Interactive visual recommendation method based on heterogeneous network information embedding model

Abstract

The interactive visual recommendation method based on the heterogeneous network information embedding model comprises the following steps: step 1, data crawling and cleaning; collecting douban user data and douban movie data from a real douban movie website, and cleaning; step 2, obtaining parameters and feature data in the model training process, constructing a heterogeneous movie information network as the input of the model for the pretreated movie data, then training using the heterogeneous embedding models HetGNN, KGAT and NIRec, and reserving relevant parameter and feature information in the training process; step 3, user interaction-based model comparison and exploration, based on the reserved path feature data, designing corresponding visual charts according to different index data and model parameters; step 4, interactive visual recommendation based on the heterogeneous embedding model, learning multiple types of objects and relationships in the heterogeneous graph, and mining the implicit rich structure and semantic information for the recommendation task. The present application solves the black box problem of recommendation and increases the explainability of the recommendation result.

Description

Technical Field

[0001] This invention proposes a visual analysis method for heterogeneous network embedding models. This method systematically explores and compares the differences in performance of three representative embedding models for downstream recommendation tasks, the common patterns existing in the network embedding process, and the source information of the recommendation results, thereby increasing the interpretability of the recommendations. Background Technology

[0002] In the real world, objects and their relationships are usually of multiple types. In order to better preserve their complex structural information and rich semantic information in data mining, researchers integrate different types of objects and their interaction information, and even information from different data sources, into an information network, which is called a heterogeneous information network.

[0003] In recent years, with the development of deep neural networks, some researchers have attempted to apply deep models to heterogeneous network embedding. Compared to shallow models, deep models use neural networks to learn embeddings from node attributes and interactions between nodes, which can better capture nonlinear relationships and thus retain more structural and semantic information. However, methods using meta-paths for heterogeneous network embedding are considered to have certain limitations due to their requirement for domain knowledge; methods using information transfer mechanisms to model higher-order relationships are considered to introduce noise because they aggregate features from all surrounding neighbors. These two methods theoretically have their own advantages and disadvantages. However, what are the differences in user features learned through different methods? Do they effectively retain user information? Especially in recommendation tasks, the similar evaluation values ​​of different models do not necessarily mean consistent recommendation effectiveness. Does the same model retain personalization in its recommendations for different users? These questions are worthy of exploration for recommendation systems.

[0004] However, most current model visualization efforts focus on analyzing a single model and process data that is primarily directly observable image data. Heterogeneous network embedding maps high-dimensional features to low-dimensional vectors, making direct comparison or intuitive understanding of the features involved in the learning process impossible. Furthermore, while existing multi-model visual comparison methods can analyze the intrinsic features and performance of a model by visualizing the activation status of neurons, this approach is not suitable for heterogeneous network embedding models. Summary of the Invention

[0005] To overcome the above-mentioned shortcomings of the prior art, this invention proposes a visual comparative analysis method for heterogeneous network embedding models.

[0006] This invention first crawls and cleans the data used to evaluate the models, then unifies the downstream tasks and evaluation metrics of each model, uses the processed data as input to each model, performs heterogeneous network embedding learning, and retains a large number of parameters and learned feature information during the learning process; then, it extracts the effective features and designs a data visualization view to display them; finally, it designs interactive operations to mine abnormal data in the embedding results in order to evaluate the effectiveness of the models and explore common patterns among the models.

[0007] An interactive visual recommendation method based on a heterogeneous network information embedding model includes the following steps:

[0008] Step 1: Crawling and cleaning the data; collecting Douban user data and Douban movie data from the real Douban movie website, and cleaning it according to certain rules to ensure the validity of the data;

[0009] 1.1 Obtain the original dataset. Randomly select a Douban user and retain them based on the percentage of movies they have watched. Then, randomly select three users from their followed list and repeat the above process. Next, count the movie viewing records of all users, retain the valid movie IDs, and crawl the movie information.

[0010] 1.2 Data preprocessing is performed on the original dataset. To make the evaluation of the model more fair and effective, i.e. to prevent invalid data, the scheme filters the data based on conditions such as users having at least >50 movie viewing records and movies being watched by more than 20 valid users. Since recommendation models usually have a cold start problem, 10% of users with <50 movie viewing records are retained as cold start data for evaluation of special cases.

