Micro map method oriented to diversified interests of users
By using graph convolutional neural network and user intention decoupling method in the micromap recommendation system, the problem of data sparseness and insufficient user interest capture capabilities are solved, more accurate and comprehensive recommendations are achieved, and the interpretability of the recommendation results is improved.
Patent Information
- Application Number
- CN202510082449.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-08-17
- Filing Date
- 2025-01-20
- Publication Date
- 2025-04-29
AI Technical Summary
In the micromap recommendation system, the data distribution is uneven because the number of maps far exceeds the number of users, and the recommendation algorithm is difficult to accurately infer and predict users' preferences and interests, and the recommendation results are low interpretability.
A graph convolution neural network is introduced to model the interactive data between users and maps, and a user intention decoupling method is added to the graph convolution layer. By establishing multiple user intention perception subgraphs and using neighborhood routing mechanisms, user representations are decoupled to obtain more fine-grained user intentions.
It effectively alleviates the problem of data sparseness, improves the accuracy and comprehensiveness of recommendations, and makes the recommendation results have good interpretability.
Smart Images
Figure CN119903239A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of micro-maps and recommendation systems, and more specifically to a micro-map method oriented to diverse interests of users. Background Art
[0002] WeMaps are grassroots maps for the public in the era of self-media. Because they do not require high mathematical foundations such as accuracy, users can participate in map production in real time without strict professional training, so massive amounts of map data are bound to be generated after they are put into use. In order to enable maps to be distributed in a timely, convenient and fast manner and for users to efficiently obtain valuable information from massive amounts of map data, it is necessary to add personalized recommendation functions to WeMaps.
[0003] In the current micro-map recommendation system, the number of maps far exceeds the number of users, and most maps have few users interacting with them, so they generally lack corresponding interaction data. This phenomenon of uneven data distribution makes it difficult for micro-map recommendation algorithms to accurately infer and predict user preferences and interests. How to deal with similar data sparsity problems is a common challenge faced by various recommendation systems at present. Graph-based deep learning methods can use the characteristics of the graph structure itself to fill in the data gaps by transferring information between nodes, thereby effectively alleviating the trouble caused by data sparsity in the recommendation process. For example, some methods use the high-order connectivity of the user and item bipartite graphs to integrate the potential "cooperative signals" in the interaction process into the embedding vector learning process of the graph nodes. Other methods combine "random walk" with graph convolutional neural network algorithms to generate feature embedding vectors covering edge and node information in the graph. The above graph-based recommendation methods have achieved certain results in solving the data sparsity problem of recommendation systems, but they are insufficient in capturing the diverse interests of users. The main reason is that users’ interests are often multidimensional and implicit. Graph models usually use the attributes of the nodes themselves or the information of neighboring nodes to comprehensively generate vector representations of nodes. However, such information cannot fully reflect the users’ implicit interests. In addition, the process of inferring and predicting through the similarity or correlation of nodes is also uncertain to a certain extent, which reduces the interpretability of the recommendation results.
[0004] In order to solve the challenges brought by the data sparsity problem in the current micro-map recommendation system, and also to capture the diverse interests of users and make the recommendation results have good interpretability, the present invention proposes a micro-map recommendation method oriented to the diverse interests of users. Summary of the invention
[0005] In view of this, the present invention provides a micro-map recommendation method for users' diverse interests. In order to solve the data sparsity problem in the existing micro-map recommendation methods, a graph convolutional neural network is introduced for the first time to model the interaction data between users and maps. In order to capture the diverse interests of users and provide users with personalized recommendations while making the recommendation results well interpretable, a user intention decoupling method is added to the graph convolution layer. Based on the constructed "user-map" bipartite graph, multiple user intention-aware subgraphs are established, and the neighborhood routing mechanism is used to iteratively update the node representations on different subgraphs to decouple user representations and obtain more fine-grained user intentions in the process of user-map interaction.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] A method for recommending micro-maps based on diverse user interests, the model includes a "user-map" interaction graph module, a graph convolution module, and a prediction and output module. The recommendation model construction method includes:
[0008] S1: Use graph data structure to model the historical data of user-micromap interaction and establish a "user-map" interaction graph;
[0009] S2: Based on the constructed “user-map” interaction graph, multiple “user-map” interaction subgraphs are constructed, each of which corresponds to an intention of the user when interacting with the map;
[0010] S3: By performing graph convolution on the “user-map” interaction subgraph, the information of different nodes on the graph can be transmitted and aggregated;
[0011] S4: Add a neighborhood routing mechanism to the graph convolution process, and perform multiple iterative updates of node information in each graph convolution process;
[0012] S5: After completing all graph convolution operations, the node vectors on each "user-map" interaction subgraph in different convolutional layers are concatenated and added together as the final node representation vector;
[0013] S6: Take the inner product of the user and map node representation vectors obtained through training to predict the user's preference for maps, sort the scores in descending order, and recommend the top K maps with the highest scores to the user to form the TopK recommendation results;
[0014] S7: End.
