A Movie Recommendation Method and System Based on Weighted Heterogeneous Information Network Representation Learning

Abstract

This invention discloses a movie recommendation method and system based on weighted heterogeneous information network representation learning, relating to the field of data mining, for better movie recommendation. The invention includes: (1) using GraphSAGE to learn the base class embeddings of nodes in a heterogeneous information network; (2) learning the edge embeddings of each node in different edge types, and aggregating all learned edge embeddings as the final edge embedding of the node; (3) fusing the encoding vector, base class embeddings, and final edge embeddings to obtain the node vector representations of the node in different types of sub-networks; (4) predicting user ratings for movies based on the node vector representations, applying a linear regression algorithm to assign weights to the user ratings of movies, and weighting to obtain the final rating; (5) selecting at least one movie with the highest final rating as the recommendation target. This invention fully utilizes the additional attribute vectors of nodes in a heterogeneous information network, improving recommendation accuracy.

Description

Technical Field

[0001] This invention relates to the field of data mining, and in particular to a movie recommendation method and system based on weighted heterogeneous information network representation learning. Background Technology

[0002] The widespread adoption of the internet has greatly facilitated people's lives and enriched their entertainment. However, the explosive growth of data has also brought about data redundancy, making it difficult for users to quickly find the information they are interested in when faced with massive amounts of online resources. This, in turn, reduces the efficiency of online data utilization. The role of recommendation systems is to help users filter out redundant information from massive amounts of data and quickly find content that interests them.

[0003] Recently, movie recommendation methods have become an important research field, attracting scientists and researchers worldwide. A major challenge facing movie recommendation systems is how to connect movies with users across various movie datasets and help users obtain useful information. Movie recommendation methods aim to provide users with accurate movie recommendations, effectively addressing the information overload problem caused by massive amounts of data.

[0004] In fact, to improve the recommendation performance of movie recommendation systems, it is almost always necessary to learn how to obtain more effective historical data embeddings (vectors) so that neural networks can learn the features of historical data more effectively, thereby making movie recommendations more efficient and accurate. Summary of the Invention

[0005] The purpose of this invention is to address all or part of the aforementioned problems by providing a movie recommendation method and system based on weighted heterogeneous information network representation learning. The main objective of this invention is to combine weighted heterogeneous information network representation learning with movie recommendation methods to more deeply uncover the implicit correlation features between users and movies, thereby enabling better movie recommendations.

[0006] The technical solution adopted in this invention is as follows:

[0007] A movie recommendation method based on weighted heterogeneous information network representation learning includes:

[0008] A weighted heterogeneous information network is constructed using a dataset. The network has users and movies as nodes and user ratings of movies as edges. The weights of each edge are abstracted into different edge types based on the user ratings of movies. The weighted heterogeneous information network is then split into subnetworks consisting of only one type of edge based on the edge type.

[0009] The encoding vectors of each node are learned using a graph neural network to obtain the base class embeddings of each node. The encoding vectors are initialization vectors obtained by encoding the node attributes.

[0010] For each node, edge embeddings in different types of sub-networks are learned, and the edge embeddings learned from all types of sub-networks are aggregated to form the final edge embedding for that node.

[0011] By fusing the encoding vector, base class embedding, and final edge embedding under the aggregate weights of each type of sub-network for each node, the node vector representation of that node under each type of sub-network is obtained;

[0012] The user's rating for the movie is predicted based on the vector representation of all nodes. A linear regression algorithm is applied to assign weights to the user's rating for the movie, and the final rating for the movie is obtained by weighting the results.

[0013] Choose at least one of the highest-rated movies as your recommendation target.

[0014] Furthermore, a graph neural network is used to learn the encoding vector of each node to obtain the base class embedding of each node, including:

[0015] The encoding vectors of the target node's neighboring nodes are aggregated at a depth of K1 and spatially transformed to obtain the base class embedding of the target node; the aggregation method includes:

[0016] For each node v, the encoded vector is used as the vector representation of the aggregation depth at layer 0;

[0017] In each aggregation depth k, the aggregation vector of node v's neighboring node N(v) at the previous aggregation depth k-1 is calculated.

