A live broadcast recommendation method based on multiple repeated behaviors

Abstract

The present invention discloses a live broadcast recommendation method based on multiple repeated behaviors. First, a live broadcast recommendation model based on multiple repeated behaviors is constructed, and the behavior data of users in the live broadcast scene are collected. Three bipartite graphs are constructed based on the collected behavior data, which respectively represent the user's behavior of entering the live broadcast room, chatting behavior, and gift giving behavior; multi-behavior modeling based on graph neural network, using improved graph attention network to learn the repeated interactive relationship between users and live broadcast rooms; integrating user and live broadcast room representations generated by different behaviors through embedding fusion strategy; adopting multi-task learning framework to train and predict the live broadcast recommendation model, and generating a live broadcast recommendation list according to the model prediction results. The above method can solve the problems of insufficient modeling of multiple behaviors of users in live broadcast scenes, limited use of repeated behaviors, and poor recommendation effect in the prior art, and can comprehensively characterize user preferences and improve the accuracy of live broadcast room recommendations and user experience.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and recommendation systems, and in particular to a live broadcast recommendation method based on multiple repeated behaviors. Background Art

[0002] With the rapid development of Internet technology, live broadcast platforms have become an important form of entertainment and social interaction. The interaction between users and anchors has become more and more frequent. Personalized live broadcast recommendations are crucial to improving user experience and platform revenue. However, traditional recommendation systems fail to effectively model the complex and repetitive behaviors of users in live broadcast scenarios, such as users repeatedly entering the same live broadcast room, frequently sending chat messages, and giving virtual gifts. These repetitive behaviors reflect the different degrees of users' preferences and participation in live broadcast content. Existing live broadcast recommendation methods mainly focus on the following aspects:

[0003] 1) Single behavior modeling: Existing methods usually only consider a single user behavior (such as clicking, gifting), ignoring the combined impact of multiple behaviors, making it difficult to fully characterize user preferences;

[0004] 2) Data sparsity problem: In live broadcast scenarios, user behavior sparsity and cold start problems are relatively serious, and single behavior modeling is difficult to meet actual needs;

[0005] 3) Insufficient modeling of repeated behaviors: Users’ behaviors in live broadcast scenarios are repetitive (such as entering the live broadcast room multiple times, chatting, or sending gifts). These repetitive behaviors can significantly reflect the users’ continued interests, but existing methods fail to fully utilize this information.

[0006] Most existing live broadcast recommendation methods only consider single behaviors and ignore users' repeated behaviors, which leads to incomplete modeling of user preferences and affects the recommendation effect. Therefore, there is an urgent need for an innovative method that can effectively utilize multiple users' repeated behaviors and improve the live broadcast recommendation effect, which is of great significance for improving user experience and increasing platform revenue. Summary of the invention

[0007] The purpose of the present invention is to provide a live broadcast recommendation method based on multiple repeated behaviors, which can solve the problems in the prior art of insufficient modeling of multiple behaviors of users in live broadcast scenarios, limited utilization of repeated behaviors, and poor recommendation effect. It can comprehensively characterize user preferences and improve the accuracy of live broadcast room recommendations and user experience.

[0008] The objective of the present invention is achieved through the following technical solutions:

[0009] A live broadcast recommendation method based on multiple repeated behaviors, the method comprising:

[0010] Step 1: First, build a live broadcast recommendation model based on multiple repeated behaviors, collect user behavior data in the live broadcast scene, and build three bipartite graphs based on the collected behavior data, which respectively represent the user's behavior of entering the live broadcast room, chatting behavior, and giving gifts;

[0011] Step 2: Multi-behavior modeling based on graph neural network, using improved graph attention network to learn the repeated interaction relationship between users and live broadcast rooms;

[0012] Step 3: Integrate the user and live broadcast room representations generated by different behaviors through embedding fusion strategy;

[0013] Step 4: Use a multi-task learning framework to train and predict the live broadcast recommendation model, and generate a live broadcast recommendation list based on the model prediction results.

