Private domain live broadcast user behavior analysis and traffic recommendation method based on deep learning

Abstract

The present invention provides a private domain live broadcast user behavior analysis and traffic diversion recommendation method based on deep learning, which relates to the field of deep learning technology, including: collecting user behavior data, constructing a multimodal deep neural network for processing, extracting user temporal behavior features, using a residual connection structure and contrastive learning to extract user text intention features, obtaining user behavior link features through a temporal graph attention network, performing user stratification through a Gaussian mixture clustering model, predicting user churn probability using a deep cross neural network, and predicting user future activity and purchasing power in combination with a recurrent neural network, matching products through a graph attention product knowledge network, using an improved dual-tower deep neural network combined with difficult sample mining and a variational autoencoder for recommendation, and searching and optimizing push strategies to obtain the optimal push strategy.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a private domain live broadcast user behavior analysis and traffic diversion recommendation method based on deep learning. Background Art

[0002] The rise of deep learning has provided powerful technical support for user behavior analysis and personalized recommendations. In the field of private live streaming, how to deeply understand user behavior, accurately predict user needs, and provide personalized product recommendations and traffic diversion services is crucial to improving user conversion rates and platform revenues.

[0003] Existing recommendation systems are usually based on static user attributes and simple behavioral data. They cannot effectively integrate multimodal user behavior data and lack effective user churn prediction and precise traffic diversion strategies. Therefore, it is difficult to accurately divert traffic based on the needs and preferences of different users, resulting in poor traffic diversion effects.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides a private domain live broadcast user behavior analysis and traffic diversion recommendation method based on deep learning, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for analyzing user behavior and traffic diversion of private domain live broadcasts based on deep learning, comprising:

[0007] Collect the behavior data of private domain live broadcast users, build a multimodal deep neural network based on the multi-head self-attention mechanism, add the behavior data to the multimodal deep neural network, perform time-series weighted processing through a bidirectional long short-term memory network module combined with a gated recursive unit and a time decay factor to obtain user time-series behavior features, extract user comment text from the behavior data and add it to the residual connection structure module, obtain user text intention features through contrastive learning processing, extract user behavior paths from the behavior data and add them to the time-series graph attention network module, obtain user behavior link features through random walk node sampling and graph neural network message passing, perform feature transformation and normalization on the user time-series behavior features, the user text intention features and the user behavior link features through a multi-layer perceptron to obtain user fusion features;

[0008] Based on the user fusion features, a Gaussian mixture clustering model is constructed, users are adaptively stratified by calculating the inter-class distance distribution, a deep cross neural network is constructed, the user fusion features are input into the deep cross neural network, the feature combination relationship is extracted through the feature interaction layer, the feature expression is optimized through the residual network layer, the user churn probability value is obtained, a recurrent neural network prediction model is constructed, the user's historical purchase data is input into the recurrent neural network prediction model, the user's future activity and purchasing power score are simultaneously predicted using a multi-task learning framework, and a drainage weight coefficient is generated based on the user churn probability value and the user's future activity and purchasing power score;

[0009] Construct a graph attention product knowledge network, input the user fusion features into the graph attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, and obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through the Thompson sampling method and optimize the push strategy in combination with the Monte Carlo tree search method to obtain the optimal push strategy.

[0010] In an optional embodiment,

[0011] The behavior data of private domain live broadcast users is collected, and a multimodal deep neural network is constructed based on a multi-head self-attention mechanism. The behavior data is added to the multimodal deep neural network, and time-series weighted processing is performed through a bidirectional long short-term memory network module in combination with a gated recursive unit and a time decay factor to obtain user time-series behavior features. User comment text is extracted from the behavior data and added to the residual connection structure module. User text intention features are obtained through contrastive learning processing. User behavior paths are extracted from the behavior data and added to the time-series graph attention network module. User behavior link features are obtained through random walk node sampling and message passing of graph neural networks. The user time-series behavior features, the user text intention features, and the user behavior link features are transformed and normalized through a multi-layer perceptron to obtain user fusion features including:

[0012] Collecting the behavior data of private domain live broadcast users, the behavior data includes user viewing time data, like data, comment data, reward data, and sharing data, normalizing and converting the behavior data into numerical values ​​to generate preprocessed data;

[0013] Construct a query vector, a key vector and a value vector, calculate a similarity matrix between the query vector and the key vector, generate a multi-head attention weight according to the similarity matrix, perform weighted summation of the multi-head attention weight and the value vector to obtain an attention feature, construct a graph structure for the user comment nodes and behavior nodes in the preprocessed data, obtain graph attention features through node aggregation operations, construct a user-product heterogeneous graph, and extract heterogeneous graph features through a node-level attention layer and a semantic-level attention layer;

[0014] Input the attention feature, the graph attention feature and the heterogeneous graph feature into a bidirectional long short-term memory network, extract bidirectional time series information through a forward propagation layer and a back propagation layer, control the update and forgetting of information flow through an update gate and a reset gate, calculate a time decay factor based on a time interval, and weight features at different time steps to obtain user time series behavior features;

[0015] Perform word segmentation and word vector conversion on the user comment text in the preprocessed data, input the word vector into the backbone network for feature extraction, retain the original features through skip connections, construct a contrastive learning loss function, minimize the feature distance of the comment text corresponding to different behaviors of the same user, maximize the feature distance of the comment text of different users, and obtain the user text intention feature;

[0016] Extracting a user behavior sequence from the preprocessed data, constructing a user behavior directed graph, performing temporal position encoding on nodes in the user behavior directed graph, performing random walk sampling on the user behavior directed graph based on an alias sampling table, generating a user behavior subgraph, performing message transmission and feature aggregation on the user behavior subgraph, and obtaining user behavior link features;

[0017] The user temporal behavior features, the user text intention features and the user behavior link features are input into the feature mapping network for dimension conversion, and the converted features are batch normalized. The nonlinear representation of features is introduced through a nonlinear activation function, the processed features are spliced ​​through a residual connection, the feature attention weights are calculated and weighted fused with the spliced ​​features to obtain user fused features.

[0018] In an optional embodiment,

[0019] The time decay factor is calculated based on the time interval as shown in the following formula:

[0020] ;

[0021] in, SS represents the time decay factor, σ represents the standard deviation of the Gaussian decay function, w 1represents the time decay weight in the Sigmoid decay function, b 1 represents the bias term of the Sigmoid decay function, β Represents the learnable Gaussian attenuation weight coefficient, △ t represents the time interval, exp() represents the exponential function, γ represents the contribution weight corresponding to the hyperbolic tangent function, w 2 represents the decay rate weight coefficient in the hyperbolic tangent function, b 2 Represents the bias term for the hyperbolic tangent function.

[0022] In an optional embodiment,

[0023] Based on the user fusion features, a Gaussian mixture clustering model is constructed, and users are adaptively stratified by calculating the inter-class distance distribution, and a deep cross neural network is constructed. The user fusion features are input into the deep cross neural network, and the feature combination relationship is extracted through the feature interaction layer. The feature expression is optimized through the residual network layer to obtain the user churn probability value, and a recurrent neural network prediction model is constructed. The user's historical purchase data is input into the recurrent neural network prediction model, and the user's future activity and purchasing power scores are simultaneously predicted using a multi-task learning framework. The drainage weight coefficient is generated based on the user churn probability value and the user's future activity and purchasing power scores, including:

[0024] Constructing a Gaussian mixture clustering model, generating an independent Gaussian distribution model for each pre-set user subgroup, each of the Gaussian distribution models including a mean vector, a covariance matrix and a mixing coefficient, constructing an expectation maximization iterative optimization module, inputting user fusion features into the Gaussian mixture clustering model, iteratively optimizing the parameters of the Gaussian distribution model by maximizing the log-likelihood function through the expectation maximization iterative optimization module, calculating the posterior probability that the sample point belongs to the user subgroup based on the Mahalanobis distance between the sample point and the center of the user subgroup, dividing the sample point into the user subgroup with the largest posterior probability, and obtaining a user stratification result;

[0025] Constructing a deep cross neural network, wherein the feature embedding module of the deep cross neural network includes a plurality of parallel embedding layers, the feature cross module of the deep cross neural network includes a plurality of cross layers connected in series, and the feature optimization module of the deep cross neural network includes a plurality of residual layers connected in series;

[0026] Input the user fusion feature into the embedding layer of the feature embedding module to map the sparse feature into a dense vector, input the dense vector into the cross layer of the feature cross module to perform a pairwise cross operation on different features to generate a high-order interactive feature, input the high-order interactive feature into the residual layer of the feature optimization module to perform element-wise addition on the dense vector to obtain an optimized feature, and pass the optimized feature through a fully connected layer and a Sigmoid activation function to obtain a user churn probability value;