[0018] Will The vector representation of node v at the previous aggregation depth k-1. After concatenation and nonlinear transformation, the vector representation of node v at aggregation depth k is obtained.

[0019] The space is transformed by using a transformation matrix to convert the vector representations of all nodes at the K1 level to a unified space.

[0020] Furthermore, The vector representation of node v at the previous aggregation depth k-1 is concatenated with the vector representation of node v at the previous aggregation depth k, and after a nonlinear transformation, the vector representation of node v at the aggregation depth k is obtained. Specifically:

[0021]

[0022] Where MEAN() represents the MEAN aggregation function, W is the aggregation weight matrix, and σ() is the Sigmoid function.

[0023] Furthermore, for each node, edge embeddings in different types of sub-networks are learned, and the edge embeddings learned from all types of sub-networks are aggregated, including:

[0024] For each node, in each type of subnetwork, the aggregation features of the neighboring nodes are used as the edge embeddings of that node in that type of subnetwork;

[0025] Edge embeddings from different types of subnetworks are spliced ​​and aggregated.

[0026] Furthermore, the attention mechanism is used to calculate the aggregate weights of the final edge embedding of nodes in different types of subnetworks.

[0027] Furthermore, the method for calculating aggregation weights using the attention mechanism is as follows:

[0028] Among them, a i，r Represents node v i Aggregate weights in a subnetwork of type r These are the parameter vector and parameter matrix to be learned, respectively, d a U represents the edge embedding length, s represents the number of subnetwork types, and U represents the number of subnetwork types. i Represents node v i The final edge embedding.

[0029] Furthermore, for each node, in each type of sub-network, the aggregated features of its neighboring nodes are used as the edge embeddings of that node in that type of sub-network, including:

[0030] For each node, based on the set aggregation depth K2, in each type of subnetwork, the aggregation vector of the neighboring nodes in the previous aggregation depth in the subnetwork of that type is used as the vector representation of the node at the current aggregation depth, and the vector representation at the highest aggregation depth is used as the edge embedding in the subnetwork of that type.

[0031] Furthermore, by fusing the encoding vector, base class embedding, and final edge embedding under the aggregated weights of each type of sub-network for each node, the node vector representation of that node under each type of sub-network is obtained, including:

[0032]

[0033] Among them, v i，r For node v i In a subnetwork of type r, the node vector representation, h z (x i ) is node vi Base class embedding, x i For node v i The vector representation of h in layer K1 z () is a transformation function that performs spatial transformation on nodes of type z, α r and β r All of these are hyperparameters of the corresponding subnetwork type r, D z It is a node v of type z i Feature transformation matrix, M r Em is a trainable transformation matrix. i Represents node v i The encoded vector, z∈(user, movie).

[0034] Furthermore, with the objective function value as the goal, the network parameters are obtained during the calculation of the node vector representation. The objective function is:

[0035]

[0036] The subnetwork of type r is denoted as G. r =(V r E r V r Represents subnetwork G r The set of nodes in the network G r The random walk path of the middle node is l is the path length, and node v pt The background vector is C = {v} pk |v pk ∈P, |kt|≤c, t≠k}, where c is the sampling window size, E r Represents subnetwork G r The set of edges; v j ∈V r P t (v) indicates that in V r The random distribution defined in c j Represents node v j Background vector, v k ~P t (v) indicates from P t Noise nodes v randomly selected from (v) k L represents the number of noise nodes, c k Represents node v k The background vector, E[x] represents the expectation of x, and E represents the objective function value.

[0037] To address the aforementioned issues, this invention also provides a movie recommendation system based on weighted heterogeneous information network representation learning, which executes the aforementioned movie recommendation method based on weighted heterogeneous information network representation learning.