[0014] It can be seen from the technical solution provided by the present invention that the above method can solve the problems in the prior art such as insufficient modeling of various behaviors of users in live broadcast scenarios, limited utilization of repeated behaviors, and poor recommendation effect. It can comprehensively characterize user preferences and improve the accuracy of live broadcast room recommendations and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0016] Figure 1 A schematic diagram of a live broadcast recommendation method based on multiple repeated behaviors provided by an embodiment of the present invention;

[0017] Figure 2 A schematic diagram of the overall architecture of a live broadcast recommendation model based on multiple repeated behaviors constructed in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying 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, which does not constitute a limitation of the present invention. 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.

[0019] like Figure 1 FIG. 1 is a schematic diagram of a live broadcast recommendation method based on multiple repeated behaviors provided by an embodiment of the present invention, wherein the method comprises:

[0020] Step 1: First, we build a multiple repeated behaviors for live streaming recommendation model (MRB4LS), collect user behavior data in live streaming scenarios, and build three bipartite graphs based on the collected behavior data, which respectively represent the user's behavior of entering the live streaming room, chatting behavior, and giving gifts;

[0021] In this step, if Figure 2 FIG. 1 is a schematic diagram of the overall architecture of a live broadcast recommendation model based on multiple repeated behaviors constructed according to an embodiment of the present invention. The live broadcast recommendation model based on multiple repeated behaviors is composed of four parts:

[0022] Embedding initialization, used to initialize user embedding and live broadcast room embedding;

[0023] Multi-behavior repeated interaction modeling uses a multi-layer improved graph attention network RepGAT (Graph Attention Network) block to learn the user and live broadcast room embeddings for each repeated interaction behavior;

[0024] Embedded fusion, integrating representations of three types of interactive behaviors of users and live broadcast rooms;

[0025] Multi-task prediction: predict the possibility of users interacting with the live broadcast room under each behavior type, and use auxiliary tasks to improve the prediction effect of the target task.

[0026] In the specific implementation, after initializing user embedding and live broadcast room embedding, the live broadcast recommendation model extracts information from different types of repeated behaviors respectively, and then fuses them; finally, two auxiliary tasks (enter and chat behavior prediction) are used to improve the performance of the target task (gift behavior prediction).

[0027] The user behavior data collected in the live broadcast scene includes user ID, live broadcast room ID and behavior type (enter, chat, send gifts). In order to ensure the quality of the data set, the embodiment of the present invention adopts a 10-core setting, which means that the retained users and live broadcast rooms have at least ten interactions. Then, the collected user behavior data is divided into a training set, a validation set and a test set according to the chronological order, with the proportions of 80%, 10% and 10% respectively;

[0028] In the training set, each sample consists of a user ID, a live broadcast room ID, and a weight, where the weight represents the number of repeated interactions;

[0029] In the test and validation sets, each sample consists of a user ID, a live broadcast room ID, and a value. If the user interacts with the chat room, the value is set to 1, otherwise the value is set to 0. Here, the value 1 or 0 is used in the test and validation sets instead of weights (the number of repeated interactions). The purpose is to simplify model evaluation by focusing on predicting whether the user will interact with the live broadcast room instead of measuring the frequency of interaction.

[0030] Then, based on the collected behavioral data, three bipartite graphs are constructed, which respectively represent the user's behavior of entering the live broadcast room, chatting behavior, and giving gifts. Each bipartite graph contains user nodes and live broadcast room nodes. The three user-live broadcast room bipartite graphs constructed are as follows:

[0031] User-room-enter graph: nodes are users and live broadcast rooms, and the edge weights represent the number of times a user enters the live broadcast room.