[0027] Constructing a recurrent neural network prediction model, wherein the sequence processing module of the recurrent neural network prediction model includes a plurality of gated recurrent units connected in series, each of the gated recurrent units includes an update gate, a reset gate and a candidate hidden state, and the multi-task prediction module of the recurrent neural network prediction model includes a feature sharing layer and two task-specific layers;

[0028] Arrange the user's historical purchase data in chronological order to form a purchase sequence, input the purchase sequence into the sequence processing module, the gated recurrent unit updates the current hidden state according to the current purchase behavior and the historical hidden state of the purchase sequence, input the current hidden state into the multi-task prediction module, the feature sharing layer extracts shared features, the two task-specific layers are respectively connected to the user activity prediction task and the user purchasing power prediction task, and the user's future activity score and the user's future purchasing power score are obtained by jointly training the user activity prediction task and the user purchasing power prediction task;

[0029] A user diversion weight coefficient is calculated according to the user stratification result, the user churn probability value, the user future activity score and the user future purchasing power score. For a user subgroup in the user stratification result whose churn risk is higher than a preset churn risk threshold, the user diversion weight coefficient is increased based on the user churn probability value. For users whose user future activity score is lower than a preset activity threshold or whose user future purchasing power score is lower than a preset purchasing power score threshold, the value of the user diversion weight coefficient is increased, wherein the value range of the user diversion weight coefficient is 0 to 1.

[0030] In an optional embodiment,

[0031] The two task-specific layers are respectively connected to the user activity prediction task and the user purchasing power prediction task, and the user future activity score and the user future purchasing power score are obtained by jointly training the user activity prediction task and the user purchasing power prediction task, including:

[0032] Residual connections are introduced in the user activity prediction task-specific layer and the user purchasing power prediction task-specific layer, and the pre-acquired shared features are element-wise added to the output of each task-specific layer to obtain the fused shared features.

[0033] The user activity prediction task specific layer maps the fused shared features to the user's future activity score through a fully connected layer and a Sigmoid activation function, and the user purchasing power prediction task specific layer maps the fused features to the user's future purchasing power score through a fully connected layer and a linear activation function;

[0034] The multi-task prediction module is used as a generator network to construct a discriminator network, wherein the discriminator network receives the user's real future activity data, real future purchasing power data, and the prediction result of the generator network;

[0035] Constructing a multi-task learning joint loss function, the multi-task learning joint loss function includes a cross entropy loss term of a user activity prediction task and a mean square error loss term of a user purchasing power prediction task, adjusting weight coefficients of the cross entropy loss term of the user activity prediction task and the mean square error loss term of the user purchasing power prediction task by dynamic weighting, and dynamically updating the weight coefficients according to the convergence speed of the task;

[0036] The generator network and the discriminator network are alternately trained by a minimax algorithm, wherein the generator network minimizes the multi-task learning joint loss function, and the discriminator network maximizes the classification accuracy of the real data and the generated data, and the training is repeated until a preset maximum number of training times is reached;

[0037] Perform multiple Monte Carlo probability sampling on the user's recent purchase sequence, input the sampled sequence into the trained multi-task prediction module to obtain multiple groups of prediction results, and obtain the final user future activity score and user future purchasing power score by integrating the multiple groups of prediction results.

[0038] In an optional embodiment,

[0039] Construct a graph attention product knowledge network, input the user fusion features into the graph attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through a Thompson sampling method and optimize the push strategy in combination with a Monte Carlo tree search method to obtain the optimal push strategy, including:

[0040] Construct a graph attention product knowledge network, take product categories, brands, and functions as nodes, and take common attributes, complementary relationships, and substitution relationships between products as edges to form a heterogeneous graph network structure, use the user fusion features as query conditions, and obtain a matching product set through multi-hop entity relationship reasoning;

[0041] Constructing a dual-tower deep neural network, the dual-tower deep neural network comprising a user tower and a product tower, the user tower receiving the user fusion feature and mapping it to a user embedding vector, the product tower receiving the features of the products in the matching product set and mapping it to a product embedding vector;

[0042] Through a pre-set difficult sample mining module, select the product with the closest product embedding vector distance to the product actually interacted by the user from the matching product set as a difficult sample, add the difficult sample to an initial negative sample queue initialized to empty in advance, repeat the selection and addition to obtain the negative sample queue, and construct a contrast loss function based on the negative sample queue to train the dual-tower deep neural network;

[0043] Construct a variational autoencoder, map the user embedding vector and the product embedding vector output by the trained dual-tower deep neural network to a latent space distribution, reconstruct the user-product interaction distribution from the latent space distribution to obtain a product recommendation probability matrix, calculate the drainage weight coefficient corresponding to the category, price, and profit of each product in the matching product set, and multiply the product recommendation probability matrix by the drainage weight coefficient to obtain a final product ranking score;

[0044] An online policy network and a target network are constructed, wherein the online policy network receives the final product ranking score and outputs the Q value of the push policy. The target network periodically copies parameters from the online policy network to estimate the Q value of the state-action pair. The posterior probability distribution of each push policy is calculated according to the Q value output by the online policy network. The Thompson sampling method is used to sample and generate the push policy from the posterior probability distribution. A push decision search tree is constructed in combination with the Monte Carlo tree search method, wherein each node represents a push decision. Simulation selection and evaluation update are repeated to obtain a push decision sequence, which is applied to the live product recommendation and the hyperparameters of the online policy network are updated to obtain the optimal push policy.

[0045] In an optional embodiment,

[0046] From the set of matching products, select the product whose product embedding vector is closest to the product that the user actually interacts with as a difficult sample as shown in the following formula:

[0047] ;

[0048] in, DIndicates the actual interaction product of the user u and products to be matched v The distance between T represents transpose, θ represents the smoothing coefficient, Indicated in t The covariance matrix at -1 time step, h Indicates t The covariance contribution of the newly added samples is W a represents the weight matrix, l a Represents the bias term.

[0049] A second aspect of an embodiment of the present invention provides a private domain live broadcast user behavior analysis and traffic recommendation system based on deep learning, including:

[0050] The first unit is used to collect the behavior data of private domain live broadcast users, build a multimodal deep neural network based on a multi-head self-attention mechanism, add the behavior data to the multimodal deep neural network, perform time-series weighted processing through a bidirectional long short-term memory network module in combination with a gated recursive unit and a time decay factor to obtain user time-series behavior features, extract user comment text from the behavior data and add it to a residual connection structure module, obtain user text intention features through contrastive learning processing, extract user behavior paths from the behavior data and add them to a time-series graph attention network module, obtain user behavior link features through random walk node sampling and message passing of a graph neural network, perform feature transformation and normalization on the user time-series behavior features, the user text intention features and the user behavior link features through a multi-layer perceptron to obtain user fusion features;

[0051] The second unit is used to construct a Gaussian mixture clustering model based on the user fusion features, adaptively stratify users by calculating the inter-class distance distribution, construct a deep cross neural network, input the user fusion features into the deep cross neural network, extract the feature combination relationship through the feature interaction layer, optimize the feature expression through the residual network layer, obtain the user churn probability value, construct a recurrent neural network prediction model, input the user's historical purchase data into the recurrent neural network prediction model, use a multi-task learning framework to simultaneously predict the user's future activity and purchasing power score, and generate a drainage weight coefficient based on the user churn probability value and the user's future activity and purchasing power score;

[0052] The third unit is used to construct a graph attention product knowledge network, input the user fusion features into the graph attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through the Thompson sampling method and optimize the push strategy in combination with the Monte Carlo tree search method to obtain the optimal push strategy.