[0015] Furthermore, the "user-map" interactive diagram module includes:
[0016] A module that defines a "user-item" interaction graph and an intent-aware subgraph. The principle is to define the map where users interact with each other as nodes in the graph, and the edges in the graph represent an interaction between nodes. In addition, K intent-aware subgraphs are constructed based on the constructed "user-item" interaction graph, and each subgraph is consistent with the original graph in form. The contribution size of the intent corresponding to each subgraph to the interaction is initialized as:
[0017] A(u,m)=(A1(u,m),...,A k (u,m),...,A K (u,m)) (0)
[0018] A graph node embedding vector initialization module, defining the embedding vector representation e of user u after decoupling u It is composed of K independent user intention vectors:
[0019] e u =Concatenate(e 1u ,...,e ku ,...,e Ku ) (2)
[0020] Furthermore, the specific steps of the graph convolution module are:
[0021] The first step is to define the total number of graph convolution layers L, and define the specific execution process of the l-th layer of graph convolution by the following formula:
[0022]
[0023] In the second step, the user intention decoupling method is introduced in each graph convolutional layer. By using the neighborhood routing mechanism on different subgraphs, the "user-map" interaction weights in the corresponding subgraphs can also be updated to obtain the contribution of different intentions to the interaction and realize the decoupling of user intentions.
[0024] Furthermore, the prediction and output module includes:
[0025] An addition operation of all the node vectors output by the graph convolution layers. The results of all the graph convolution layers of the k-th subgraph are added to generate the final node representation vector. After performing L graph convolution operations, the user and map embedding feature vectors can be calculated by formula (4):
[0026]
[0027] An operation that concatenates the vectors of each subgraph node and performs prediction. The user and map vectors in the K user intent subgraphs are concatenated to obtain the embedding vector representation e of the final user u and map m. u 、e m, calculate the inner product of the final vector representation obtained to calculate the preference of user u for map m:
[0028]
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] 1. This invention introduces a graph convolutional neural network for the first time in micro-map recommendation to model the interaction process between users and maps, and uses the high-order connectivity between nodes to discover the collaborative signals between nodes and their high-hop neighbor nodes, thereby alleviating the problem of difficulty in forming effective recommendations due to sparse data on user-map interaction in the recommendation process.
[0031] 2. In order to accurately grasp the specific intentions of users when interacting with maps and capture the diverse interests of users, so as to make more accurate and comprehensive recommendations, the present invention adds a user intention decoupling method to the graph neural network model. By establishing multiple intention-aware subgraphs, the neighborhood routing mechanism is iteratively executed in the intention-aware subgraphs to decouple the intentions of users interacting with maps, and the contribution of different intention factors to the interaction between users and maps is obtained. The introduction of the user intention decoupling method provides effective support for improving the recommendation effect of micromaps and conducting interpretable analysis of the recommendation process. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only schematic diagrams of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0033] Figure 1 A schematic diagram of a flow chart of a method for recommending micro-maps based on diverse interests of users in the present invention;
[0034] Figure 2 Schematic diagram of the overall structure of a recommendation model for user diversified interests in an embodiment of the present invention;
[0035] Figure 3 It is a structural diagram of a method for decoupling user intention in an embodiment of the present invention; DETAILED DESCRIPTION
[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.
[0037] like Figure 2 As shown, a method for recommending micromaps based on diverse interests of users includes the following steps:
[0038] S1: Use graph data structure to model the historical data of user-micromap interaction and establish a "user-map" interaction graph;
[0039] S2: Based on the constructed “user-map” interaction graph, multiple “user-map” interaction subgraphs are constructed, each of which corresponds to an intention of the user when interacting with the map;
[0040] S3: By performing graph convolution on the “user-map” interaction subgraph, the information of different nodes on the graph can be transmitted and aggregated;
[0041] S4: Add a neighborhood routing mechanism to the graph convolution process, and perform multiple iterative updates of node information in each graph convolution process;
[0042] S5: After completing all graph convolution operations, the node vectors on each "user-map" interaction subgraph in different convolutional layers are concatenated and added together as the final node representation vector;
[0043] S6: Take the inner product of the user and map node representation vectors obtained through training to predict the user's preference for maps, sort the scores in descending order, and recommend the top K maps with the highest scores to the user to form the TopK recommendation results;
[0044] S7: End.
[0045] In order to further optimize the above technical solution, in step S2, K intent-aware subgraphs are constructed according to the set number of decoupled user intentions K. The structure, number of nodes and edges of each subgraph are consistent with the original graph, except that the representation of nodes and weights of edges on each subgraph are different in later training. Furthermore, we initialize the constructed intent-aware subgraph, which mainly includes two processes.