[0011] 1.3 Fair evaluation of the model: In order to make the comparative evaluation of the model effective and fair, the downstream objectives embedded in the model are unified as recommendation tasks, and their evaluation metrics are reconstructed into the most common evaluation metrics in recommendation tasks, namely accuracy, recall and AUC, which are calculated as follows.

[0012]

[0013]

[0014]

[0015] Where TP represents correctly classified positive samples, TN represents correctly classified negative samples, FP represents misclassified positive samples, FN represents misclassified negative samples, M is the number of positive samples, N is the number of negative samples, and rank... i This represents the predicted score of positive sample i.

[0016] In addition, a personalized index for the recommendation results was added, along with a similarity assessment of the recommendation results. This index calculates the cosine similarity between the recommendation lists of different users. The personalized score calculation process for user i is shown in formula (4):

[0017]

[0018] Where U represents the user set, q i Let represent a vector consisting of a recommendation list for user i.

[0019] Step 2: Obtain the parameters and feature data during the model training process;

[0020] The preprocessed film data from the above steps is used to construct a heterogeneous film information network as input to the model. Then, three representative heterogeneous network embedding models—HetGNN, KGAT, and NIRec—are used for training, preserving relevant parameters and feature information from the training process. The processing and data storage of the heterogeneous network embedding model includes the following steps:

[0021] 2.1 Obtain the meta-path features of the NIRec model;

[0022] To explore the differences in neighbor features learned by heterogeneous network embedding models based on meta-path neighbors using end-to-end interactive modeling methods, the path instance information enumerated by the heterogeneous network embedding models when calculating path attention weight parameters is preserved and used as visualization data.

[0023] The path information is divided into two parts: one is node feature information based on the instance path. Second, path feature information based on instance paths, the calculation methods are as follows.

[0024]

[0025]

[0026] h j ρ This represents instance path information based on path ρ in the interaction matrix, where W, α, β, and b are all trainable parameters, and N... ρj It represents neighborhood and single meta-path-guided, semantically specific aggregated information.

[0027] The path information above consists of the feature paths of all positive and negative samples. To ensure the fairness of the model comparison, the final recommendation result, i.e., the positive sample, is used to filter the paths, and only the instance path information that can reach the target node from the source node is retained as the visualization data.

[0028] 2.2 Obtain the attention weight features of the KGAT model;

[0029] To explore the differences between neighbor features learned by non-custom soft paths formed by combining a heterogeneous network embedding model based on information transfer mechanism with attention mechanism, the bidirectional attention weight parameter π(h,r,t) in the training process of this model is retained. This parameter represents the importance of neighbor nodes to source nodes, and the specific calculation formula is as follows.

[0030]

[0031] Where h represents the source node, t represents the tail node, r represents the relationship from the tail node to the source node, and N h This represents the set of all neighboring nodes surrounding the source node.

[0032] Next, the average attention value between different types of nodes is calculated based on the obtained data, and this value is used as the threshold for selecting neighbor nodes to construct metapaths. Neighbor nodes with values ​​higher than the threshold are considered important nodes, while those with values ​​lower than the threshold are discarded. Finally, a depth-first search algorithm is used to traverse the important neighbor nodes of each source node to generate metapaths of less than 4 that can reach the target node. These metapaths contain feature information from the source node to the target node, and are therefore used as visualization data for model comparison.

[0033] 2.3 Obtaining the meta-path features of the HetGNN model;

[0034] To explore the feature differences learned by the meta-path-based heterogeneous network embedding model during the neighbor node aggregation process, the path instances involved in the model learning process were preserved as visualization data for model comparison.

[0035] In the initialization process of the heterogeneous network embedding model, when generating the initial node embedding using the random walk algorithm with restart, sampling is performed proportionally according to the relative proportion of each type of node to the total number of nodes. For each node, sampling is performed on its movie neighbors, user neighbors, type neighbors, actor neighbors, and director neighbors. The instance path formed by these neighbors is calculated using the following formula.

[0036]

[0037] The above three models are used to learn from the original data. The node entity paths used by the NIREC model during the learning process are preserved as path fusion features; the attention weights of each node pair in the KGAT model during the learning process are preserved, and entity paths are constructed through network connection relationships as fusion features; the entity path information sampled by the HetGNN model during the learning process is preserved as path fusion features.

