A method and device for discovering same-type users in a heterogeneous network based on multi-granularity fusion

Abstract

This invention belongs to the field of user recommendation, and specifically relates to a heterogeneous network representation method and apparatus based on multi-granularity fusion. The method includes collecting and completing / aligning the feature information and relationship data of relevant users; aggregating user nodes directly contacted by relevant users to fuse the low-order structure of the user network; aggregating the meta-path information of relevant users to fuse the high-order semantics of the user network; using the idea of ​​multi-granularity cognitive computing, performing complementary aggregation of low-order structure vectors and high-order semantic vectors to obtain low-dimensional vectors of relevant users; projecting the low-dimensional vectors of relevant users onto a two-dimensional coordinate system, and recommending the N users closest to the low-dimensional vectors of relevant users to the relevant users. This invention fuses low-order structure encoding and high-order semantic encoding to discover potential information in the user network, thereby improving the accuracy of user relationship prediction and user recommendation efficiency.

Description

Technical Field

[0001] This invention belongs to the field of user recommendation, and specifically relates to a method and apparatus for discovering users of the same type in heterogeneous networks based on multi-granularity fusion. Background Technology

[0002] Heterogeneous networks, due to their diverse node and edge structures, can express richer information, enabling the modeling of various scenarios and interactions within those scenarios. In real-world applications, heterogeneous networks have numerous uses, such as modeling different types of users on a website and learning their low-dimensional vectors. This low-dimensional vectors can then be used for close user recommendations among relevant users.

[0003] However, in real-world scenarios, modeling and mining complex networks is no easy task. Relevant information within the network may be overlooked, while erroneous information may be included, creating confusion. Manually filtering and analyzing this information consumes enormous human and material resources, and the collected information is not always complete, further complicating network modeling.

[0004] User networks are also a type of social network. Due to the different types of users and relationships, and the inclusion of various attributes (such as user age, gender, and region), modeling them as heterogeneous networks better describes the information within them. After modeling, the model can more accurately uncover the latent information within the user network, accelerating the integration of effective information in user recommendations and thus improving recommendation accuracy.

[0005] Current research on heterogeneous networks largely focuses on their structural, semantic, and attribute information. However, effectively extracting neighboring structural information and accurately extracting semantic information remain unresolved issues. Furthermore, structural and semantic information contribute differently to the model's structure. Applying these technical challenges to complex network analysis reveals that models may overlook crucial information, such as user location, age, and gender, while simultaneously incorporating irrelevant information as essential. Summary of the Invention

[0006] To address the above problems, this invention proposes a method and apparatus for discovering users of the same type in heterogeneous networks based on multi-granularity fusion. The method specifically includes the following steps:

[0007] Collect relevant user feature information and relationship data, and complete and align any missing or incomplete parts of the collected information;

[0008] Aggregate user nodes that are directly contacted by relevant users and merge the low-level structure of the user network;

[0009] Aggregate the meta-path information of relevant users and integrate the higher-order semantics of the user network;

[0010] Using the approach of multi-granularity cognitive computing, low-order structural vectors and high-order semantic vectors are complementaryly aggregated to obtain the user's low-dimensional vector;

[0011] Project the low-dimensional vectors of the user network onto a two-dimensional coordinate system, identify the N users whose low-dimensional vectors are closest to the current user as similar users, and recommend similar users to the current user.

[0012] Furthermore, based on the acquired information, the similarity distance between the relevant user and the people they come into contact with is calculated using the low-dimensional vector of the user network, and several close user sets of the relevant user are predicted based on the similarity distance.

[0013] Furthermore, acquire relevant users The process of refining content attribute information involves generating a random walk sequence, updating the parameters using a skip-gram model based on the generated sequence, and generating the final node representation matrix, including:

[0014]

[0015]

[0016] in, Therefore A random walk sequence starting from , Represents a heterogeneous neural network, where t is the length of the walk sequence; S∈R |V|×d This is a user attribute content matrix, where each row of the matrix represents... For relevant users A content attribute; d is the node embedding size, w is the wandering window size; |V| is the number of users in the network user set V.

