering decision. The number of interest clusters in each layer is determined by the graph pooling ratio hyperparameter. The control is the integer part of the product of the number of interest nodes in the current layer and the graph pooling ratio; S43. Based on the soft cluster allocation matrix generated in each layer of the graph pooling module, perform adaptive interest aggregation on the node interest embedding vectors. Adaptive interest aggregation obtains the result by weighting and summing the interest embedding vectors of all interest nodes according to their corresponding cluster allocation probabilities. Layer node interest embedding vector The adjacency structure of the original graph is mapped according to the assigned probability to obtain the first... Layered interest cluster graph structure vector ; S44. Construct an interest structure fidelity loss function in each graph pooling module. The interest structure fidelity loss function is used to measure the degree of structural deviation between the adjacency structure of the original graph and the compressed interest cluster graph. S45. In the top-level interest cluster graph representation of the hierarchical graph neural network model, a cross-layer residual connection mechanism is used to fuse the interest cluster embedding vectors of different layers. The fusion process involves multiplying the node interest embedding vectors output by each layer by the learnable residual connection weights and then weighting and stacking them layer by layer. A global average pooling operation is then used to summarize all node interest embedding vectors to obtain the high-order interest graph embedding vector of the population profile interest graph corresponding to the graph snapshot. ; S46. High-order interest graph embedding vectors that aggregate all time window snapshots Construct a hierarchical structure vector set of time windows for subsequent interest representation fusion and supervised learning. .

6. The method for enhancing the modeling of user profile interest graphs based on deep learning according to claim 1, characterized in that, Step S5 includes the following steps: S51. The high-order interest graph embedding vectors corresponding to each snapshot in the time window hierarchical structure vector set are concatenated in chronological order to form a time series interest vector matrix. The time series interest vector matrix is ​​used to preserve the evolution trajectory of the interest graph structure of the population profile under different time windows. Each row represents a high-order interest graph embedding vector of a time snapshot, and each column represents an interest embedding dimension. S52. Construct a sparse attention fusion module to process the time series interest vector matrix. Attention weights are assigned by embedding snapshots of each image, resulting in a time-weighted interest fusion representation vector. ; S53. Use the interest fusion representation vector generated by the sparse attention fusion module as the training input to construct an offline supervised learning model. The offline supervised learning model is defined as an offline teacher model. The offline teacher model is a multi-layer neural network with a high-capacity structure. Its input is the interest fusion representation vector and its output is the prediction result, which is used to restore the recommendation task results of users in specific interest behavior scenarios. S54. Based on the true label value of each sample in the training dataset and the prediction result of the offline teacher model, a supervised loss function is constructed. The supervised loss function is used to measure the degree of difference between the predicted value and the true label. The supervised loss function uses the form of binary cross-entropy to calculate the logarithmic error between the predicted value and the label value at the sample level, and then averages it over all samples. The smaller the supervised loss function, the stronger the prediction ability of the offline teacher model. S55. The gradient descent optimization algorithm is used to minimize the supervised loss function, and the parameter set of the offline teacher model is iteratively optimized. The training process ends when the supervised loss function reaches its minimum value or the termination condition is met.

7. The method for enhancing the modeling of user profile interest graphs based on deep learning according to claim 6, characterized in that, Step S6 includes the following steps: S61. Construct a lightweight student model with a simplified structure. The lightweight student model is a deep compressed graph neural network structure. S62. Extract the output distribution of each training sample from the offline teacher model as a soft label. At the same time, combined with the true label of each sample Construct a knowledge distillation loss function The knowledge distillation loss function consists of two parts: one part is the Kullback-Leibler divergence loss between the soft label and the student model prediction, and the other part is the cross-entropy loss between the true label and the student model prediction. S63. Utilizing the knowledge distillation loss function Optimize the parameter set of the lightweight student model The knowledge distillation loss function is iteratively minimized using backpropagation and gradient descent. To obtain the lightweight student model structure after distillation; S64. Deploy the lightweight student model trained by distillation to the online real-time recommendation system, build an inference interface to accept user request data from real-time traffic scenarios, and input data includes user current context scene features, user historical behavior interest tag sequence and timestamp compression features. After calculation by the lightweight student model, the recommendation task results are output. S65. In real-time traffic scenarios, the recommendation score is based on the output of the lightweight student model. Based on preset classification decision rules, determine the recommended status label for the user's current behavior: If recommended score If the user behavior context is a cold start environment, then the current user state is marked as having unknown interests; If recommended score If there are sparse time jumps in the historical interest tag sequence, then the user's status is marked as an interest change; If recommended score If the similarity between the current interest tag and the historical tag exceeds the threshold, then the user's state is marked as stable interest.

Images