[0032] Chat behavior (user-room-chat) graph: nodes are users and live broadcast rooms, and the edge weight is the number of messages sent by the user in the live broadcast room;

[0033] Gift-giving behavior (user-room-gift) graph: nodes are users and live broadcast rooms, and the weight of the edge is the number of times the user gives virtual gifts in the live broadcast room;

[0034] Among them, users and live broadcast rooms are represented by nodes, different types of behaviors are represented by multiple types of edges, and the number of repeated interactions between users and live broadcast rooms is embedded in the graph as edge weights.

[0035] For example, let’s take the user-room-gift diagram as an example to explain in detail:

[0036] The user-room-gift graph is a bipartite graph, denoted by , which is used to capture the information of gift-giving behavior between users and live broadcast rooms, where interaction means that users give virtual gifts in the live broadcast room; and They are the user set and live broadcast room set that have gift-giving behavior; The edge set representing the gift-giving behavior, where the edge weight Record the interaction history between users and the live broadcast room.

[0037] The existing multi-behavior recommendation model uses a binary method to record whether the user has interacted with the item, that is, ignoring repeated interactions. The value of is 1 or 0. In order to model repeated interactions, the present invention defines as follows:

[0038]

[0039] in In addition to this interactive information, there are two other behaviors that can also reflect the user's preference for the live broadcast room. The user-room-chat graph and the user-room-enter graph are defined in the same way. They are represented as and .

[0040] Step 2: Multi-behavior modeling based on graph neural network, using improved graph attention network to learn the repeated interaction relationship between users and live broadcast rooms;

[0041] In this step, we first initialize the behavior embedding. and are user and live studio sets respectively, where n and m are the number of users and live studios respectively; following the existing graph-based recommendation model, we first establish an embedding lookup table for each node in the user and live studio sets in the following way:

[0042] , in, , is the embedding matrix of all users and live broadcast rooms; d represents the dimension of latent features; and Representing users And live broadcast room To ensure the scalability of this method, each user or live broadcast room is described by a unique ID and represented by a one-hot vector. and Representing the one-hot vector matrices of all users and live broadcast rooms, use matrix multiplication to obtain users And live broadcast room The embedding is as follows:

[0043] , in, and The users are And live broadcast room The one-hot vector of ;

[0044] For the three behavior graphs, the improved graph attention network (RepGAT) is used for feature extraction to learn the repeated interaction relationship between users and live broadcast rooms. The specific process is as follows:

[0045] We first describe the single-layer structure of the improved graph attention network, and then generalize it to multiple consecutive layers. ,use Representation Node The first-order neighbors of this layer generate nodes Output characteristics for:

[0046] in, is the activation function; is the weight matrix; Is a node ' and is the embedding dimension; Representation Node For Node The normalized attention coefficient is The number of repeated interactions between nodes is introduced during the calculation, so that the model can use the number of repeated interactions to deeply learn the user's preference for the live broadcast room and the heterogeneous interaction relationship between the user and the live broadcast room. The calculation method is as follows:

[0047] in, Indicates a join operation; the superscript T is a transpose operation; is a trainable parameter; the LeakyReLU activation function allows the message to encode both positive and small negative signals; Is a node and nodes The edge weights between is the weight matrix of the linear transformation applied to the concatenated vector; Represents the node and nodes The two vectors are connected; Represents the node and nodes Two vectors connected; node Representative Node Neighbor nodes of Is a node and nodes The edge weights between

[0048] In order to enhance the model’s expressiveness and learning ability on graph data, and thus better capture the various relationships and features in the input sequence, the present invention uses a multi-head attention mechanism in the improved graph attention network. Specifically, K independent attention mechanisms perform the transformation of the single-layer structure of the graph attention network, and then their features are connected in series to produce the following output feature representation: :

[0049] In the formula The normalized attention coefficient calculated for the kth attention mechanism; is the weight matrix of the corresponding input linear transformation;