[0017] Furthermore, the process of obtaining low-level structural information about the user network includes:

[0018] Get relevant users The set of first-order neighbor nodes, and the category of the set of contacted users;

[0019] The users contacted in each category are merged separately according to the category to obtain the merging result for each category;

[0020] The fusion results for each category are compared with those of the relevant users. The data is then fused to obtain the low-order structure vector of the network.

[0021] Furthermore, by merging the users contacted in each category separately, the categories are: The user integration process can be represented as follows:

[0022]

[0023] in, Indicates category as The results of user integration. To reach users For relevant users Importance For relevant users The category of the contact user set is One of the nodes, Indicates contact with users Among the users contacted, the type is The set of nodes; For relevant users The node content attributes, Indicates relevant users The type of the contact user set is Similar users.

[0024] Furthermore, if there are K other types of users who have been contacted, then the relevant users... The process of integrating with all types of users is represented as follows:

[0025]

[0026] in, This represents the low-order structural information vector of the network; For type Contact users to relevant users The degree of association is calculated through an attention mechanism; K is the number of types. This refers to the nth type out of a total of K types, or simply type.

[0027] Furthermore, the process of aggregating the metapath information of relevant users includes:

[0028] The meta-path is encoded using GCN, and the result of the last layer of GCN is used as the encoding based on the meta-path.

[0029] The encodings of all original paths are fused to obtain the high-order semantic vector of the user network.

[0030] Furthermore, during the encoding of the metapath using GCN, the output of the (l+1)th layer of GCN is represented as:

[0031]

[0032] in, For relevant users Features in layer l For relevant user v j Features in the l-th layer; For relevant users The neighbor set includes itself; PReLU() represents the activation function; C ij W represents the normalization factor. (l) b (l) All are trainable matrices, W (l) Let b be the weight of the l-th layer. (l) This is the bias term for the l-th layer.

[0033] Furthermore, through contrastive learning, low-order structural vectors and high-order semantic vectors are complementary and aggregated. During the aggregation process, the cosine distance between the low-order structural vector and the high-order semantic vector is used as a loss function for optimization until the cosine distance between the optimized low-order structural vector and the high-order semantic vector is less than a set threshold. This yields fused low-order structural vectors and high-order semantic vectors, one of which is used as the final low-dimensional vector of the user network. This invention also proposes a heterogeneous network representation device based on multi-granularity hierarchical fusion, including a feature extraction module, an attribute completion module, a structural encoder module, a semantic encoder module, a biased contrastive fusion module, an encoding fusion module, and a recommendation module, wherein:

[0034] The feature extraction module is used to extract content attribute information of relevant users;

[0035] The attribute completion module is used to collect and complete the content attributes of relevant users;

[0036] The structure encoder module is used to aggregate the first-order neighbor nodes of relevant users and merge the user's low-order structure;

[0037] The semantic encoder module is used to aggregate the meta-path information of relevant users and integrate the user's higher-order semantics;

[0038] The biased contrast fusion module is used to perform complementary aggregation of low-order structure and high-order semantics, and finally obtain the user's low-dimensional vector.

[0039] The recommendation module projects the low-dimensional vectors of relevant users onto a two-dimensional coordinate system, identifies the N users whose low-dimensional vectors are closest to those of the relevant users as similar users, and recommends similar users to the relevant users.

[0040] This invention utilizes the concept of multi-granularity cognitive computing to fuse structural and semantic information from complex networks. Then, it employs a biased contrastive fusion method to finally fuse low-order structural vectors and high-order semantic vectors to obtain embedding results. This process uncovers latent information within user networks, thereby improving the accuracy of user relationship prediction and enhancing user recommendation efficiency. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the process for discovering users of the same type in a heterogeneous network based on multi-granularity fusion according to the present invention.

[0042] Figure 2 This is a schematic diagram of the structure of the three modules in this invention: structural encoder, semantic encoder, and biased contrast fusion.