[0038] Step 3: Exploratory design based on model comparison using user interaction;

[0039] Based on the path feature data retained in the above steps, corresponding visualization charts will be designed according to different indicator data and model parameters. The specific steps are as follows:

[0040] 3.1 The data selection and detailed overview entry points for the visualization design are as follows: Based on the requirements of model researchers for the overview of model evaluation metrics, a circular stacked bar chart is used to display the evaluation metrics of the heterogeneous embedding models for the recommendation task. Each bar represents the recommendation evaluation result of a user, and each color represents a model. To facilitate data filtering, more interactive methods need to be designed, such as selecting a specific target by using a selection tool. In addition, a Venn diagram showing the relationship between recommendation results of different models is designed. The corresponding results are randomly scattered in the color areas corresponding to the models mentioned above, and the scatter points that match the user's viewing history are marked with a white border, indicating that they belong to the correctly predicted movies.

[0041] 3.2 Visualizing and Comparing Model Details: Based on the design and model developers' needs in step 3.1 to view detailed information about the embedded model during the neighbor aggregation process, a scalable force guidance graph was designed to display the meta-paths between the selected target and its recommendation results. These paths roughly include UMUM (user-movie-user-movie), UMGM (user-movie-genre-movie), UMAM (user-movie-actor-movie), and UMDM (user-movie-director-movie). The path width represents the importance of the path to the target node, and the node radius represents the higher the importance as the number of entity nodes it contains.

[0042] Step 4: Interactive visual recommendation based on heterogeneous embedding model;

[0043] Heterogeneous network embedding models aim to learn multiple types of objects and relationships in heterogeneous graphs, and mine the rich structural and semantic information hidden within them for recommendation tasks.

[0044] The heterogeneous embedding methods NIRec and HetGNN based on meta-paths for recommendation systems involve three steps: First, experts design effective meta-paths based on domain knowledge; second, the heterogeneous embedding model uses neurons to learn user neighborhood features through the aforementioned meta-paths, obtaining embedding vectors for both users and items to mine user preferences in the user-item interaction graph; third, the similarity between the user's embedding vector and all item embedding vectors is calculated, excluding training data, and the top K items with the highest similarity are used as the recommendation results.

[0045] KGAT, a heterogeneous embedding method based on information transmission mechanism, utilizes the natural connectivity of heterogeneous graphs to learn features from tail nodes connected to the target node through edges. The specific steps are as follows: 1. Obtain the embedding vectors of each node in the heterogeneous graph using the TransR method, common in the knowledge graph field; 2. Integrate the features of the target node's neighbors into the node's embedding vector through a propagation mechanism, and use an attention mechanism to calculate the importance of each neighbor node, thereby distinguishing the degree of integration of each feature; 3. After iterating through higher-order propagation, multiply the higher-order embedding vectors of the user and the item to obtain the final recommendation probability, and select the top K items with the highest recommendation probability as recommendations.

[0046] Finally, the recommendation result tracing view displays the path information between the recommended movies and the target user for different models. The meta-path-based method provides interpretability through meta-paths; the heterogeneous network embedding model uses these paths to find what other movies people who watched the same movies as the target user have watched, using this as a basis for recommendations and mining the preference features of similar users. The propagation mechanism-based method provides interpretability through attention scores in neighbor relationships; for example, nodes connected to the user with categories like "romance" and "drama" have higher attention weights, and the final recommended movies include these category features, indicating that the heterogeneous network embedding model's recommendations align with user preferences.

[0047] Preferably, in step 1, 20,000 Douban user data sets and 20,000 Douban movie data sets are collected from the actual Douban movie website.

[0048] Preferably, in step 2.3, the number of movie neighbors, user neighbors, genre neighbors, actor neighbors, and director neighbors sampled for each node are 25, 25, 25, 15, and 10, respectively.

[0049] The interaction method described in step 3.1 includes slider range selection.

[0050] This invention primarily addresses the following three issues: ① It enables a systematic comparison and analysis of three different heterogeneous information network embedding models to provide an understanding of the impact of different neighbor feature aggregation mechanisms on the embedding results; ② Through a bottom-up visualization method, it demonstrates the relationship between recommendation results and the model learning process, providing contextual information about the model and helping non-model developers understand the generation process of recommendation results; ③ By combining real Douban movie heterogeneous graph data with feature data from heterogeneous embedding models in the recommendation process, it solves the black-box problem of recommendation and increases interpretability.