[0038] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0039] 1. This invention makes full use of the additional attribute vectors of nodes in heterogeneous information networks. The vector representation of the nodes fully preserves the network structure and node attribute information, and further explores the contribution value of historical data to movie recommendation.

[0040] 2. This invention learns the mapping from node attributes to base class embeddings and edge embeddings. When a new node is added, it can also learn the base class embedding and edge embedding of the new node based on the attribute information of the new node, so as to recommend movies more accurately and effectively solve the cold start problem in movie recommendation.

[0041] 3. This invention fully considers the differences in the influence of different edge types and uses an attention mechanism to calculate the weight of each edge embedding, which can better simulate the user's individual interests and help improve the accuracy of recommendations. Attached Figure Description

[0042] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:

[0043] Figure 1 This is a structural diagram of a heterogeneous information network provided in an embodiment of the present invention.

[0044] Figure 2 This is a data flow diagram illustrating a movie recommendation method based on weighted heterogeneous information network representation learning, provided in an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of the user feature and movie feature processing network of the present invention. Detailed Implementation

[0046] All features disclosed in this specification, or steps in all methods or processes disclosed herein, may be combined in any way, except for mutually exclusive features and / or steps.

[0047] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0048] Example 1

[0049] A movie recommendation method based on weighted heterogeneous information network representation learning includes the following steps:

[0050] (1) A weighted heterogeneous information network is constructed using the dataset. This network uses users and movies as nodes, and user ratings of movies as edges. The edge weights are abstracted into different types based on user ratings of movies. The ratings can be standardized to 1-N, with each rating representing a different edge type, such as... Figure 1 In the example shown, N is 5, meaning the edges of the weighted heterogeneous information network are abstracted into 5 types. Then, based on the edge types, the weighted heterogeneous information network is split into subnetworks consisting of only one type of edge.

[0051] (2) Use the GraphSAGE algorithm (a type of graph neural network) to learn the encoding vector of each node and obtain the base class embedding of each node.

[0052] The encoding vector of a node is the initialization vector obtained by encoding the node's attributes. The encoding method can be numeric encoding, Mutli-Hot encoding, etc.

[0053] In some embodiments, this step includes: performing K1-depth aggregation (aggregation depth K1 is obtained by configuration) and spatial transformation on the encoding vectors of the target node's neighboring nodes to obtain the base class embedding of the target node.

[0054] The aggregation methods include:

[0055] For each node v, the encoded vector is used as the vector representation of the aggregation depth at layer 0;

[0056] In each aggregation depth k, the aggregation vector of node v's neighboring node N(v) at the previous aggregation depth k-1 is calculated.

[0057] Will The vector representation of node v at the previous aggregation depth k-1. After concatenation and nonlinear transformation, the vector representation of node v at aggregation depth k is obtained. In practice, this step is implemented as follows:

[0058]

[0059] Where MEAN() represents the MEAN aggregation function, W is the aggregation weight matrix, and σ() is the Sigmoid function.

[0060] Additionally, this step is performed using the function h. z (x i To achieve spatial transformation, where x iFor node v i The vector representation of h in layer K1 z () is a transformation function that performs spatial transformation on nodes of type z so that nodes of different types are embedded and projected into the same space, z∈(user, movie).

[0061] (3) Learn edge embeddings in different types of subnetworks for each node, and aggregate the edge embeddings learned from all types of subnetworks as the final edge embedding for that node.

[0062] In some embodiments, this step includes:

[0063] For each node, in each type of subnetwork, the aggregation features of the neighboring nodes are used as the edge embeddings of that node in that type of subnetwork;

[0064] Edge embeddings from different types of subnetworks are spliced ​​and aggregated.