[0053] According to a third aspect of the embodiments of the present invention,

[0054] An electronic device is provided, comprising:

[0055] processor;

[0056] a memory for storing processor-executable instructions;

[0057] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0058] A fourth aspect of the embodiments of the present invention is:

[0059] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0060] In the present invention, by using a multimodal deep neural network, combining a variety of deep learning models (such as a bidirectional long short-term memory network, a residual connection structure, a temporal graph attention network, etc.) and contrastive learning methods, it is possible to more comprehensively and deeply analyze the user's temporal behavior, text intention, behavior link and other multi-dimensional information, so as to more accurately understand user needs and preferences, and adaptively stratify users through a Gaussian mixture clustering model, and combine a deep cross neural network to predict the probability of user churn, and use a recurrent neural network to predict the user's future activity and purchasing power, so as to generate a more accurate drainage weight coefficient, use a graph attention product knowledge network to match products, and combine an improved dual-tower deep neural network and a variational autoencoder to learn the user-product interaction distribution, and obtain a more accurate product recommendation result by combining the drainage weight and the product recommendation probability, and use a dual deep Q network combined with Thompson sampling and Monte Carlo tree search methods to optimize the push strategy for the final product ranking score, so as to select the best push strategy, improve the conversion rate of product push, and improve the operational efficiency and revenue of private domain live broadcast. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A flowchart of a method for analyzing user behavior and traffic diversion of private domain live broadcasts based on deep learning according to an embodiment of the present invention;

[0062] Figure 2 This is a structural diagram of a private domain live broadcast user behavior analysis and traffic diversion recommendation system based on deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0065] Figure 1 Schematic diagram of the process of the method for analyzing user behavior and traffic diversion of private domain live broadcast based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] S1. Collect the behavior data of private domain live broadcast users, build a multimodal deep neural network based on the multi-head self-attention mechanism, add the behavior data to the multimodal deep neural network, perform time-series weighted processing through a bidirectional long short-term memory network module, combined with a gated recursive unit and a time decay factor, to obtain user time-series behavior features, extract user comment text from the behavior data and add it to the residual connection structure module, obtain user text intention features through contrastive learning processing, extract user behavior paths from the behavior data and add them to the time-series graph attention network module, obtain user behavior link features through random walk node sampling and graph neural network message passing, perform feature transformation and normalization on the user time-series behavior features, the user text intention features and the user behavior link features through a multi-layer perceptron, and obtain user fusion features;

[0067] The multimodal deep neural network refers to a deep learning model that can process and fuse information from different modalities (such as images, text, voice, etc.), extracts features from different modalities and fuses them together through shared learning to achieve more accurate prediction or classification tasks. The gated recurrent unit is an improved recurrent neural network structure that aims to solve the gradient vanishing and explosion problems that traditional RNNs may encounter in the long sequence learning process. The time decay factor is a weight adjustment factor used in time series modeling to assign different weights to historical data. The user text intention feature is a feature that reflects the intention extracted from the text information provided by the user, which is usually implemented through natural language processing technology (such as word vectors, syntactic analysis, etc.), which helps to understand the user's needs and behaviors so as to provide personalized recommendations or services for them. The random walk node sampling is a graph data processing method for effectively sampling subgraphs from large-scale graphs. Samples are generated by randomly walking in the graph and selecting the nodes passed through. The user behavior link feature refers to a feature that describes the user's behavior sequence over a period of time, which helps the model understand the user's behavior pattern and is often used in scenarios such as recommendation systems and advertising to improve the accuracy of personalized recommendations.

[0068] In an optional embodiment,

[0069] The behavior data of private domain live broadcast users is collected, and a multimodal deep neural network is constructed based on a multi-head self-attention mechanism. The behavior data is added to the multimodal deep neural network, and time-series weighted processing is performed through a bidirectional long short-term memory network module in combination with a gated recursive unit and a time decay factor to obtain user time-series behavior features. User comment text is extracted from the behavior data and added to the residual connection structure module. User text intention features are obtained through contrastive learning processing. User behavior paths are extracted from the behavior data and added to the time-series graph attention network module. User behavior link features are obtained through random walk node sampling and message passing of graph neural networks. The user time-series behavior features, the user text intention features, and the user behavior link features are transformed and normalized through a multi-layer perceptron to obtain user fusion features including:

[0070] Collecting the behavior data of private domain live broadcast users, the behavior data includes user viewing time data, like data, comment data, reward data, and sharing data, normalizing and converting the behavior data into numerical values ​​to generate preprocessed data;

[0071] Construct a query vector, a key vector and a value vector, calculate a similarity matrix between the query vector and the key vector, generate a multi-head attention weight according to the similarity matrix, perform weighted summation of the multi-head attention weight and the value vector to obtain an attention feature, construct a graph structure for the user comment nodes and behavior nodes in the preprocessed data, obtain graph attention features through node aggregation operations, construct a user-product heterogeneous graph, and extract heterogeneous graph features through a node-level attention layer and a semantic-level attention layer;

[0072] Input the attention feature, the graph attention feature and the heterogeneous graph feature into a bidirectional long short-term memory network, extract bidirectional time series information through a forward propagation layer and a back propagation layer, control the update and forgetting of information flow through an update gate and a reset gate, calculate a time decay factor based on a time interval, and weight features at different time steps to obtain user time series behavior features;

[0073] Perform word segmentation and word vector conversion on the user comment text in the preprocessed data, input the word vector into the backbone network for feature extraction, retain the original features through skip connections, construct a contrastive learning loss function, minimize the feature distance of the comment text corresponding to different behaviors of the same user, maximize the feature distance of the comment text of different users, and obtain the user text intention feature;

[0074] Extracting a user behavior sequence from the preprocessed data, constructing a user behavior directed graph, performing temporal position encoding on nodes in the user behavior directed graph, performing random walk sampling on the user behavior directed graph based on an alias sampling table, generating a user behavior subgraph, performing message transmission and feature aggregation on the user behavior subgraph, and obtaining user behavior link features;

[0075] The user temporal behavior features, the user text intention features and the user behavior link features are input into the feature mapping network for dimension conversion, and the converted features are batch normalized. The nonlinear representation of features is introduced through a nonlinear activation function, the processed features are spliced ​​through a residual connection, the feature attention weights are calculated and weighted fused with the spliced ​​features to obtain user fused features.

[0076] The behavior node refers to a node in a graph, which represents a certain behavior or activity, and is usually associated with the user's behavior record. The user-product heterogeneous graph is a graph structure in which nodes are divided into different types, such as user nodes and product nodes, and edges represent the interactive relationship between users and products. The jump connection is a connection method in a neural network, which allows information to be transmitted across layers in the network to avoid information loss in deep networks. It is usually used to improve the training effect of deep networks and is commonly seen in architectures such as residual networks (ResNet). The temporal position encoding is used to represent the position of each element in a sequence when processing sequence data. The alias sampling table is a technique for accelerating sampling and is often used to deal with node sampling problems in large-scale graph data. By pre-calculating a mapping table, alias sampling can sample from a probability distribution with approximately constant time complexity without performing complex probability calculations each time. The user behavior subgraph is a subgraph extracted from a complete user-behavior graph, containing behavior data of a specific user.

[0077] Collect the behavioral data of private domain live broadcast users, including user viewing time, number of likes, comment content, reward amount, number of shares, etc. For example, user A watched for 30 minutes, liked twice, commented "the anchor's explanation is very clear", rewarded 10 yuan, and shared once; user B watched for 10 minutes, without likes, comments, rewards, or sharing. Normalize these data, such as converting viewing time into a value between 0 and 1, converting the number of likes into like frequency, and converting the reward amount into reward level. And convert non-numerical data, such as comment content, into numerical data, such as using word vector representation. Finally, generate preprocessed data.

[0078] Construct a multi-head self-attention mechanism. Convert the preprocessed data into query vector, key vector and value vector respectively. For example, the viewing time data of user A, 0.8, is converted into query vector [0.7, 0.1, 0.2], key vector [0.2, 0.6, 0.2], and value vector [0.5, 0.3, 0.2]. Calculate the similarity matrix between the query vector and the key vector. For example, the similarity between the query vector and the key vector of user A's viewing time is 0.6. Generate multi-head attention weights according to the similarity matrix. Perform weighted summation of the multi-head attention weights and the value vector to obtain the attention feature. At the same time, construct a graph structure based on user comment nodes and behavior nodes. For example, user A's comment "The anchor's explanation is very clear" can be connected to behavior nodes such as "viewing time" and "reward". Obtain graph attention features through node aggregation operations. In addition, construct a user-product heterogeneous graph. For example, if user A watches the live broadcast of product C, there is a connection between user A and product C. Extract heterogeneous graph features through node-level attention layer and semantic-level attention layer.

[0079] The attention features, graph attention features, and heterogeneous graph features are input into the bidirectional long short-term memory network, and the bidirectional temporal information is extracted through the forward propagation layer and the back propagation layer. The update gate and the reset gate are used to control the update and forgetting of the information flow. For example, the viewing time data of user A at different time points will be input into the bidirectional long short-term memory network in chronological order, and the time decay factor will be calculated based on the time interval. For example, the weight of user A's viewing behavior three days ago is lower than that of yesterday's viewing behavior. The features at different time steps are weighted to obtain the user's temporal behavior features.

[0080] Perform word segmentation and word vector conversion on the user comment text in the preprocessed data. For example, user A's comment "The anchor's explanation is very clear" is segmented into "anchor", "explanation", "very", "clear", and converted into corresponding word vectors. Input the word vector into the backbone network for feature extraction, while retaining the original features through jump connections. Construct a contrastive learning loss function to minimize the feature distance between the comment texts corresponding to different behaviors of the same user, such as minimizing the feature distance between user A's "like" behavior and the comment "The anchor's explanation is very clear". Maximize the feature distance between different users' comment texts, such as maximizing the feature distance between the comments of user A and user B. Obtain the user text intention feature.