[0051] The advantages of this invention can be summarized as follows: (1) Simple operation and strong expressiveness. Users only need to make simple selection operations on the visual interface; the system will display the recommendation results and their rich contextual features, and make intuitive comparisons and evaluations. (2) Strong innovation. There is currently no visual comparison system for heterogeneous network embedding models. This method can systematically compare and analyze three different architecture embedding models, which can help developers understand the impact of different neighbor feature aggregation mechanisms on the embedding results. Through case analysis, the neighbor interaction learning problem of the NIRec model and the neighbor sampling fusion problem of the HetGNN model were discovered, proving the effectiveness of the system in mining the model neighbor fusion mechanism. (3) Practical significance. The recommendation result traceability graph can help non-model developers quickly understand the origin of the recommendation results and the relationship between the model and the recommendation results. Non-developers can combine the viewing history heatmap and the correlation histogram to understand the contextual information of the recommended movies and the association between the recommended movies and user preferences, which solves the black box problem of recommendation and increases the interpretability of the recommendation results. Attached Figure Description

[0052] Figure 1 This is the overall flowchart of the method of the present invention.

[0053] Figure 2 This is the overall framework diagram of the KGAT model for visual comparison of the present invention.

[0054] Figure 3 This is the overall framework diagram of the NIRec model used for visual comparison in this invention.

[0055] Figure 4 This is the overall framework diagram of the HetGNN model for visual contrast in this invention.

[0056] Figure 5 This is the actual system interface of the method of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0058] This embodiment provides an interactive visual recommendation method for movie recommendations on the Internet based on a heterogeneous network information embedding model.

[0059] Referring to the overall flowchart of the technical solution, this invention comprises four stages: raw data crawling and processing, data preprocessing, model training to obtain model parameters and feature data, and visual recommendation. The data acquisition stage primarily involves crawling and cleaning data from the Douban Movie website, removing user data with limited viewing history and movie data with few views by valid users. The data preprocessing stage includes constructing a heterogeneous information network from the raw data, unifying the downstream tasks of various models, and defining model evaluation metrics. The model training stage uses the constructed heterogeneous movie information network as input data to train HetGNN, KGAT, and NIRec models, retaining relevant parameters and feature information from the training process. Finally, a visualization design is implemented for the relevant features and results from the model training process, adding corresponding interactive functions to achieve interpretable movie recommendations and compare recommendation results between different models.

[0060] The interactive visual recommendation method based on heterogeneous network information embedding model has the following specific steps:

[0061] Step 1: Data crawling and cleaning; We collected 20,000 Douban user data and 20,000 Douban movie data from the real Douban movie website, and cleaned them according to certain rules to ensure the validity of the data.

[0062] 1.1 Obtain the original dataset. Randomly select a Douban user and retain them based on the percentage of movies they have watched. Then, randomly select three users from their followed list and repeat the above process. Next, count the movie viewing records of all users, retain the valid movie IDs, and crawl the movie information.

[0063] 1.2 Data preprocessing is performed on the original dataset. To make the evaluation of the model more fair and effective, i.e. to prevent invalid data, the scheme filters the data based on conditions such as users having at least >50 movie viewing records and movies being watched by more than 20 valid users. Since recommendation models usually have a cold start problem, 10% of users with <50 movie viewing records are retained as cold start data for evaluation of special cases.

[0064] 1.3 Fair evaluation of the models: The downstream task of the three embedding models used for comparison is unified as recommendation, and their evaluation metrics are modified to the most common evaluation metrics in recommendation tasks, namely precision (P), recall (R), and AUC. In addition, this scheme adds personalized metrics for recommendation results and enhances the similarity evaluation of recommendation results.

[0065] Step 2: Obtain the parameters and feature data during the model training process; construct a heterogeneous information movie network using the pre-processed movie data as the input of the model, train it through three representative heterogeneous embedding networks HetGNN, KGAT and NIRec, retain the relevant parameters and feature information during the training process, and design a data visualization view to display the information;

[0066] 2.1 Training the NIRec model to obtain parameter and feature data

[0067] NIRec designed a neighborhood-based interaction model to enhance object representation. Building upon the recommender system, it utilizes an ensemble approach to guide the selection of neighbors at different steps and of different types. A heterogeneous interaction module is designed to acquire rich interaction information, and a heterogeneous aggregation module acquires rich object embeddings. Therefore, we retain the path instance information enumerated by the model when calculating path attention weight parameters and use it as visualization data.