[0065] Different edge embeddings correspond to different types of subnetworks, and different types of subnetworks have varying influences on movie recommendations. Therefore, it is necessary to consider the contribution of edge embeddings under different types of subnetworks. By calculating the aggregation weights under different types of subnetworks, and then weighting the final edge embeddings of nodes according to the aggregation weights, we can obtain the contribution of the final edge embeddings of nodes to the node vector representation under different types of subnetworks. In some embodiments, an attention mechanism is used to calculate the aggregation weights of the final edge embeddings of nodes in different types of subnetworks. The calculation method of the aggregation weights is as follows: Among them, a i，r Represents node v i Aggregate weights in a subnetwork of type r These are the parameter vector and parameter matrix to be learned, respectively, d a U represents the edge embedding length, s represents the number of subnetwork types, and U represents the number of subnetwork types. i Represents node v i The final edge embedding.

[0066] In the specific implementation of this step, for each node, according to the set aggregation depth K2, in each type of sub-network, the aggregation vector of the neighboring node at the previous aggregation depth is used as the vector representation of the node at the current aggregation depth, and the vector representation at the highest aggregation depth is used as the edge embedding in the sub-network of that type.

[0067] (4) By fusing the encoding vector, base class embedding, and final edge embedding of each node under the aggregate weights of each type of subnetwork, the node vector representation of the node under each type of subnetwork is obtained.

[0068] In some embodiments, this step is implemented as follows:

[0069]

[0070] Among them, v i，r For node v i In a subnetwork of type r, the node vector representation, h z (x i ) is node v i Base class embedding, x i For node v i In the vector representation of layer K1, h z () is a transformation function that performs spatial transformation on nodes of type z, α r and β r All of these are hyperparameters of the corresponding subnetwork type r, D z It is a node v of type z i Feature transformation matrix, M r Em is a trainable transformation matrix. i Represents node v i The encoded vector.

[0071] The node vector representation of a node involves multiple network parameters, which are obtained through training. In some embodiments, this is achieved by setting a target function:

[0072]

[0073] The network parameters are trained to compute the node vector representation by minimizing the objective function value. The subnetwork of type r is represented as G. r =(V r E r V r Represents subnetwork G r The set of nodes in the network G r The random walk path of the middle node is l is the path length, and node v pt The background vector is C = {v} pk |v pk ∈P, |kt|≤c, t≠k}, where c is the sampling window size, E r Represents subnetwork G r The set of edges; v j ∈V r P t (v) indicates that in V r The random distribution defined in c j Represents node v j Background vector, v k ~P t (v) indicates from P t Noise nodes v randomly selected from (v) kL represents the number of noise nodes, c k Represents node v k The background vector, E[x] represents the expectation of x, and E represents the objective function value.

[0074] (5) Predict the user's rating for the movie based on the node vector representations of all nodes. For example, take the inner product of the node vector representations of the movie node and the node vector representations of the user node to obtain the predicted rating. Apply a linear regression algorithm to assign weights to the user's predicted rating for the movie, and calculate the final rating for the movie using a weighted average.

[0075] (6) Select at least one of the highest-rated movies as the recommended movie.

[0076] Example 2

[0077] A movie recommendation method based on weighted heterogeneous information network representation learning, such as Figure 2 As shown, it includes the following steps:

[0078] (1) A weighted heterogeneous information network is constructed using a dataset containing movie attributes and user attributes. The weighted heterogeneous information network has users and movies as nodes and user ratings of movies as edges.

[0079] (2) The GraphSAGE algorithm is used to learn the base class embeddings of each node in a weighted heterogeneous information network. This mainly includes the following steps:

[0080] a. For node v i The node attributes (user attributes or movie attributes) are encoded accordingly to obtain node v. i The encoded vector is used as the initial input vector Em for GraphSAGE. i .