[0050] Next, the present invention stacks more attention embedding propagation layers to mine high-order connectivity information, which is crucial for encoding collaborative signals to estimate the relevance scores between users and live broadcast rooms. Attention embedding propagation layer, nodes can receive Hop-Neighbor Propagation Message, Improved Graph Attention Network in Behavior High-order embedding propagation is performed on the repeated interaction graph of , which can be expressed as:

[0051] in, Indicates that the node is After layer propagation, the behavior The updated embedding of the next node; the embedding matrix of layer 0, which is the initial embedding state of the node ; , is the adjacency matrix of the interaction graph, represents the identity matrix, is a diagonal matrix, where the diagonal elements ; It is behavior Next The trainable weight matrix of the layer;

[0052] For nodes , in behavior In the repeated interaction diagram, After layer propagation, its multi-layer representation is obtained, that is, ,Since the representations obtained in different layers emphasize the messages transmitted through different connections, they have different contributions in reflecting user preferences, so they are connected to form nodes In behavior The final embedding :

[0053] The initial embedding is enriched by embedding propagation layers, and by adjusting the number of layers To control the scope of dissemination.

[0054] Step 3: Integrate the user and live broadcast room representations generated by different behaviors through embedding fusion strategy;

[0055] In this step, we first obtain the behavior embedding, and design a connection-based method to fuse the user and live broadcast room representations generated by different behaviors. In order to make full use of different interactive behaviors to comprehensively represent user and live broadcast room nodes, the present invention fuses the embeddings learned from multiple types of repeated behaviors. After modeling multiple repeated interactions, each node will have three embeddings from its own interaction graph, which represent the feature representations learned under different behaviors. , and get its three embedding vectors , which come from enter behavior, chat behavior and gift behavior respectively;

[0056] Concatenates the node representations from the three interaction graphs into a single vector :

[0057] in, Indicates a connection operation;

[0058] It is assumed here that all interactions are necessary to understand the node The contribution of each interaction behavior is equal, but since the impact of different interactive behaviors may vary greatly, this assumption may not be optimal. Therefore, inspired by the attention mechanism, the present invention assigns a weight to each interactive behavior so that different interactive behaviors have different contributions to the potential factors of users and live broadcast rooms. Specifically:

[0059] Design an attention-based method to learn the importance of different behaviors, perform weighted fusion on the representations of different behaviors, and fuse the representations of multiple repeated behaviors by weighted summation:

[0060] in, Indicates different interaction behaviors on nodes The attention weights of the latent factors’ contributions are parameterized using an attention network :

[0061] choose As an activation function; and is a trainable parameter.

[0062] Step 4: Use a multi-task learning framework to train and predict the live broadcast recommendation model, and generate a live broadcast recommendation list based on the model prediction results.

[0063] In this step, during the training and prediction process of the live broadcast recommendation model, the model is trained using the collected live broadcast interaction data, and the probability of interaction between the user and a specific live broadcast room is predicted based on the trained model, and live broadcast content is recommended based on this. Specifically:

[0064] A multi-task learning framework is used to simultaneously predict the user's entry, chat, and gift-giving behaviors, and the auxiliary tasks (entry and chat behavior prediction) are used to enhance the prediction effect of the target task (gift-giving behavior prediction);

[0065] For user nodes and live studio nodes , after embedding fusion, the node representation is obtained and , because the final embedding involves the interaction between multiple different vector spaces, and the simple inner product cannot capture the complex nonlinear relationship between different vector spaces. and Represent the user set and live broadcast room set respectively, and describe the live broadcast recommendation task as follows:

[0066] Input: User collection And live room collection Graph structure interaction data of three repeated behaviors ;

[0067] Output: predicted user node Under the gift behavior (target behavior) and the live broadcast room node Probability of interaction ;

[0068] The unified embeddings of the two nodes are concatenated, and a two-layer multilayer perceptron (MLP) model is used to learn the interaction between these embeddings. The multilayer perceptron model is used to calculate the user node and live studio nodes The score between:

[0069] in, , and They represent the first The layer's weight matrix, bias vector, and activation function;

[0070] Multi-task learning is a joint training paradigm for different but related tasks. It improves the performance of each task by updating shared parameters or shared models. In order to better learn parameters from multiple types of repeated behavior data, this paper designs a loss function for each behavior to supervise the learning process of behavior features, and the optimization of each task adopts binary cross entropy loss. The loss function of gift, chat, and enter behaviors is , , They are:

[0071] in, express x Whether the behavioral interaction occurs, if x Behavioral interactions are observed. is marked as 1, otherwise it is marked as 0. x Behaviors include gift, chat, and enter behaviors, corresponding to superscripts g, c, and e, respectively; yes The prediction score of Indicated in x Samples observed in behavior, Indicated in x Unobserved samples in the behavior: For each live broadcast room with which the user has had positive interactions, 4 unobserved live broadcast rooms are randomly selected as negative samples;

[0072] In order to make full use of the signals in a variety of repeated behavior data, a multi-task learning strategy is adopted to regard the learning of each behavior feature as a task. The final loss is the summary of all losses of different behaviors. The formula is as follows:

[0073] in, , Respectively represent the loss function adjustment coefficients of gift and chat behaviors, which are used to adjust the proportion of losses of different behaviors;

[0074] Then, by calculating the interaction probability score, the recommendation results are sorted in descending order according to the score to generate a Top-K live broadcast recommendation list.

[0075] It is worth noting that the contents not described in detail in the embodiments of the present invention belong to the prior art known to professional and technical personnel in the field.

[0076] In order to evaluate the effect of the Top-K and preference ranking of the live recommendation list described in the embodiment of the present invention, two widely used evaluation indicators are adopted: NDCG@K and Recall@K. We repeated each experiment 20 times and reported the average results for each metric in the test set, which requires the model to output a predicted Top-K recommendation list and compare it with the real data.

[0077] Through the above method, the present invention significantly improves the performance of the recommendation system in the live broadcast scenario. Based on the modeling of multiple behaviors and repeated behaviors, it can effectively capture user preferences and significantly improve the recommendation accuracy (NDCG and Recall are improved by more than 6%), while alleviating data sparsity and cold start problems. In addition, the optimized graph neural network structure is adopted to realize real-time recommendation list generation, which meets the actual needs of dynamic scenarios of live broadcast platforms.

[0078] In summary, the method described in the embodiment of the present invention has the following advantages:

[0079] 1. Personalized recommendation: Through the joint modeling of multiple behavior information and repeated behaviors, the prediction ability of the recommendation system is improved to provide live broadcast recommendations that better meet the personalized needs of users;

[0080] 2. Improve interaction rate: Through more accurate recommendations, increase the frequency of interaction between users and the live broadcast room and improve user experience;

[0081] 3. Enhance platform revenue: Increase user consumption during live broadcasts, such as giving virtual gifts, thereby improving the platform's profitability;

[0082] 4. Data sparsity mitigation: Through the multi-task learning framework, the information of auxiliary tasks is used to enhance the prediction effect of the main task, effectively alleviating data sparsity and cold start problems;

[0083] 5. Real-time support: It adopts an efficient graph neural network structure to quickly generate recommendation results and adapt to the real-time needs of live broadcast scenarios.

[0084] In addition, a person skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be implemented by instructing related hardware through a program, and the corresponding program can be stored in a computer-readable storage medium. The above-mentioned storage medium can be a read-only memory, a disk or an optical disk, etc.

[0085] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims. The information disclosed in the background technology section of this article is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as an admission or in any form that the information constitutes prior art known to those skilled in the art.