[0043] Figure 3 This is a schematic diagram of the high-order semantic encoding of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] This invention proposes a method for discovering users of the same type in heterogeneous networks based on multi-granularity fusion, specifically including the following steps:

[0046] Collect relevant user feature information and relationship data, and complete and align any missing or incomplete parts of the collected information;

[0047] Aggregate user nodes that are directly contacted by relevant users and merge the low-level structure of the user network;

[0048] Aggregate the meta-path information of relevant users and integrate the higher-order semantics of the user network;

[0049] Using the approach of multi-granularity cognitive computing, low-order structural vectors and high-order semantic vectors are complementaryly aggregated to obtain the user's low-dimensional vector;

[0050] Project the user's low-dimensional vector onto a two-dimensional coordinate system, identify the N users whose low-dimensional vectors are closest to the relevant user as similar users, and recommend similar users to the relevant user.

[0051] In this embodiment, as Figure 1After collecting user information (i.e., collecting content attributes), an attribute completion module is used to complete missing or incomplete parts of the collected information. The completed data is then input into a low-order structure encoder and a high-order semantic encoder. The encoded low-order structure and high-order semantics are then subjected to biased contrast fusion to obtain a final low-dimensional vector. This final low-dimensional vector is applied to fields such as link prediction and friend recommendation. This invention primarily uses the final encoding of relevant users in the recommendation field, for example, recommending similar users or related products to similar users to relevant users. Specifically, it includes:

[0052] (I) Data Collection and Data Completion

[0053] After collecting relevant user content attributes, the attribute completion module uses random walk and SkipGram methods to complete missing or incomplete attributes. Features of relevant users are extracted, i.e., relevant user content attributes, such as age, location, and gender. The data obtained in this process may lack node content attribute information or the content attribute information may be incomplete. To enable unified processing of user attributes, we use the topological relationships between users as a guide, completing the relevant user attributes based on the overall topology of the user network.

[0054] Get relevant users The process of perfecting content attribute information includes:

[0055]

[0056]

[0057] in, Therefore A random walk sequence starting from , Let be the entire heterogeneous neural network, and t be the length of the walk sequence. Then, based on the generated walk sequence, the parameters are updated using a skip-gram model, and the final relevant user representation matrix is ​​generated. Embed the node with size, where w is the size of the traversal window, and each row For relevant users Content attributes.

[0058] (ii) Obtaining the low-level structure of relevant users

[0059] Integrate relevant users The system identifies the contact users and obtains the low-order structure vector of the network. As an optional implementation, in this embodiment, the contact users can refer to neighboring users directly connected to the relevant user, or users within a certain range from the relevant user. This certain range is set by those skilled in the art based on actual circumstances. After determining the set of contact users for the relevant user, the contact users in this set are classified according to their type. Then, contact users of each type are aggregated separately. The process of aggregating users through contact can be represented as follows:

[0060]

[0061] in, For relevant users The node content attributes; To reach users For relevant users The correlation degree is calculated using existing methods for calculating the importance of one user to another. As an optional implementation, this embodiment chooses to use an attention mechanism for calculation.

[0062] After each type is merged, if there are a total of K types of users who have been in contact with the user, the merging process for these K types of users includes:

[0063]

[0064] in, For type Contact users The importance of [something] is calculated through an attention mechanism; This refers to the low-order structure vector of the network; For type Contact users The vector representation of .

[0065] (III) Obtaining the higher-order semantics of nodes

[0066] In obtaining the higher-order semantic vectors of nodes, the metapath is first encoded using GCN. The encoding process includes:

[0067]

[0068] in, For relevant users Features in layer l For relevant users The result of the last layer, including the neighboring user set including oneself. That is, based on metapath The node encoding.

[0069] Based on the number of original paths, if there are M original paths, then merge the relevant users. The process of obtaining M meta-path information includes:

[0070]

[0071] in, The importance weight of a meta-path; This is the result of encoding the higher-order semantic information of the network; It is a vector representation of a node on a metapath.