[0081] Extract user behavior sequences from preprocessed data, for example, user A's behavior sequence is "watch" - "like" - "comment" - "reward". Construct a directed graph of user behavior, for example, from "watch" to "like". Encode the temporal position of the nodes in the graph. Perform random walk sampling on the user behavior directed graph based on the alias sampling table to generate a user behavior subgraph. Perform message passing and feature aggregation on the subgraph to obtain the user behavior link features.

[0082] Input the user temporal behavior features, user text intent features, and user behavior link features into the feature mapping network for dimension conversion. Perform batch normalization on the converted features. Introduce nonlinear representation of features through nonlinear activation functions. Concatenate the processed features through residual connections. Calculate the feature attention weights and perform weighted fusion with the concatenated features to obtain user fusion features.

[0083] In this embodiment, by integrating multi-dimensional user behavior data, including viewing, liking, commenting, rewarding, sharing, etc., and combining multimodal deep learning technology, it is possible to more comprehensively capture user behavior characteristics and avoid the limitations of a single data source. By adopting advanced technologies such as multi-head self-attention mechanism, contrastive learning, and graph neural network, it is possible to more accurately extract user temporal behavior characteristics, text intent characteristics, and behavior link characteristics, effectively improving the accuracy of user feature expression. Through more comprehensive and accurate user feature expression, it is possible to better understand user interests and needs, thereby improving the accuracy and effectiveness of personalized recommendations, and ultimately improving user satisfaction and platform revenue.

[0084] In an optional embodiment,

[0085] The time decay factor is calculated based on the time interval as shown in the following formula:

[0086] ;

[0087] in, SS represents the time decay factor, σ represents the standard deviation of the Gaussian decay function, w 1 represents the time decay weight in the Sigmoid decay function, b 1 represents the bias term of the Sigmoid decay function, β Represents the learnable Gaussian attenuation weight coefficient, △ t represents the time interval, exp() represents the exponential function, γ represents the contribution weight corresponding to the hyperbolic tangent function, w 2 represents the decay rate weight coefficient in the hyperbolic tangent function, b 2 Represents the bias term for the hyperbolic tangent function.

[0088] In this embodiment, three attenuation modes are comprehensively considered, which can more accurately reflect the changing pattern of the influence of events over time. Through learnable parameters, the attenuation speed and weight can be adjusted according to different scenarios to improve the adaptability of the model. By analyzing the contribution of different attenuation functions, the inherent mechanism of time attenuation can be better understood.

[0089] S2. Based on the user fusion features, a Gaussian mixture clustering model is constructed, users are adaptively stratified by calculating the inter-class distance distribution, a deep cross neural network is constructed, the user fusion features are input into the deep cross neural network, the feature combination relationship is extracted through the feature interaction layer, the feature expression is optimized through the residual network layer, the user churn probability value is obtained, a recurrent neural network prediction model is constructed, the user's historical purchase data is input into the recurrent neural network prediction model, a multi-task learning framework is used to simultaneously predict the user's future activity and purchasing power score, and a drainage weight coefficient is generated based on the user churn probability value and the user's future activity and purchasing power score;

[0090] The inter-class distance distribution refers to the distribution of distances or similarities between different categories in a multi-class classification task. Calculating the distance distribution between samples of different categories can help the model understand the distribution characteristics of each category and help improve classification performance, especially when facing data imbalance or category overlap. The deep cross neural network is a neural network structure that performs complex feature learning by crossing features of multiple layers. The user churn probability value refers to the probability of predicting that a user may churn (i.e. stop using a product or service) in user behavior analysis.

[0091] In an optional embodiment,

[0092] Based on the user fusion features, a Gaussian mixture clustering model is constructed, and users are adaptively stratified by calculating the inter-class distance distribution, and a deep cross neural network is constructed. The user fusion features are input into the deep cross neural network, and the feature combination relationship is extracted through the feature interaction layer. The feature expression is optimized through the residual network layer to obtain the user churn probability value, and a recurrent neural network prediction model is constructed. The user's historical purchase data is input into the recurrent neural network prediction model, and the user's future activity and purchasing power scores are simultaneously predicted using a multi-task learning framework. The drainage weight coefficient is generated based on the user churn probability value and the user's future activity and purchasing power scores, including:

[0093] Constructing a Gaussian mixture clustering model, generating an independent Gaussian distribution model for each pre-set user subgroup, each of the Gaussian distribution models including a mean vector, a covariance matrix and a mixing coefficient, constructing an expectation maximization iterative optimization module, inputting user fusion features into the Gaussian mixture clustering model, iteratively optimizing the parameters of the Gaussian distribution model by maximizing the log-likelihood function through the expectation maximization iterative optimization module, calculating the posterior probability that the sample point belongs to the user subgroup based on the Mahalanobis distance between the sample point and the center of the user subgroup, dividing the sample point into the user subgroup with the largest posterior probability, and obtaining a user stratification result;

[0094] Constructing a deep cross neural network, wherein the feature embedding module of the deep cross neural network includes a plurality of parallel embedding layers, the feature cross module of the deep cross neural network includes a plurality of cross layers connected in series, and the feature optimization module of the deep cross neural network includes a plurality of residual layers connected in series;

[0095] Input the user fusion feature into the embedding layer of the feature embedding module to map the sparse feature into a dense vector, input the dense vector into the cross layer of the feature cross module to perform a pairwise cross operation on different features to generate a high-order interactive feature, input the high-order interactive feature into the residual layer of the feature optimization module to perform element-wise addition on the dense vector to obtain an optimized feature, and pass the optimized feature through a fully connected layer and a Sigmoid activation function to obtain a user churn probability value;

[0096] Constructing a recurrent neural network prediction model, wherein the sequence processing module of the recurrent neural network prediction model includes a plurality of gated recurrent units connected in series, each of the gated recurrent units includes an update gate, a reset gate and a candidate hidden state, and the multi-task prediction module of the recurrent neural network prediction model includes a feature sharing layer and two task-specific layers;

[0097] Arrange the user's historical purchase data in chronological order to form a purchase sequence, input the purchase sequence into the sequence processing module, the gated recurrent unit updates the current hidden state according to the current purchase behavior and the historical hidden state of the purchase sequence, input the current hidden state into the multi-task prediction module, the feature sharing layer extracts shared features, the two task-specific layers are respectively connected to the user activity prediction task and the user purchasing power prediction task, and the user's future activity score and the user's future purchasing power score are obtained by jointly training the user activity prediction task and the user purchasing power prediction task;

[0098] A user diversion weight coefficient is calculated according to the user stratification result, the user churn probability value, the user future activity score and the user future purchasing power score. For a user subgroup in the user stratification result whose churn risk is higher than a preset churn risk threshold, the user diversion weight coefficient is increased based on the user churn probability value. For users whose user future activity score is lower than a preset activity threshold or whose user future purchasing power score is lower than a preset purchasing power score threshold, the value of the user diversion weight coefficient is increased, wherein the value range of the user diversion weight coefficient is 0 to 1.

[0099] The user subgroup refers to different groups in user data that users are divided into based on certain common features (such as behavior patterns, interests, regions, etc.). The Gaussian mixture clustering model is a probabilistic model based on the Gaussian distribution assumption, which is used to cluster data. It assumes that the data set consists of multiple Gaussian distribution subsets, and the data clustering is achieved by estimating the probability that each data point belongs to each Gaussian distribution. It is suitable for processing data sets with different distributions. The Mahalanobis distance is a method of measuring the distance between two points, taking into account the covariance matrix of the data. Compared with the Euclidean distance, the Mahalanobis distance can better handle the correlation between data, and is particularly suitable for calculating distribution differences in multidimensional data.

[0100] Stratify users. Use the Gaussian mixture clustering model to group users with similar features into the same group. Specifically, assume that user features include age, gender, and amount of consumption. First, initialize multiple Gaussian distribution models, each of which contains a mean vector, a covariance matrix, and a mixture coefficient. For example, initialize three Gaussian distributions to represent low-consumption, medium-consumption, and high-consumption user groups. Then, use the expectation-maximization algorithm to iteratively optimize the parameters of these Gaussian distribution models. In each iteration, calculate the posterior probability that each user belongs to each user group, which is based on the Mahalanobis distance between the user feature and the center of the user group. For example, the probability that a user belongs to the low-consumption, medium-consumption, and high-consumption user groups is calculated to be 0.1, 0.3, and 0.6, respectively. Finally, divide the user into the user group with the largest posterior probability. In this example, the user will be classified into the high-consumption user group.