[0081] Tables 1 and 2 show the composition of user files and movie files. Therefore, a feature processing network was designed, such as... Figure 3As shown, the network is divided into a user feature network and a movie feature network. The user features UserID, Gender, Age, and Job can be considered categorical data. The preprocessing method here is to encode these data as numbers and use them as indices in the embedding matrix. This avoids the problem of one-hot encoding, commonly used for categorical data, which can result in each value being encoded as a high-dimensional sparse vector, leading to a high computational cost for the neural network. The processing method for UserID is the same as for Age, Gender, and Job. First, the input is used as an index to extract the corresponding feature vector from the embedding matrix, where UserID is encoded as a 32-dimensional vector, and the other features are encoded as 16-dimensional vectors. Then, a fully connected layer (fc) and a dropout layer are added after each feature vector. The fully connected layer has 32 neurons. Finally, the four 32-dimensional vectors are concatenated to form a 128-dimensional vector, which is used as the input to the fully connected layer. The final output is a 200-dimensional user feature vector. For movie features, MovieID can be considered categorical data, and the preprocessing method is the same as above. However, movie title and movie genre are special; they can be regarded as multi-valued attributes with varying lengths. For movie genres, since there are not many types, Mutli-Hot encoding can be used directly. In the neural network, the input of different lengths is achieved by multiplying the encoded vector with the embedding matrix. The processing of movie title is slightly more complex. First, a word->int mapping dictionary is created, then a list of numbers is used for encoding and padded to the same length. This is passed through an LSTM network, and the average of all network outputs is calculated to obtain the movie title feature. The processing method for MovieID and Genres is the same as that for user features. The only difference is that the first layer of Genres uses matrix multiplication instead of lookup. Each title is encoded into a [15, 32]-dimensional feature after passing through the embedding layer, and then passed through an LSTM with a hidden layer of 128 neurons. The average of the outputs of these 15 LSTM units is calculated (note that the padding value is not included in the calculation), and finally a 128-dimensional feature vector is obtained. This vector is concatenated with the output vectors of MovieID and Genres and used as the input to the fully connected layer, resulting in a 200-dimensional vector that serves as the movie feature vector.

[0082] Table 1. User File Diagram

[0083]

[0084] Table 2. Schematic diagram of movie files

[0085] MovieID Movie ID Title Movie title Genres Film Genre Various film genres are separated by "|".

[0086] b. Set the aggregation depth K1.

[0087] c. At each aggregation depth k (k∈[1, K1]), for a node v in the network, select a fixed number of its neighbors, and aggregate the attributes of the neighboring nodes to the aggregation at the k-th layer.

[0088] d. will The (k-1)th layer vector representation of node v The data is concatenated and then subjected to a nonlinear transformation to generate the k-th layer embedding representation of vertex v. Here, the aggregation function for node v can be the average aggregation function: Alternatively, you can choose pooling aggregation functions: Where N(v) represents the neighboring nodes of node v, MEAN() represents the MEAN aggregation function, W is the aggregation weight matrix, and σ() is the Sigmoid function.

[0089] e. Finally, the base class embedding of the node is obtained. i =h z (x i ), where h z (x i ) is node v i Base class embedding, x i For node v i In the vector representation of layer K1, h z () is a transformation function that performs spatial transformation on nodes of type z, where z∈(user, movie).

[0090] (3) The edge weights in the weighted heterogeneous information network are abstracted into five different types of edges, such as Figure 1 As shown, the ratings for movies are standardized to 1-5 based on user ratings, with each score representing a different edge type. The original network is then split into subnetworks containing only one edge type based on the edge type.

[0091] (4) Learn the edge embeddings of each node in different edge type subnetworks, and aggregate the learned edge embeddings in all edge types as the final edge embedding of the node.

[0092] The specific steps include:

[0093] a. Node v i Aggregate the k-th order neighbors of edge type r (r∈[1,5]) to obtain the edge embeddings in the subnetwork of type r.

[0094]

[0095] Where N i，r It is node v iNeighboring nodes in a subnetwork of type r. Node v i The vector corresponding to the subnetwork of type r Random initialization. The aggregation function can be the mean aggregation:

[0096] b. Aggregate the edge embeddings of nodes in different types of subnetworks: U i =(u i，1 u i，2 ,...u i，m ), which serves as the final edge embedding for that node.