[0068] 2.2 Training the KGAT model to obtain parameter and feature data

[0069] The KGAT model addresses the modeling of higher-order relations through recursive embedding propagation. It updates node embeddings based on neighbor embeddings, capturing higher-order connectivity in linear time complexity. It learns the weights of each neighbor during propagation using an attention-based aggregation method. Therefore, we retain the bidirectional attention weight parameters and local neighbor features from the model's training process as visualization data.

[0070] 2.3 Training the HetGNN model to obtain parameters and feature data

[0071] The HetGNN model effectively captures the heterogeneity of structure and content, considering both structural and node attribute information of heterogeneous networks, and separates the learning process from downstream tasks to meet the downstream tasks of various HINs. Therefore, we preserved the path instances involved in the learning process during the embedding vector initialization of this model as visualization data for model comparison.

[0072] The above three models are used to learn from the original data. The node entity paths used by the NIREC model during the learning process are retained as path fusion features; the attention weights of each node pair in the KGAT model during the learning process are retained, and entity paths are constructed through network connection relationships as fusion features; the entity path information sampled by the HetGNN model during the learning process is retained as path fusion features.

[0073] Step 3: Exploratory design based on model comparison using user interaction;

[0074] Based on the path feature data retained in the above steps, this invention will design corresponding visualization charts according to different indicator data and model parameters. The specific steps are as follows:

[0075] 3.1 The data selection and detailed overview entry points for the visualization design are as follows: Based on the requirements of model researchers for an overview of model evaluation metrics, this invention uses a circular stacked bar chart to display the evaluation metrics of heterogeneous embedded models for recommendation tasks. Each bar represents the recommendation evaluation result of a specific user, and each color represents a model. To facilitate data filtering, more interactive methods need to be designed, such as slider range selection and selection brush box selection of specific targets. In addition, a Venn diagram showing the relationship between recommendation results of different models is designed. The corresponding results are randomly scattered in the color areas corresponding to the models mentioned above, and the scatter points that match the user's viewing history are marked with a white border, indicating that they belong to the correctly predicted movies.

[0076] 3.2 Visualizing and Comparing Model Details: Based on the design and model developers' needs in step 3.1 to view detailed information about the embedded model during the neighbor aggregation process, a scalable force guidance graph was designed to display the meta-paths between the selected target and its recommended results. These paths generally include UMUM, UMGM, UMAM, and UMDM. The path width represents the importance of the path to the target node, and the node radius indicates the higher the importance as it contains more entity nodes.

[0077] Step 4: Interactive visual recommendation based on heterogeneous embedding models; This invention allows developers of heterogeneous embedding models to intuitively compare the differences in user features learned through different methods in practical applications, thereby solving the black-box problem of heterogeneous embedding models in recommendation tasks and improving the effectiveness of recommendations.

[0078] The heterogeneous embedding methods NIRec and HetGNN based on meta-paths for recommendation systems involve three steps: First, experts design effective meta-paths based on domain knowledge; second, the heterogeneous embedding model uses neurons to learn user neighborhood features through the aforementioned meta-paths, obtaining embedding vectors for both users and items to mine user preferences in the user-item interaction graph; third, the similarity between the user's embedding vector and all item embedding vectors is calculated, excluding training data, and the top K items with the highest similarity are used as the recommendation results.

[0079] The heterogeneous embedding method KGAT based on the information transmission mechanism has the following specific steps: 1. Obtain the embedding vector of each node in the heterogeneous graph using the TransR method; 2. Integrate the features of the neighbors around the target node into the node's embedding vector through the propagation mechanism, and use the attention mechanism to calculate the importance of the neighboring nodes; 3. After iterating through higher-order propagation, multiply the higher-order embedding vectors of the user and the item to obtain the final probability of the recommendation, and select the top K items with the highest recommendation probability as the recommendations.

[0080] Finally, the recommendation result source view displays the path information between the recommended movie and the target user for different models. Meta-path-based methods provide interpretability through meta-paths; for example, the path UMUM mines the preference features of similar users. Propagation mechanism-based methods provide interpretability through attention scores in neighbor relationships.