[0101] A deep cross neural network is constructed to predict the probability of user churn, which includes a feature embedding module, a feature cross module, and a feature optimization module. Taking a user as an example, its fusion features include age (25 years old), gender (female), and the amount of the most recent purchase (500 yuan). The feature embedding module converts these sparse features into dense vector representations. For example, age 25 is embedded as a 128-dimensional vector. Then, the feature cross module performs a pairwise cross operation on different features to generate high-order interactive features. For example, the cross feature of age and gender can represent the combined effect of age and gender. The feature optimization module uses a residual network structure to add high-order interactive features to the original dense vector to obtain an optimized feature expression. Finally, the user churn probability value is output through a fully connected layer and a Sigmoid activation function. For example, the churn probability value of this user is 0.2, indicating that the user has a 20% chance of churn.

[0102] Construct a recurrent neural network prediction model to predict the user's future activity and purchasing power. Assume that a user's historical purchase data is the purchase amount of the past three months, which is 100 yuan, 200 yuan and 300 yuan respectively. Input these data into the sequence processing module of the recurrent neural network in chronological order. This module consists of multiple gated recurrent units, each of which updates the current hidden state according to the current purchase behavior and historical hidden state. For example, the first unit processes a purchase amount of 100 yuan and generates a hidden state; the second unit processes a purchase amount of 200 yuan and the hidden state of the previous unit to generate a new hidden state, and so on. The final hidden state is input into the multi-task prediction module, which contains a feature sharing layer and two task-specific layers. The feature sharing layer extracts shared features, and the two task-specific layers predict the user's future activity score and purchasing power score respectively. For example, the user's future activity score is 0.8 and the purchasing power score is 0.7.

[0103] The user drainage weight coefficient is calculated based on the user stratification results, churn probability, future activity and purchasing power scores. Assume that the churn risk threshold is set to 0.3, the activity threshold is set to 0.5, and the purchasing power threshold is set to 0.6. For users who belong to the high churn risk user group, if their churn probability is higher than 0.3, their drainage weight coefficient is increased. For users whose future activity score is lower than 0.5 or whose future purchasing power score is lower than 0.6, their drainage weight coefficient is also increased. The value range of the drainage weight coefficient is 0 to 1. For example, a user who belongs to the high churn risk user group has a churn probability of 0.4 (higher than the threshold of 0.3), a future activity score of 0.4 (lower than the threshold of 0.5), and a future purchasing power score of 0.7, then his drainage weight coefficient will be increased, for example, set to 0.8.

[0104] In this embodiment, by stratifying and accurately predicting users, resources are concentrated on users with high churn risk, thus avoiding resource waste and improving traffic diversion efficiency. By identifying and intervening in high churn risk users in advance, the user churn rate is effectively reduced and the user scale is maintained. Based on user characteristics and behavior predictions, personalized traffic diversion strategies are formulated to improve user retention and activity.

[0105] In an optional embodiment,

[0106] The two task-specific layers are respectively connected to the user activity prediction task and the user purchasing power prediction task, and the user future activity score and the user future purchasing power score are obtained by jointly training the user activity prediction task and the user purchasing power prediction task, including:

[0107] Residual connections are introduced in the user activity prediction task-specific layer and the user purchasing power prediction task-specific layer, and the pre-acquired shared features are element-wise added to the output of each task-specific layer to obtain the fused shared features.

[0108] The user activity prediction task specific layer maps the fused shared features to the user's future activity score through a fully connected layer and a Sigmoid activation function, and the user purchasing power prediction task specific layer maps the fused features to the user's future purchasing power score through a fully connected layer and a linear activation function;

[0109] The multi-task prediction module is used as a generator network to construct a discriminator network, wherein the discriminator network receives the user's real future activity data, real future purchasing power data, and the prediction result of the generator network;

[0110] Constructing a multi-task learning joint loss function, the multi-task learning joint loss function includes a cross entropy loss term of a user activity prediction task and a mean square error loss term of a user purchasing power prediction task, adjusting weight coefficients of the cross entropy loss term of the user activity prediction task and the mean square error loss term of the user purchasing power prediction task by dynamic weighting, and dynamically updating the weight coefficients according to the convergence speed of the task;

[0111] The generator network and the discriminator network are alternately trained by a minimax algorithm, wherein the generator network minimizes the multi-task learning joint loss function, and the discriminator network maximizes the classification accuracy of the real data and the generated data, and the training is repeated until a preset maximum number of training times is reached;

[0112] Perform multiple Monte Carlo probability sampling on the user's recent purchase sequence, input the sampled sequence into the trained multi-task prediction module to obtain multiple groups of prediction results, and obtain the final user future activity score and user future purchasing power score by integrating the multiple groups of prediction results.

[0113] The cross layer is a technology used in deep neural networks to enhance the expressiveness of models. It refers to information crossing between different layers of the network, helping the model learn richer feature representations by sharing features or influencing each other. The multi-task learning joint loss function combines the loss functions of multiple related tasks and optimizes the objectives of all tasks simultaneously during training. The minimax algorithm is an optimization method used in game theory, usually applied to zero-sum games. The Monte Carlo probability sampling is a method of estimating complex probability distributions through random sampling. Through a large number of random sampling, the probability distribution or approximation of the expected value of a problem can be obtained. It is widely used in optimization, statistical analysis, physical simulation and other fields.

[0114] Pre-acquire the shared features of users. These shared features can include basic information, historical behavior data, social relationship data, etc. For example, the user's age, gender, registration time, browsing history, click records, number of friends, etc. can all be used as shared features. Specifically, we can convert the user's age into an age range, gender into a numerical code, registration time into a time interval from the current time, browsing history and click records into embedded vectors of products or categories, and the number of friends into a normalized numerical value.

[0115] Construct a multi-task prediction module. The task prediction module contains two task-specific layers, which are connected to the user activity prediction task and the user purchasing power prediction task respectively. Introduce residual connections in each task-specific layer, and add the pre-acquired shared features to the output of each task-specific layer element-wise to obtain fused shared features. The user activity prediction task-specific layer maps the fused shared features to the user's future activity score through a fully connected layer and a Sigmoid activation function, with a value range of 0 to 1. The user purchasing power prediction task-specific layer maps the fused features to the user's future purchasing power score through a fully connected layer and a linear activation function, with a value range of positive real numbers. For example, assuming that the fused shared feature is a 128-dimensional vector, the user activity prediction task-specific layer can include a 128-to-1 fully connected layer and a Sigmoid activation function, outputting a value between 0 and 1, indicating the probability of the user's future activity. The user purchasing power prediction task-specific layer can include a 128-to-1 fully connected layer and a linear activation function, outputting a positive real number, indicating the predicted value of the user's future purchasing power.

[0116] Construct a discriminator network and use the multi-task prediction module as the generator network. The discriminator network receives the user's real future activity data, real future purchasing power data, and the prediction results of the generator network. The goal of the discriminator network is to distinguish between real data and generated data. For example, the discriminator network can be a multi-layer perceptron. The input is the user's real future activity, real future purchasing power, and predicted activity and purchasing power. The output is a value between 0 and 1, indicating the probability that the input data is real data.

[0117] Construct a joint loss function for multi-task learning, including the cross entropy loss term for the user activity prediction task and the mean square error loss term for the user purchasing power prediction task. Adjust the weight coefficients of these two loss terms by dynamic weighting, and dynamically update the weight coefficients according to the convergence speed of the task. For example, the weight coefficients of the two loss terms can be set to 0.5 initially, and then the weight coefficients can be dynamically adjusted according to the changes in the training loss of each task, so that the task with a slower convergence speed gets a larger weight.

[0118] The generator network and the discriminator network are trained alternately through the minimax algorithm. The goal of the generator network is to minimize the joint loss function of multi-task learning, and the goal of the discriminator network is to maximize the classification accuracy of real data and generated data. Repeat the training until the preset maximum number of training times is reached, for example, 1000 times.

[0119] Perform multiple Monte Carlo probability sampling on the user's recent purchase sequence, for example, 10 times. Input the sampled sequence into the trained multi-task prediction module to obtain multiple sets of prediction results. By integrating multiple sets of prediction results, such as taking the average, the final user's future activity score and user's future purchasing power score are obtained. Assuming that the user recently purchased three products A, B, and C, the products that the user may purchase in the future can be sampled based on the user's historical purchase behavior, for example, sampling to obtain multiple sequences such as A, B, D, B, C, E, A, C, F, etc.