[0097] c. Considering the different impacts of different types of edges, the aggregation weights for different types of subnetworks are calculated using an attention mechanism: Among them, a i，r Represents node v i Aggregate weights in a subnetwork of type r d is the parameter to be learned. a represents the edge embedding length, and s represents the number of subnetwork types.

[0098] (5) Combining the encoding vector obtained in step (2), the base class embedding, and the final edge embedding obtained in step (4), we obtain node v. i The final node vector representation in a subnetwork of type r:

[0099]

[0100] Among them, v i，r For node v i In a subnetwork of type r, the node vector representation, α r and β r All of these are hyperparameters of the corresponding subnetwork type r, D z It is a node v of type z i Feature transformation matrix, M r is a trainable transformation matrix.

[0101] As can be seen from the expression of the node vector representation, the process of the model extracting the node vector representation involves multiple parameters to be learned. This invention uses a meta-path-based random walk method and skip-gram to learn the model parameters.

[0102] Given a subnetwork G of type r r =(V r E r V r Represents subnetwork G r The set of nodes in E r Represents subnetwork G rThe set of edges. And a given metapath Γ of length l: V1→V2→…→V t →…→V l Then the transition probability at step t is:

[0103]

[0104] That is, an edge can only be sampled if it belongs to type r and the node type matches the node type in the metapath.

[0105] Assume network G r The random walk path of the middle node is If l is the path length, then node v pt The background vector is C = {v} pk |v pk ∈P, |kt|≤c, t≠k}, where c is the sampling window size, i.e., the background vector consists of nodes within a range of length c to the left and right of the current node in the random walk path. Therefore, given node v i Given the background vector C from the random walk path, the objective function can be expressed as:

[0106]

[0107] Where θ represents all network parameters, v j ∈V r c j Represents node v j The background vector, c k Represents node v k The background vector.

[0108] Finally, heterogeneous negative sampling is used to construct the objective function: Among them, P t (v) indicates that in V r The random distribution defined in the document, v k ~P t (v) indicates from P t Noise nodes v randomly selected from (v) k L represents the number of noise nodes, E[x] represents the expectation of x, and E represents the objective function value.

[0109] By minimizing the above objective function value, the model parameters can be trained and obtained.

[0110] (6) Predict the user's rating of the movie based on the node vector representation. For example, use a collaborative filtering algorithm, or take the inner product of the node vector representation of the movie node and the node vector representation of the user node to obtain the user's preliminary predicted rating of the movie. Then apply a linear regression algorithm to assign weights to the user's rating of the movie, and sum the weights to obtain the final rating.

[0111] (7) The movie with the highest final rating is the movie recommended to the user.

[0112] Example 3

[0113] This embodiment corresponds to Embodiment 1 or 2 above, and proposes a movie recommendation system based on weighted heterogeneous information network representation learning. This system executes the movie recommendation method based on weighted heterogeneous information network representation learning in the above embodiments.

[0114] Specifically, the system is configured with a storage medium and a processor. The storage medium stores a computer program that can be run to execute the methods described in the above embodiments. The processor is connected to the storage medium to run the computer program stored in the storage medium.

[0115] Implementation 4

[0116] This embodiment proposes a computer software product that is configured via a computer program to execute the methods in Embodiment 1 or 2 above.