[0046] In the node embedding vector initialization process, when initializing the user and map embedding vector representation, the representation of a user embedding vector is divided into K parts according to the number of decoupled user intentions K set in the model training, as shown in formula (2). During initialization, the embedding vector representation in each subgraph is randomly initialized, and then the embedding vector representation of the node in the subgraph will be iteratively updated multiple times by executing the neighborhood routing mechanism in the graph convolution process.
[0047] The initialization of the contribution of intention to the generated user and map interaction is actually the initialization of formula (1). During the initial training of the model, it is assumed that the contribution of each user intention to the interaction between users u and m is the same, and the initial value of A(u,m) is initialized to:
[0048] A(u,m)=(1,...,1,...,1) (6)
[0049] The method involved in the present invention is further described below.
[0050] The present invention introduces a neighborhood routing mechanism to transfer information between nodes during the training process of subgraph node embedding vectors. In this mechanism, each node only interacts with its first-order neighbor nodes, reducing the complexity and communication cost of network interaction. The main workflow of the neighborhood routing mechanism includes:
[0051] In the first step, in the lth graph convolution layer, the embedding vector representation of user u in the kth intent-aware subgraph is The contribution of this intention to the interaction between user u and map m It is obtained by performing T times of neighborhood routing iterations. The iteration process is as follows: Figure 3 For the convenience of representation, the tth iteration in the T-iteration process corresponds to and Recorded as
[0052] In the second step, during the tth iteration of the neighborhood routing mechanism, the contribution scores of user intentions to the interactions in all subgraphs are first normalized using the Softmax function. The calculation process is shown in formula (7):
[0053]
[0054] Where: Represents the contribution of the intent to the interaction between user u and map m in the k-th intent-aware subgraph at the t-th iteration of the neighborhood routing mechanism. After integrating the contribution of all intentions of user u to the interaction, The result obtained by normalization. Through this process, we can intuitively get which intention of the user contributes most to the interaction when the user interacts with the map.
[0055] Secondly, information propagation of nodes in the subgraph is performed through formula (8):
[0056]
[0057] Where: By aggregating user u’s first-order neighbor nodes N at the tth iteration ku ={m|(u,m)∈G k} information is obtained, is the vector representation of its first-order neighbor node, that is, the map node that has directly interacted with user u. When executing the neighborhood routing mechanism, the reason for only aggregating its first-order neighbor node information is to eliminate the interference of its high-hop neighbor node information and retain relatively complete local context information. yes The Laplace matrix obtained after degree normalization is calculated as shown in formula (9):
[0058]
[0059] Where: is the corresponding degree of user u node and map m node in the kth subgraph. Related research has shown that degree normalization can reduce the impact of node degree differences on training results, making the model more robust and generalizable.
[0060] The third step is that in the same intent-aware subgraph, the user's interaction with the map is based on the same intent. Therefore, during the training process, the model needs to encourage the node vector representations in each intent-aware subgraph to be as similar as possible. Based on this, S k The calculation method of iterative update is shown in formula (10):
[0061]
[0062] Where: It can be understood as calculating the affinity between user u and map m. Tanh is a nonlinear activation function introduced to increase the representation ability of the model.
[0063] The final output of the neighborhood routing mechanism after T iterations is That is, the output result of the lth graph convolutional layer. For any intent k∈{1,...,K}, the node vector representation and weight matrix of the corresponding subgraph output after multiple rounds of training are important bases for interpretability analysis of the model.
[0064] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0065] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for recommending micro-maps based on diverse interests of users, characterized in that: A new recommendation model is proposed for recommending micro-map tasks of interest to users in a micro-map recommendation system. The model includes a "user-map" interaction graph module, a graph convolution module, and a prediction and output module. The recommendation model construction method includes: S1: Use graph data structure to model the historical data of user-micromap interaction and establish a "user-map" interaction graph; S2: construct multiple "user-map" interaction subgraphs based on the constructed "user-map" interaction graph, each subgraph corresponds to an intention of the user when interacting with the map; S3: By executing the graph convolution method on the "user-map" interaction subgraph, the information of different nodes on the graph can be transmitted and aggregated; S4: Add a neighborhood routing mechanism to the graph convolution process, and perform multiple iterative updates of node information in each graph convolution process; S5: After completing all graph convolution operations, concatenate and add the node vectors on each "user-map" interaction subgraph in different convolutional layers as the final node representation vector; S6: Take the inner product of the user and map node representation vectors obtained through training to predict the user's preference for maps, sort the scores in descending order, and recommend the top K maps with the highest scores to the user to form the TopK recommendation results; S7: End. The method for recommending micromaps based on diverse interests of users according to claim 1 is characterized in that: In step S4, the specific steps of iteratively updating the node vector representation using the neighborhood routing mechanism are: Step 1: Normalize the contribution scores of user intentions to interactions in all subgraphs; Step 2: Use the Laplacian matrix obtained after degree normalization to aggregate the information of the node and its first-order neighbor nodes into the node itself; Step 3: Calculate the contribution of the corresponding intention to the user's interaction with the map by updating the node vector representation once.