[0120] In this embodiment, through multi-task learning joint training, the model can learn the potential correlation between user activity and purchasing power, thereby improving the accuracy of prediction. The introduction of adversarial training mechanism enables the model to better cope with noise and abnormal data, enhancing the robustness of the model. The Monte Carlo sampling strategy can capture the randomness of user behavior, thereby achieving more personalized user behavior prediction.

[0121] S3. Construct a graph-attention product knowledge network, input the user fusion features into the graph-attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, and obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through the Thompson sampling method, and optimize the push strategy in combination with the Monte Carlo tree search method to obtain the optimal push strategy.

[0122] The graph attention product knowledge network is a model based on a graph neural network architecture, which is used for product recommendation and knowledge graph construction. The multi-hop entity relationship reasoning is a reasoning method based on a graph model, which aims to infer complex relationships between entities through multiple jumps (i.e., through multiple intermediate nodes). The difficult sample mining module is a technology that automatically identifies and pays attention to samples that are difficult for the model to correctly classify or predict during the training process. The contrastive learning negative sample queue is a queue used to store negative samples in the contrastive learning framework. The dual deep Q network is a reinforcement learning algorithm that solves the problem of overestimation of value in traditional Q-learning by introducing two Q networks. The Thompson sampling method is an algorithm for Bayesian optimization and multi-armed bandit problems. It selects the action that is currently considered to be optimal by sampling the posterior distribution of each action.

[0123] In an optional embodiment,

[0124] Construct a graph attention product knowledge network, input the user fusion features into the graph attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through a Thompson sampling method and optimize the push strategy in combination with a Monte Carlo tree search method to obtain the optimal push strategy, including:

[0125] Construct a graph attention product knowledge network, take product categories, brands, and functions as nodes, and take common attributes, complementary relationships, and substitution relationships between products as edges to form a heterogeneous graph network structure, use the user fusion features as query conditions, and obtain a matching product set through multi-hop entity relationship reasoning;

[0126] Constructing a dual-tower deep neural network, the dual-tower deep neural network comprising a user tower and a product tower, the user tower receiving the user fusion feature and mapping it to a user embedding vector, the product tower receiving the features of the products in the matching product set and mapping it to a product embedding vector;

[0127] Through a pre-set difficult sample mining module, select the product with the closest product embedding vector distance to the product actually interacted by the user from the matching product set as a difficult sample, add the difficult sample to an initial negative sample queue initialized to empty in advance, repeat the selection and addition to obtain the negative sample queue, and construct a contrast loss function based on the negative sample queue to train the dual-tower deep neural network;

[0128] Construct a variational autoencoder, map the user embedding vector and the product embedding vector output by the trained dual-tower deep neural network to a latent space distribution, reconstruct the user-product interaction distribution from the latent space distribution to obtain a product recommendation probability matrix, calculate the drainage weight coefficient corresponding to the category, price, and profit of each product in the matching product set, and multiply the product recommendation probability matrix by the drainage weight coefficient to obtain a final product ranking score;

[0129] An online policy network and a target network are constructed, wherein the online policy network receives the final product ranking score and outputs the Q value of the push policy. The target network periodically copies parameters from the online policy network to estimate the Q value of the state-action pair. The posterior probability distribution of each push policy is calculated according to the Q value output by the online policy network. The Thompson sampling method is used to sample and generate the push policy from the posterior probability distribution. A push decision search tree is constructed in combination with the Monte Carlo tree search method, wherein each node represents a push decision. Simulation selection and evaluation update are repeated to obtain a push decision sequence, which is applied to the live product recommendation and the hyperparameters of the online policy network are updated to obtain the optimal push policy.

[0130] The dual-tower deep neural network is a commonly used structure in a recommendation system. It learns the representations of users and products respectively through two independent neural networks (i.e., "towers"), and then interacts with these two representations to generate prediction results. The difficult samples refer to samples that are misclassified or predicted with low confidence by the model during the training process. The latent space distribution refers to the data distribution represented by the latent space (or latent space) in a generative model (such as a variational autoencoder or a generative adversarial network). The drainage weight coefficient refers to the relative weight used in the recommendation system to represent the drainage (attraction) of each recommended item to the user during the recommendation process. The Monte Carlo tree search method is a search algorithm for decision-making processes, which is particularly suitable for reinforcement learning problems with large-scale state spaces.

[0131] Construct a graph attention product knowledge network, with product categories, brands, functions, etc. as nodes, and common attributes between products (such as "electronic products"), complementary relationships (such as "mobile phones" and "mobile phone cases"), and substitution relationships (such as "iPhone14" and "Huawei P60") as edges, to form a heterogeneous graph network structure. Fusion of multi-dimensional features of users, such as historical browsing records, purchase records, collection records, demographic characteristics, etc., into user fusion feature vectors. Using the user fusion feature vector as the query condition, multi-hop entity relationship reasoning is performed on the graph attention product knowledge network. For example, if a user browses "sports shoes", the system will find products related to "sports shoes" along the edges in the graph, such as "sports socks", "sportswear", etc., and may even find "insoles" that complement specific brands of sports shoes or substitute "sports shoes" of other brands. Through multi-hop reasoning, the system can obtain a set of products that better match the user's interests. Suppose the user browses "laptop computer", through two-hop reasoning, related products such as "mouse", "keyboard", and "computer bag" can be obtained.

[0132] Construct an improved dual-tower deep neural network, including a user tower and a product tower. The user tower receives the user fusion feature vector as input and maps it to a user embedding vector. The product tower receives the features of each product in the matching product set obtained in the previous step, such as product title, description, price, sales volume, etc., and maps each product to a product embedding vector. For example, the user tower maps the user's historical browsing records "dress", "high heels", etc. into a vector, and the product tower maps products such as "red dress", "black high heels" into corresponding vectors.

[0133] In order to improve the recommendation effect of the model, a difficult sample mining module and contrastive learning are introduced. The product with the closest embedding vector distance to the product that the user actually interacts with is selected from the matching product set as a difficult sample. For example, if the user buys a "blue shirt", the system will find the product that is most similar to the "blue shirt" from the matching product set, such as "blue T-shirt" or "white shirt" as a difficult sample. These difficult samples are added to the initially empty negative sample queue. By continuously selecting and adding difficult samples, a negative sample queue is finally obtained. Based on the negative sample queue, a contrast loss function is constructed for training the dual-tower deep neural network.

[0134] In order to learn the user-product interaction distribution, a variational autoencoder is constructed. The user embedding vector and product embedding vector output by the trained dual-tower deep neural network are mapped to the latent space distribution. The user-product interaction distribution is reconstructed from the latent space distribution to obtain the product recommendation probability matrix. For example, if the user embedding vector and the product embedding vector are close in the latent space, the recommendation probability of the product will be higher. At the same time, the drainage weight coefficient corresponding to the category, price, and profit of each product in the matching product set is calculated. For example, the drainage weight coefficient of "new mobile phone" may be higher than that of "old mobile phone". Multiply the product recommendation probability matrix by the drainage weight coefficient to obtain the final product ranking score.

[0135] Construct a push strategy optimization module based on reinforcement learning, and construct an online strategy network and a target network. The online strategy network receives the final product ranking score and outputs the Q values ​​of different push strategies, such as the Q values ​​of strategies such as "push by score sorting" and "push by price sorting". The target network regularly copies parameters from the online strategy network to estimate the Q value of the state-action pair. According to the Q value output by the online strategy network, the posterior probability distribution of each push strategy is calculated. The Thompson sampling method is used to sample and generate push strategies from the posterior probability distribution. For example, if the Q value of "push by score sorting" is the highest, the probability of this strategy being selected is the highest. Combined with the Monte Carlo tree search method, a push decision search tree is constructed, and each node represents a push decision, such as which product to push and in what order to push. Repeat the simulation selection and evaluation update to obtain the push decision sequence, and apply the sequence to the live product recommendation, and update the hyperparameters of the online strategy network to finally obtain the optimal push strategy. For example, after multiple simulations, the system may find that the best effect is to push "high-profit products" interspersed between "high-click rate products".

[0136] In this embodiment, the graph attention product knowledge network can capture the complex relationship between products and achieve more accurate product matching, thereby improving the accuracy of recommendations. The introduction of difficult sample mining and comparative learning enables the model to better learn the essential characteristics of user-product interaction, improve the robustness of the model, and reduce the risk of overfitting. The push strategy optimization module based on reinforcement learning can dynamically adjust the push strategy according to real-time user feedback to maximize the recommendation effect.

[0137] In an optional embodiment,

[0138] From the set of matching products, select the product whose product embedding vector is closest to the product that the user actually interacts with as a difficult sample as shown in the following formula:

[0139] ;

[0140] in, DIndicates the actual interaction product of the user u and products to be matched v The distance between T represents transpose, θ represents the smoothing coefficient, Indicated in t The covariance matrix at -1 time step, h Indicates t The covariance contribution of the newly added samples is W a represents the weight matrix, l a Represents the bias term.