[0072] In this embodiment, Vector representations of nodes in different scenarios. This represents the vector representation of a node during the fusion process of nodes according to category, or the user's low-order structure. This means that during the process of merging nodes according to metapaths, the vector representation of a node can also be called a metapath-based representation. The node encoding, or higher-order semantics of the user.

[0073] like Figure 2 This embodiment uses a biased contrast fusion method to process low-order structure vectors. and higher-order semantic vectors The structural vectors and semantic vectors are fused. Since both obtain user network information from different perspectives, they each contain information that the other lacks. When fused, they complement each other, resulting in a more complete user network information. This embodiment uses the cosine function to calculate... and The degree of similarity, when and When the degree of similarity increases during the fusion process, it indicates... and The better, the more complementary information you obtain from the other party. and The information will then be more complete.

[0074] Low-order structure vectors and higher-order semantic vectors The formula for the cosine phase function between them is:

[0075]

[0076] Furthermore, define the loss function:

[0077]

[0078] in, As a preferred embodiment, the weight matrix is ​​initialized as a diagonal matrix with the values ​​on the diagonal set to 2. To connect with relevant users via metapath A connected set of users. In machine learning, the goal is for the loss function to decrease during training, eventually converging to a constant value. During this process, and Information complementarity is also being achieved. When training ends, the two vectors tend to be complete. As a preferred implementation scheme, this embodiment ultimately selects [the appropriate vector] based on experimental data. For the low-dimensional vector of the user network, Compare It contains more information.

[0079] As an optional technical solution, this embodiment provides an application of obtaining potential information about target users through the low-dimensional vectors of the final user network. Based on the learned low-order vectors of the user network... It is possible to uncover latent information about the target user class, thereby recommending several close user sets of the relevant users, for example... Figure 3 Each circle represents a user, and the value inside the circle is the user's label. When we need to study related users, for example... Figure 3 User 5 in the network can be encoded using the low-order vector of the user network, i.e. Figure 3 The left side of the middle represents the relationships between users. Data mining is then performed. In this embodiment, users directly connected to relevant users in the low-order vector encoding of the user network are considered as the contact users of those relevant users. For example... Figure 3 Users 1, 2, 3, 4, and 12 are considered close users. Users who are connected to related users but not connected to related users are considered close users. Figure 3 When making recommendations, users 5 and 11 will be recommended to each other's close users.

[0080] This embodiment also proposes a heterogeneous network representation device based on multi-granularity fusion, including a feature extraction module, an attribute completion module, a structural encoder module, a semantic encoder module, a biased contrast fusion module, an encoding fusion module, and a recommendation module, wherein:

[0081] The feature extraction module is used to extract content attribute information of relevant users;

[0082] The attribute completion module is used to collect and complete the content attributes of relevant users;

[0083] The structure encoder module is used to aggregate the first-order neighbor nodes of relevant users and merge the user's low-order structure;

[0084] The semantic encoder module is used to aggregate the meta-path information of relevant users and integrate the user's higher-order semantics;

[0085] The biased contrast fusion module is used to perform complementary aggregation of low-order structure and high-order semantics, and finally obtain the user's low-dimensional vector.

[0086] The recommendation module projects the user's low-dimensional vector onto a two-dimensional coordinate system, identifies the N users whose low-dimensional vector is closest to the current user as similar users, and recommends similar users to the current user.

[0087] As an optional implementation, when recommending relevant users based on the similarity distance calculated from the low-dimensional vector, the relevant user approach is adopted. With user v j The Euclidean distance, used as the similarity distance, is expressed as:

[0088]

[0089] in, Indicates relevant users With user v j The Euclidean distance; n is the dimension of the low-dimensional vector, related users The low-dimensional vector is (x1, x2, ..., x n ),user The low-dimensional vector is (y1, y2, ..., y n After calculating the Euclidean distance between the relevant user and all other users, sort the users by this Euclidean distance in ascending order, and select the top N users as the relevant users. Close users.

[0090] As another alternative implementation, after obtaining similar users of the current user according to this application, similar users can be recommended to the current user in a social network, and products purchased by similar users can be recommended to the current user in a shopping app.