[0081] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. An interactive visual recommendation method based on a heterogeneous network information embedding model, comprising the following steps: Step 1: Data crawling and cleaning; collect Douban user data and Douban movie data from the real Douban movie website and clean them to ensure the validity of the data; (1.1) Obtain the original data set, randomly select a Douban user, and randomly retain according to the percentage of the number of movies watched, and randomly select three users from the list of users they follow, repeat the above process; then, count all users' viewing records, retain valid movie numbers, and crawl movie information; (1.2) Data preprocessing for the original data set, to make the evaluation of the model more fair and effective, that is, to prevent invalid data, filter the data under the condition that the user's viewing record is at least > 50 and the number of times the movie is watched by valid users is > 20; Due to the cold start problem of the recommendation model, 10% of the users with <50 viewing records are reserved as cold start data for special situation evaluation; (1.3) Fair evaluation of the model, in order to make the comparison of the model effective and fair, the downstream target of the embedding model is unified as the recommendation task, and the evaluation indicators are reconstructed as the most common evaluation indicators in the recommendation task, namely accuracy, recall rate and AUC, the calculation method is as follows: （1） （2） （3） Where, TP true negatives, TN false positives, FP false negatives, FN, M number of positive samples, N number of negative samples, rank i predicatescorei represents the predicted score for positive sample i; In addition, the individualization index of the recommendation result is added, and the similarity evaluation of the recommendation result is added, which calculates the cosine similarity between the recommendation lists of different users; the individualization score of user i is calculated as shown in formula (4): （4） where U denotes the set of users, represents a vector of some recommendation list for user i; Step 2: Obtain the parameters and feature data in the model training process; Construct a heterogeneous movie information network as the input of the model using the above step-preprocessed movie data, then use three representative heterogeneous network embedding models HetGNN, KGAT and NIRec for training, and retain the relevant parameters and feature information in the training process; The processing and data saving of the model includes the following steps: (2.1) Obtain the meta-path feature of the NIRec model; In order to explore the differences in neighbor features learned by the heterogeneous network embedding model based on meta-path neighbors in the use of end-to-end interaction modeling method, the path instance information enumerated by the heterogeneous network embedding model when calculating the path attention weight parameter is retained as visualization data; The path instance information is divided into two parts, one is the node feature information based on the instance path The other is the path feature information based on the instance path, and the calculation methods are shown as follows. （5） (6) denotes the instance path information based on paths in the interaction matrix W, a, b, b are trainable parameters, denotes the aggregated information based on meta-path guided neighborhood, single meta-path and semantics; The above path information is the feature path of all positive and negative samples, in order to ensure the fairness of the model comparison, the final recommendation result, i.e. the positive sample, is used to filter the path, only the instance path information of the source node reachable to the target node is retained as the visualization data; (2.2) Obtain the attention weight feature of the KGAT model; In order to explore the difference between the neighbor features learned by the non-customized soft path constituted by the heterogeneous network embedding model based on the information transmission mechanism combined with the attention mechanism, the bidirectional attention weight parameter in the training process of the heterogeneous network embedding model is reserved , which represents the importance of the neighbor node to the source node, and the specific calculation formula is as follows; （7） where h denotes the source node, t denotes the tail node, r denotes the relationship from the tail node to the source node, N h denotes the set of all neighbor nodes around the source node; Then, the average attention value between different types of nodes is calculated according to the obtained data, and the value is taken as the neighbor node screening threshold for constructing the meta path. The neighbor nodes higher than the threshold are regarded as important nodes, and vice versa. Finally, the important neighbor nodes of each source node are traversed by using the depth-first algorithm to generate the meta path reaching the target node with a length less than 4. The path contains the feature information from the source node to the target node, and is taken as the visualization data for model comparison. (2.3) Obtain the meta path feature of the HetGNN model; In order to explore the feature difference learned by the meta path-based heterogeneous network embedding model in the neighbor node aggregation process, the path instances participating in the model learning process are reserved as the visualization data for model comparison. In the initialization process of the heterogeneous network embedding model, when generating the initial embedding of the nodes by using the random walk algorithm with restart, the nodes are sampled in proportion according to the proportion of each type of node in the total number of nodes, and the movie neighbors, user neighbors, type neighbors, actor neighbors and director