[0141] In this embodiment, by selecting products that are close to but not completely the same as the products that the user actually interacts with as negative samples, the recommendation model can be better trained to distinguish user preferences, thereby improving the accuracy of recommendations. The dynamic covariance adjustment mechanism can avoid always selecting very similar products as negative samples, thereby increasing the diversity of recommendation results and improving user experience. Dynamically updating the covariance matrix can enable the model to adapt to changing user behavior and product information, thereby improving the robustness of the model.

[0142] Figure 2 Schematic diagram of the structure of a private domain live broadcast user behavior analysis and traffic recommendation system based on deep learning in an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0143] The first unit is used to collect the behavior data of private domain live broadcast users, build a multimodal deep neural network based on a multi-head self-attention mechanism, add the behavior data to the multimodal deep neural network, perform time-series weighted processing through a bidirectional long short-term memory network module in combination with a gated recursive unit and a time decay factor to obtain user time-series behavior features, extract user comment text from the behavior data and add it to a residual connection structure module, obtain user text intention features through contrastive learning processing, extract user behavior paths from the behavior data and add them to a time-series graph attention network module, obtain user behavior link features through random walk node sampling and message passing of a graph neural network, perform feature transformation and normalization on the user time-series behavior features, the user text intention features and the user behavior link features through a multi-layer perceptron to obtain user fusion features;

[0144] The second unit is used to construct a Gaussian mixture clustering model based on the user fusion features, adaptively stratify users by calculating the inter-class distance distribution, construct a deep cross neural network, input the user fusion features into the deep cross neural network, extract the feature combination relationship through the feature interaction layer, optimize the feature expression through the residual network layer, obtain the user churn probability value, construct a recurrent neural network prediction model, input the user's historical purchase data into the recurrent neural network prediction model, use a multi-task learning framework to simultaneously predict the user's future activity and purchasing power score, and generate a drainage weight coefficient based on the user churn probability value and the user's future activity and purchasing power score;

[0145] The third unit is used to construct a graph attention product knowledge network, input the user fusion features into the graph attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through the Thompson sampling method and optimize the push strategy in combination with the Monte Carlo tree search method to obtain the optimal push strategy.

[0146] According to a third aspect of the embodiments of the present invention,

[0147] An electronic device is provided, comprising:

[0148] processor;

[0149] a memory for storing processor-executable instructions;

[0150] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0151] A fourth aspect of the embodiments of the present invention is:

[0152] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0153] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing user behavior and traffic diversion in private live streaming based on deep learning, characterized in that: include: Collect the behavior data of private domain live broadcast users, build a multimodal deep neural network based on the multi-head self-attention mechanism, add the behavior data to the multimodal deep neural network, perform time-series weighted processing through a bidirectional long short-term memory network module combined with a gated recursive unit and a time decay factor to obtain user time-series behavior features, extract user comment text from the behavior data and add it to the residual connection structure module, obtain user text intention features through contrastive learning processing, extract user behavior paths from the behavior data and add them to the time-series graph attention network module, obtain user behavior link features through random walk node sampling and graph neural network message passing, perform feature transformation and normalization on the user time-series behavior features, the user text intention features and the user behavior link features through a multi-layer perceptron to obtain user fusion features; Based on the user fusion features, a Gaussian mixture clustering model is constructed, users are adaptively stratified by calculating the inter-class distance distribution, a deep cross neural network is constructed, the user fusion features are input into the deep cross neural network, the feature combination relationship is extracted through the feature interaction layer, the feature expression is optimized through the residual network layer, the user churn probability value is obtained, a recurrent neural network prediction model is constructed, the user's historical purchase data is input into the recurrent neural network prediction model, the user's future activity and purchasing power score are simultaneously predicted using a multi-task learning framework, and a drainage weight coefficient is generated based on the user churn probability value and the user's future activity and purchasing power score; Construct a graph attention product knowledge network, input the user fusion features into the graph attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, and obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through the Thompson sampling method, and optimize the push strategy in combination with the Monte Carlo tree search method to obtain the optimal push strategy.

2. The method according to claim 1, characterized in that The behavior data of private domain live broadcast users is collected, and a multimodal deep neural network is constructed based on a multi-head self-attention mechanism. The behavior data is added to the multimodal deep neural network, and time-series weighted processing is performed through a bidirectional long short-term memory network module in combination with a gated recursive unit and a time decay factor to obtain user time-series behavior features. User comment text is extracted from the behavior data and added to the residual connection structure module. User text intention features are obtained through contrastive learning processing. User behavior paths are extracted from the behavior data and added to the time-series graph attention network module. User behavior link features are obtained through random walk node sampling and message passing of graph neural networks. The user time-series behavior features, the user text intention features, and the user behavior link features are transformed and normalized through a multi-layer perceptron to obtain user fusion features including: Collecting the behavior data of private domain live broadcast users, the behavior data includes user viewing time data, like data, comment data, reward data, and sharing data, normalizing and converting the behavior data into numerical values ​​to generate preprocessed data; Construct a query vector, a key vector and a value vector, calculate a similarity matrix between the query vector and the key vector, generate a multi-head attention weight according to the similarity matrix, perform weighted summation of the multi-head attention weight and the value vector to obtain an attention feature, construct a graph structure for the user comment nodes and behavior nodes in the preprocessed data, obtain graph attention features through node aggregation operations, construct a user-product heterogeneous graph, and extract heterogeneous graph features through a node-level attention layer and a semantic-level attention layer; Input the attention feature, the graph attention feature and the heterogeneous graph feature into a bidirectional long short-term memory network, extract bidirectional time series information through a forward propagation layer and a back propagation layer, control the update and forgetting of information flow through an update gate and a reset gate, calculate a time decay factor based on a time interval, and weight features at different time steps to obtain user time series behavior features; Perform word segmentation and word vector conversion on the user comment text in the preprocessed data, input the word vector into the backbone network for feature extraction, retain the original features through skip connections, construct a contrastive learning loss function, minimize the feature distance of the comment text corresponding to different behaviors of the same user, maximize the feature distance of the comment text of different users, and obtain the user text intention feature; Extracting a user behavior sequence from the preprocessed data, constructing a user behavior directed graph, performing temporal position encoding on nodes in the user behavior directed graph, performing random walk sampling on the user behavior directed graph based on an alias sampling table, generating a user behavior subgraph, performing message transmission and feature aggregation on the user behavior subgraph, and obtaining user behavior link features; The user temporal behavior features, the user text intention features and the user behavior link features are input into the feature mapping network for dimension conversion, and the converted features are batch normalized. The nonlinear representation of features is introduced through a nonlinear activation function, the processed features are spliced ​​through a residual connection, the feature attention weights are calculated and weighted fused with the spliced ​​features to obtain user fused features.

3. The method according to claim 2, characterized in that The time decay factor is calculated based on the time interval as shown in the following formula: ； in, SS represents the time decay factor, σ represents the standard deviation of the Gaussian decay function, w 1 Represents the time decay weight in the Sigmoid decay function, b 1 represents the bias term of the Sigmoid decay function, β Represents the learnable Gaussian attenuation weight coefficient, △ t represents the time interval, exp() represents the exponential function, γ represents the contribution weight corresponding to the hyperbolic tangent function, w 2 represents the decay rate weight coefficient in the hyperbolic tangent function, b 2 Represents the bias term for the hyperbolic tangent function.