[0091] Existing research on complex network analysis largely focuses on homogeneous networks. However, real-world network data is complex and highly variable. Treating these networks as homogeneous leads to the loss of much relevant data. Heterogeneous networks better describe the various networks in real-world scenarios. Furthermore, effectively extracting proximity structure information and accurately extracting semantic information from networks remains an open problem. This embodiment also provides a specific implementation process:

[0092] Figure 3This is a schematic diagram of a user network embodiment in this invention. Assume that in a certain social networking site, user 5 is our relevant user. He / She has direct interaction with users 1, 2, 3, 4, and 12, meaning they are users who have been in contact with user 5. After a multi-granularity fusion method, a low-dimensional vector of the user network is obtained. We can find that although users 6 and 11 are not directly connected to relevant user 5, they have a potential connection. Therefore, users 6 and 11 may be close users of relevant user 5. Based on this, we can make friend recommendations to relevant user 5 in this social networking site. The specific steps of using this method are as follows:

[0093] 1. Collection of user node content attributes

[0094] Retrieve user attribute information and interaction information, such as gender, age, and location, from social networking site databases.

[0095] 2. Network Modeling

[0096] Based on the specific circumstances of this social networking site, we will treat each user as a node and the interaction relationships between users as edges to establish a heterogeneous network G. This network will be the focus of subsequent discussions.

[0097] 3. Attribute completion

[0098] Since the obtained node attributes may be incomplete (e.g., user 1 chooses to hide gender and age), the node content attribute information needs to be completed and preprocessed according to the network topology before each node in network G can be processed uniformly.

[0099] 4. Low-order structure encoding

[0100] User 5 has a first-order structure, such as 1-5, 2-5, 3-5, 4-5, 5-12, indicating that User 1 has direct interaction with the other users mentioned above. This interaction may include adding friends, liking, chatting, etc. Based on the first-order structure, this embodiment can obtain the low-order structure vector of User 5 through a low-order structure encoder.

[0101] 5. High-order semantic coding

[0102] by Figure 3 For example, user 5 has several meta-paths starting from 5, such as 5-1-6 and 5-12-11. These meta-paths have different semantics: user 5 and user 6 have the same friend, user 1; user 5 and user 11 both follow user 12, etc. Based on these meta-paths, this embodiment can obtain the high-order semantic vector of user 5 through a high-order semantic encoder.

[0103] 6. Biased contrast fusion

[0104] Based on the low-order structure vector and high-order semantic vector of user 5 obtained above, these two vectors are complementaryly fused using a biased contrastive fusion method to obtain the final vector of user 5. For all users 1, 2, 3, ..., n, their embedded low-dimensional vectors are obtained in the same way.

[0105] 7. Applications of low-dimensional vectors

[0106] Based on the low-dimensional vectors of all users obtained above, the similarity distance between related users and other users can be determined. Then, based on the similarity distance, several close user sets of the related users are recommended. Subsequently, following a similar pattern, N potentially related close users are selected and stored. Finally, the N close users recommended by the social networking site's backend data calculation server are displayed on the user interface as a reference for friend recommendations.

[0107] After obtaining similar users based on this application, similar users can be recommended to the current user on social networks, and products purchased by similar users can be recommended to the current user on shopping apps.