neighbors of each node are sampled. The instance path formed by the above neighbors is calculated by the following formula ; （8） Through the above three models, the original data is learned. The node entity path used by the NIRec model in the learning process is reserved as the path fusion feature; the attention weight of each node pair used by the KGAT model in the learning process is reserved, and the entity path is constructed as the fusion feature through the network connection relationship; the entity path information sampled by the HetGNN model in the learning process is reserved as the path fusion feature. Step 3: Model comparison and exploration design based on user interaction; Based on the path feature data reserved in the above steps, according to different index data and model parameters, corresponding visualization charts are designed. The specific steps are as follows: (3.1) Data selection entry and detail overview entry of visualization design. According to the index overview requirements of the model researchers for model evaluation, a ring-shaped stacked column chart is used to display the evaluation indexes of the heterogeneous network embedding model for the recommendation task. Each bar chart represents the recommendation evaluation result of a certain user, and each color represents a model. In order to facilitate data filtering, more interaction methods need to be designed, such as selecting a specified target by brushing. In addition, a Wean chart is designed to show the relationship between the recommendation results of different models. The corresponding results are randomly scattered in the color area corresponding to the above model, and the scatter results hitting the user's viewing history are attached with a white border, indicating that they belong to the correctly predicted movies. (3.2) Model details of visualization comparison design. Based on the design of step 3.1 and the demand of model developers for viewing the details of the embedding model in the aggregation process of neighbors, a scalable force guide chart is designed to show the meta path between the selected target and its recommendation result. The path includes UMUM (user-movie-user-movie), UMGM (user-movie-genre-movie), UMAM (user-movie-actor-movie), and UMDM (user-movie-director-movie). The width of the path represents the importance of the path to the target node. The larger the node radius, the more entity nodes it contains, and the higher the importance. Step 4: Interactive visual recommendation based on heterogeneous network embedding model Heterogeneous network embedding model aims to learn multiple types of objects and relationships in heterogeneous graph, and mine the implicit rich structure and semantic information for recommendation task; The meta-path-based heterogeneous embedding methods NIRec and HetGNN for recommendation system need three steps, first, experts design effective meta-paths according to domain knowledge; second, the heterogeneous network embedding model uses neurons to learn user neighborhood features through the above meta-paths, and obtains user and item embedding vectors respectively, in order to mine user preferences in user-item interaction graph; third, similarity calculation is performed between user embedding vector and all item embedding vectors, excluding training data, and the top K items with the highest similarity are recommended as the recommended results; The heterogeneous embedding method KGAT based on information propagation mechanism uses the natural connection properties of heterogeneous graph to learn the features of the tail nodes connected to the target node through the edge, the specific steps are as follows: first, use the common TransR method in knowledge graph field to obtain the embedding vector of each node in the heterogeneous graph; second, the neighbor features around the target node are fused into the embedding vector of the node through the propagation mechanism, and the attention mechanism is used to calculate the importance of each neighbor node, so as to distinguish the fusion degree of each feature; third, after iteration of high-order propagation, the high-order embedding vectors of users and items are multiplied to obtain the final probability of recommendation, and the top K items with the highest recommendation probability are recommended; Finally, in the recommendation result traceability view, the path information between the recommended movies and the target user is displayed for different models; the meta-path-based method provides the explainability of recommendation through the meta-path, the embedding model finds out what movies the same person has watched as the target user through this path, and uses it as recommendation, and mines the preference features of similar users; the method based on propagation mechanism provides the explainability of recommendation through the attention score in the neighbor relationship, and the final recommended movie contains this type of feature, indicating that the recommendation of the heterogeneous network embedding model meets the user's preference.

2. The method of claim 1, wherein the method is based on a heterogeneous network information embedding model. In step 1, 20,000 pieces of Douban user data and 20,000 pieces of Douban movie data were collected from the real Douban movie website.

3. The interactive visual recommendation method based on a heterogeneous network information embedding model as described in claim 1, characterized in that: In step 2.3, the number of movie neighbors, user neighbors, type neighbors, actor neighbors, and director neighbors for each node was sampled as 25, 25, 25, 15, and 10 respectively.

4. The interactive visual recommendation method based on a heterogeneous network information embedding model as described in claim 1, characterized in that: The interaction method described in step 3.1 includes sliding bar interval selection.

Images