4. The method according to claim 1, characterized in that: Based on the user fusion features, a Gaussian mixture clustering model is constructed, and users are adaptively stratified by calculating the inter-class distance distribution, and a deep cross neural network is constructed. The user fusion features are input into the deep cross neural network, and the feature combination relationship is extracted through the feature interaction layer. The feature expression is optimized through the residual network layer to obtain the user churn probability value, and a recurrent neural network prediction model is constructed. The user's historical purchase data is input into the recurrent neural network prediction model, and the user's future activity and purchasing power scores are simultaneously predicted using a multi-task learning framework. The drainage weight coefficient is generated based on the user churn probability value and the user's future activity and purchasing power scores, including: Constructing a Gaussian mixture clustering model, generating an independent Gaussian distribution model for each pre-set user subgroup, each of the Gaussian distribution models including a mean vector, a covariance matrix and a mixing coefficient, constructing an expectation maximization iterative optimization module, inputting user fusion features into the Gaussian mixture clustering model, iteratively optimizing the parameters of the Gaussian distribution model by maximizing the log-likelihood function through the expectation maximization iterative optimization module, calculating the posterior probability that the sample point belongs to the user subgroup based on the Mahalanobis distance between the sample point and the center of the user subgroup, dividing the sample point into the user subgroup with the largest posterior probability, and obtaining a user stratification result; Constructing a deep cross neural network, wherein the feature embedding module of the deep cross neural network includes a plurality of parallel embedding layers, the feature cross module of the deep cross neural network includes a plurality of cross layers connected in series, and the feature optimization module of the deep cross neural network includes a plurality of residual layers connected in series; Input the user fusion feature into the embedding layer of the feature embedding module to map the sparse feature into a dense vector, input the dense vector into the cross layer of the feature cross module to perform a pairwise cross operation on different features to generate a high-order interactive feature, input the high-order interactive feature into the residual layer of the feature optimization module to perform element-wise addition on the dense vector to obtain an optimized feature, and pass the optimized feature through a fully connected layer and a Sigmoid activation function to obtain a user churn probability value; Constructing a recurrent neural network prediction model, wherein the sequence processing module of the recurrent neural network prediction model includes a plurality of gated recurrent units connected in series, each of the gated recurrent units includes an update gate, a reset gate and a candidate hidden state, and the multi-task prediction module of the recurrent neural network prediction model includes a feature sharing layer and two task-specific layers; Arrange the user's historical purchase data in chronological order to form a purchase sequence, input the purchase sequence into the sequence processing module, the gated recurrent unit updates the current hidden state according to the current purchase behavior and the historical hidden state of the purchase sequence, input the current hidden state into the multi-task prediction module, the feature sharing layer extracts shared features, the two task-specific layers are respectively connected to the user activity prediction task and the user purchasing power prediction task, and the user's future activity score and the user's future purchasing power score are obtained by jointly training the user activity prediction task and the user purchasing power prediction task; A user diversion weight coefficient is calculated according to the user stratification result, the user churn probability value, the user future activity score and the user future purchasing power score. For a user subgroup in the user stratification result whose churn risk is higher than a preset churn risk threshold, the user diversion weight coefficient is increased based on the user churn probability value. For users whose user future activity score is lower than a preset activity threshold or whose user future purchasing power score is lower than a preset purchasing power score threshold, the value of the user diversion weight coefficient is increased, wherein the value range of the user diversion weight coefficient is 0 to 1.

5. The method according to claim 4, characterized in that The two task-specific layers are respectively connected to the user activity prediction task and the user purchasing power prediction task, and the user future activity score and the user future purchasing power score are obtained by jointly training the user activity prediction task and the user purchasing power prediction task, including: Residual connections are introduced in the user activity prediction task-specific layer and the user purchasing power prediction task-specific layer, and the pre-acquired shared features are element-wise added to the output of each task-specific layer to obtain the fused shared features. The user activity prediction task specific layer maps the fused shared features to the user's future activity score through a fully connected layer and a Sigmoid activation function, and the user purchasing power prediction task specific layer maps the fused features to the user's future purchasing power score through a fully connected layer and a linear activation function; The multi-task prediction module is used as a generator network to construct a discriminator network, wherein the discriminator network receives the user's real future activity data, real future purchasing power data, and the prediction result of the generator network; Constructing a multi-task learning joint loss function, the multi-task learning joint loss function includes a cross entropy loss term of a user activity prediction task and a mean square error loss term of a user purchasing power prediction task, adjusting weight coefficients of the cross entropy loss term of the user activity prediction task and the mean square error loss term of the user purchasing power prediction task by dynamic weighting, and dynamically updating the weight coefficients according to the convergence speed of the task; The generator network and the discriminator network are alternately trained by a minimax algorithm, wherein the generator network minimizes the multi-task learning joint loss function, and the discriminator network maximizes the classification accuracy of the real data and the generated data, and the training is repeated until a preset maximum number of training times is reached; Perform multiple Monte Carlo probability sampling on the user's recent purchase sequence, input the sampled sequence into the trained multi-task prediction module to obtain multiple groups of prediction results, and obtain the final user future activity score and user future purchasing power score by integrating the multiple groups of prediction results.

6. The method according to claim 1, characterized in that Construct a graph attention product knowledge network, input the user fusion features into the graph attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through a Thompson sampling method and optimize the push strategy in combination with a Monte Carlo tree search method to obtain the optimal push strategy, including: Construct a graph attention product knowledge network, take product categories, brands, and functions as nodes, and take common attributes, complementary relationships, and substitution relationships between products as edges to form a heterogeneous graph network structure, use the user fusion features as query conditions, and obtain a matching product set through multi-hop entity relationship reasoning; Constructing a dual-tower deep neural network, the dual-tower deep neural network comprising a user tower and a product tower, the user tower receiving the user fusion feature and mapping it to a user embedding vector, the product tower receiving the features of the products in the matching product set and mapping it to a product embedding vector; Through a pre-set difficult sample mining module, select the product with the closest product embedding vector distance to the product actually interacted by the user from the matching product set as a difficult sample, add the difficult sample to an initial negative sample queue initialized to empty in advance, repeat the selection and addition to obtain the negative sample queue, and construct a contrast loss function based on the negative sample queue to train the dual-tower deep neural network; Construct a variational autoencoder, map the user embedding vector and the product embedding vector output by the trained dual-tower deep neural network to a latent space distribution, reconstruct the user-product interaction distribution from the latent space distribution to obtain a product recommendation probability matrix, calculate the drainage weight coefficient corresponding to the category, price, and profit of each product in the matching product set, and multiply the product recommendation probability matrix by the drainage weight coefficient to obtain a final product ranking score; An online policy network and a target network are constructed, wherein the online policy network receives the final product ranking score and outputs the Q value of the push policy. The target network periodically copies parameters from the online policy network to estimate the Q value of the state-action pair. The posterior probability distribution of each push policy is calculated according to the Q value output by the online policy network. The Thompson sampling method is used to sample and generate the push policy from the posterior probability distribution. A push decision search tree is constructed in combination with the Monte Carlo tree search method, wherein each node represents a push decision. Simulation selection and evaluation update are repeated to obtain a push decision sequence, which is applied to the live product recommendation and the hyperparameters of the online policy network are updated to obtain the optimal push policy.

7. The method according to claim 6, characterized in that From the set of matching products, select the product whose product embedding vector is closest to the product that the user actually interacts with as a difficult sample as shown in the following formula: ； in, D Indicates the actual interaction product of the user u and products to be matched v The distance between T represents transpose, θ represents the smoothing coefficient, Indicated in t The covariance matrix at -1 time step, h Indicates t The covariance contribution of the newly added samples is W a represents the weight matrix, l a Represents the bias term.

8. A private domain live broadcast user behavior analysis and traffic recommendation system based on deep learning, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect the behavior data of private domain live broadcast users, build a multimodal deep neural network based on a multi-head self-attention mechanism, add the behavior data to the multimodal deep neural network, perform time-series weighted processing through a bidirectional long short-term memory network module in combination with a gated recursive unit and a time decay factor to obtain user time-series behavior features, extract user comment text from the behavior data and add it to a residual connection structure module, obtain user text intention features through contrastive learning processing, extract user behavior paths from the behavior data and add them to a time-series graph attention network module, obtain user behavior link features through random walk node sampling and message passing of a graph neural network, perform feature transformation and normalization on the user time-series behavior features, the user text intention features and the user behavior link features through a multi-layer perceptron to obtain user fusion features; The second unit is used to construct a Gaussian mixture clustering model based on the user fusion features, adaptively stratify users by calculating the inter-class distance distribution, construct a deep cross neural network, input the user fusion features into the deep cross neural network, extract the feature combination relationship through the feature interaction layer, optimize the feature expression through the residual network layer, obtain the user churn probability value, construct a recurrent neural network prediction model, input the user's historical purchase data into the recurrent neural network prediction model, use a multi-task learning framework to simultaneously predict the user's future activity and purchasing power score, and generate a drainage weight coefficient based on the user churn probability value and the user's future activity and purchasing power score; The third unit is used to construct a graph attention product knowledge network, input the user fusion features into the graph attention product knowledge network, obtain a set of matching products through multi-hop entity relationship reasoning, construct an improved dual-tower deep neural network and use the user fusion features and the matching product set as input data, construct a contrast learning negative sample queue through a difficult sample mining module, learn the user-product interaction distribution through a variational autoencoder module, obtain a product recommendation probability matrix, perform weighted calculation on the product recommendation probability matrix and the drainage weight coefficient to obtain a final product ranking score, input the final product ranking score into a pre-constructed dual deep Q network, select a push strategy through the Thompson sampling method and optimize the push strategy in combination with the Monte Carlo tree search method to obtain the optimal push strategy.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images