[0117] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A movie recommendation method based on weighted heterogeneous information network representation learning, characterized in that, include: A weighted heterogeneous information network is constructed using a dataset. The network has users and movies as nodes and user ratings of movies as edges. The weights of each edge are abstracted into different edge types based on the user ratings of movies. The weighted heterogeneous information network is then split into subnetworks consisting of only one type of edge based on the edge type. The base class embeddings of each node are obtained by learning the encoding vectors of each node using a graph neural network. This includes: performing K1-depth aggregation and spatial transformation on the encoding vectors of the target node's neighboring nodes to obtain the base class embeddings of the target node; the aggregation method includes: For each node v, the encoded vector is used as the vector representation of the 0th aggregation depth; in each aggregation depth k, the aggregation vectors of node v's neighboring nodes N(v) at the previous aggregation depth k-1 are calculated. , ;Will The vector representation of node v at the previous aggregation depth k-1. After concatenation and nonlinear transformation, the vector representation of node v at aggregation depth k is obtained. The space is transformed into a unified space by using a transformation matrix to convert the vector representations of all nodes at the K1 level; the encoded vector is an initialization vector obtained by encoding the node attributes; For each node, based on the set aggregation depth K2, in each type of sub-network, the aggregation vector of the neighboring nodes in the previous aggregation depth in the sub-network of that type is used as the vector representation of the node in the current aggregation depth, and the vector representation under the highest aggregation depth is used as the edge embedding in the sub-network of that type. The edge embeddings in different types of sub-networks are concatenated and aggregated to become the final edge embedding of the node. By fusing the encoding vector, base class embedding, and final edge embedding under the aggregate weights of each type of sub-network for each node, the node vector representation of that node under each type of sub-network is obtained; The user's rating for the movie is predicted based on the vector representation of all nodes. A linear regression algorithm is applied to assign weights to the user's rating for the movie, and the final rating for the movie is obtained by weighting the results. Choose at least one of the highest-rated movies as your recommendation target.

2. The movie recommendation method based on weighted heterogeneous information network representation learning as described in claim 1, characterized in that, Will The vector representation of node v at the previous aggregation depth k-1. After concatenation and nonlinear transformation, the vector representation of node v at aggregation depth k is obtained. Specifically: ， in, This represents the MEAN aggregation function, where W is the aggregation weight matrix. This is the Sigmoid function.

3. The movie recommendation method based on weighted heterogeneous information network representation learning as described in claim 1, characterized in that, The attention mechanism is used to calculate the aggregate weights of the final edges of nodes embedded in different types of subnetworks.

4. The movie recommendation method based on weighted heterogeneous information network representation learning as described in claim 3, characterized in that, The method for calculating aggregate weights using the attention mechanism is as follows: ,in, Represents a node Aggregate weights in a subnetwork of type r These are the parameter vector and parameter matrix to be learned, respectively. Indicates the edge embedding length. Indicates the number of sub-network types. Represents a node The final edge embedding.

5. The movie recommendation method based on weighted heterogeneous information network representation learning as described in claim 4, characterized in that, By fusing the encoded vector, base class embedding, and final edge embedding under the aggregated weights of each type of sub-network for each node, the node vector representation of that node under each type of sub-network is obtained, including: ， in, For nodes In type r The node vector representation in the subnetwork, For nodes Base class embedding, For nodes Vector representation in layer K1, This is a transformation function that performs spatial transformations on nodes of type z. and These are all hyperparameters corresponding to the sub-network type r. It is a node of type z The feature transformation matrix, For a trainable transformation matrix, Represents a node The encoded vector, .

6. The movie recommendation method based on weighted heterogeneous information network representation learning as described in claim 5, characterized in that, The network parameters are trained to calculate the node vector representation with the goal of minimizing the objective function value. The objective function is: Among them, the type is r The subnetwork is represented as , Subnetwork The set of nodes in the network The random walk path of the middle node is , For path length, node The background vector is , The sampling window size, Subnetwork The set of edges; , Indicates in The random distribution defined in [the document], Represents a node The background vector, Indicates from Noise nodes randomly selected from the data L represents the number of noise nodes. Represents a node The background vector, Indicates "correct" expect, This represents the objective function value.

7. A movie recommendation system based on weighted heterogeneous information network representation learning, characterized in that, The system executes the movie recommendation method based on weighted heterogeneous information network representation learning as described in any one of claims 1-6.

Images