Claims

1. A live broadcast recommendation method based on multiple repeated behaviors, characterized in that: The method comprises: Step 1: First, build a live broadcast recommendation model based on multiple repeated behaviors and collect user behavior data in live broadcast scenarios. In step 1, the collected user behavior data is divided into a training set, a validation set, and a test set according to the chronological order; In the training set, each sample consists of a user ID, a live broadcast room ID, and a weight, where the weight represents the number of repeated interactions; Then, three bipartite graphs are constructed based on the collected behavioral data as follows: Entering behavior graph: nodes are users and live broadcast rooms, and the weight of the edge represents the number of times the user enters the live broadcast room; Chat behavior graph: nodes are users and live broadcast rooms, and the edge weight is the number of messages sent by the user in the live broadcast room; Gift-giving behavior graph: the nodes are users and live broadcast rooms, and the edge weights are the number of times a user gives virtual gifts in the live broadcast room; Step 2: Multi-behavior modeling based on graph neural network, using improved graph attention network to learn the repeated interaction relationship between users and live broadcast rooms; The process of step 2 is specifically as follows: First, initialize the behavior embedding. and are the sets of users and live broadcast rooms respectively, where n and m are the numbers of users and live broadcast rooms respectively; Create an embedding lookup table for each node in the user and live studio collection as follows: in, is the embedding matrix of all users and live broadcast rooms; d represents the dimension of latent features; and Representing users And live broadcast room Initialization embedding of; For the three behavior graphs, the improved graph attention network is used for feature extraction to learn the repeated interaction relationship between users and live broadcast rooms. The specific process is as follows: Calculate the normalized attention coefficient α of node i to node j ij : in, is a trainable parameter; d′ is the embedding dimension; LeakyReLU is the activation function; w i，j is the edge weight between node i and node j; W A is the weight matrix of the linear transformation applied to the concatenated vector; [e i ||e j ] represents the connection between two vectors of node i and node j; [e i ||e k ] represents the connection between two vectors of node i and node k; represents the first-order neighbors of node i; w i,k is the edge weight between node i and node k; Calculate the output feature representation e′ i : In the formula is the normalized attention coefficient calculated by the qth attention mechanism; W q is the weight matrix of the corresponding input linear transformation; σ(·) is the activation function; Q represents the number of independent attention mechanisms; By stacking l attention embedding propagation layers, a node receives messages propagated from its l-hop neighbors, and the improved graph attention network performs high-order embedding propagation on the repeated interaction graph of behavior b, expressed as: Among them, E (b，l) represents the updated embedding of the node under behavior b after the node propagates in layer l; the embedding matrix of layer 0, i.e. the initial embedding state of the node A is the adjacency matrix of the interaction graph, I is the identity matrix, and D is a diagonal matrix where the mth diagonal element is W (b，l-1) is the trainable weight matrix of the l-1th layer under behavior b; For node i, after L layers of propagation in the repeated interaction graph of behavior b, its multi-layer representation is obtained, that is, Connect them to form the final embedding of node i in behavior b Step 3: Integrate the user and live broadcast room representations generated by different behaviors through embedding fusion strategy; In step 3, we first obtain the behavior embedding and design a connection-based method to fuse the user and live broadcast room representations generated by different behaviors. After modeling multiple repeated interactions, each node has three embeddings from its interaction graph, which represent the feature representations learned under different behaviors. For node i, we get its three embedding vectors They come from entry behavior, chat behavior, and gift-giving behavior respectively; The representation of multiple repeated behaviors is fused by weighted summation: in, tanh(·) is the activation function; W b and p are trainable parameters; Step 4: Use a multi-task learning framework to train and predict the live broadcast recommendation model, and generate a live broadcast recommendation list based on the model prediction results.

2. The live broadcast recommendation method based on multiple repeated behaviors according to claim 1, characterized in that: In step 4, during the training and prediction process of the live broadcast recommendation model, the collected live broadcast interaction data is used to train the model, and the probability of interaction between the user and a specific live broadcast room is predicted based on the trained model, and the live broadcast content is recommended based on this.

Images