[0108] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for discovering users of the same type in heterogeneous networks based on multi-granularity fusion, characterized in that, Includes the following steps: Collect relevant user feature information and relationship data, and complete and align any missing or incomplete parts of the collected information; Aggregate user nodes that are directly contacted by relevant users and merge the low-level structure of the user network; The process of obtaining low-level structural information about a user network includes: Get relevant users The set of first-order neighbor nodes, and the category of the set of contacted users; The users contacted in each category are merged separately according to the category to obtain the merging result for each category; The fusion results for each category are compared with those of the relevant users. The data is then fused to obtain the low-order structure vector of the network. Aggregate the meta-path information of relevant users and integrate the higher-order semantics of the user network; Using a multi-granularity cognitive computing approach, low-order structural vectors and high-order semantic vectors are complementaryly aggregated to obtain low-dimensional vectors of relevant users; the process of aggregating the meta-path information of relevant users includes: The metapath is encoded using GCN, and the result of the last layer of GCN is used as the encoding based on that metapath. The encodings of all metapaths are fused to obtain the high-order semantic vector of the user network; Biased contrastive fusion enables complementary aggregation of low-order structural vectors and high-order semantic vectors. During the aggregation process, a weight matrix is ​​used to make the fusion focus biased towards high-order semantic vectors. At the same time, the cosine distance between the low-order structural vectors and high-order semantic vectors is used as a loss function for optimization until the cosine distance between the optimized low-order structural vectors and high-order semantic vectors is less than the set threshold. The fused low-order structural vectors and high-order semantic vectors are obtained, and one of them is used as the final low-dimensional vector of the user network. Project the low-dimensional vectors of relevant users onto a two-dimensional coordinate system, identify the N users whose low-dimensional vectors are closest to the relevant users as similar users, and recommend similar users to the relevant users.

2. The method for discovering users of the same type in a heterogeneous network based on multi-granularity fusion according to claim 1, characterized in that, Based on the acquired information, the similarity distance between the relevant user and the people they come into contact with is calculated using the low-dimensional vector of the user network. Based on the similarity distance, several close user sets of the relevant user are predicted.

3. The method for discovering users of the same type in a heterogeneous network based on multi-granularity fusion according to claim 1, characterized in that, Get relevant users The process of refining content attribute information involves generating a random walk sequence, updating the parameters using a skip-gram model based on the generated sequence, and generating the final node representation matrix, including: in, Therefore A random walk sequence starting from , This represents a heterogeneous neural network, where t is the length of the walk sequence; This is a user attribute content matrix, where each row of the matrix represents... For relevant users A content attribute; is the node embedding size, w is the wandering window size; |V| is the number of network users.

4. The method for discovering users of the same type in a heterogeneous network based on multi-granularity fusion according to claim 1, characterized in that, If the users contacted in each category are merged separately according to the category, then the categories are: The user integration process can be represented as follows: in, Indicates category as The results of user integration. To reach users For relevant users Importance For relevant users The category of the contact user set is One of the nodes; Indicates relevant users The type of the contact user set is Similar users.

5. The method for discovering users of the same type in a heterogeneous network based on multi-granularity fusion according to claim 4, characterized in that, If there are K other types of users who have been contacted, then the relevant users The process of integrating with all types of users is represented as follows: in, This represents the low-order structural information vector of the network; For type Contact users to relevant users The degree of correlation is calculated through an attention mechanism.

6. The method for discovering users of the same type in a heterogeneous network based on multi-granularity fusion according to claim 1, characterized in that, In the process of encoding the metapath using GCN, the output of the (l+1)th layer of GCN is represented as: in, For relevant users Features in layer l For user v j Features in the l-th layer; For relevant users The neighbor set includes itself; PReLU() represents the activation function; C ij W represents the normalization factor. (l) b is the weight of the l-th layer; (l) is the intercept of the l-th layer.

7. A device for discovering users of the same type in a heterogeneous network based on multi-granularity fusion, characterized in that, The method for discovering users of the same type in a heterogeneous network based on multi-granularity fusion as described in claim 1 includes a feature extraction module, an attribute completion module, a structural encoder module, a semantic encoder module, a biased contrastive fusion module, an encoding fusion module, and a recommendation module, wherein: The feature extraction module is used to extract content attribute information of relevant users; The attribute completion module is used to collect and complete the content attributes of relevant users; The structure encoder module is used to aggregate the first-order neighbor nodes of relevant users and merge the user's low-order structure; The semantic encoder module is used to aggregate the meta-path information of relevant users and integrate the user's higher-order semantics; The biased contrast fusion module is used to perform complementary aggregation of low-order structure and high-order semantics, and finally obtain the low-dimensional vector of the relevant user. The recommendation module projects the user's low-dimensional vector onto a two-dimensional coordinate system, identifies the N users whose low-dimensional vector is closest to the current user as similar users of related users, and recommends similar users to related users.

Images