A heterogeneous information network recommendation method of automatically extracting meta paths

Abstract

The application provides a heterogeneous information network recommendation method for automatically extracting a meta path, and comprises the following steps: S1, automatically extracting a meta path: firstly, mining path instances with different semantics between entities, and fully utilizing the entity relationship in the HIN; then, classifying and arranging all path instances according to the entity types passed by the path, so as to obtain the automatically mined meta path; S2, inputting the meta path into a hierarchical attention network, and obtaining a project recommendation list of a user. The application automatically mines meaningful path instances from the interaction scene based on the heterogeneous information network, the path instances can represent the recommendation process and the recommendation reason; in addition, the preferences of the user to each path instance are considered, then the context based on the meta path is incorporated into the heterogeneous information network recommendation model of the three-party interaction modeling of the interaction, so that the most suitable information for the user is matched.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of heterogeneous information network recommendation, and particularly to a method for automatically extracting meta-path-based heterogeneous information network recommendation. BACKGROUND

[0002] With the rapid development of the Internet era, a large amount of information is constantly emerging, which greatly facilitates the communication and information selection of users. However, it also brings the problem of information overload, which greatly affects the efficiency of people obtaining useful information. The recommendation system aims to mine the potential interests of users and match the most suitable information for users, so as to solve the problem of information overload and mine long-tail demand and improve user experience. After decades of development, the recommendation system has been successfully applied to e-commerce, multimedia and other service scenarios.

[0003] Heterogeneous information network (HIN) is a general method of fusing multi-source data. The recommendation system based on heterogeneous information network has significant advantages in information fusion and exploring structural semantics. It not only can effectively alleviate the problem of data sparsity and cold start, but also helps to improve the accuracy and interpretability of the recommendation system, so it has attracted wide attention and application.

[0004] Most of the existing recommendation methods based on heterogeneous information network use meta-path to obtain rich user and item auxiliary information from heterogeneous information network. For example, PathSim and its variant HeteSim are similarity measurement methods based on meta-path. Pathsim can find peer objects in the network (for example, find similar fields and authors with similar reputation), which is more meaningful in many cases than similarity measurement based on random walk. Because meta-path introduces more background information, and the designed meta-path satisfies good properties such as symmetry. These advantages enable meta-path to integrate more semantic information, but meta-path needs to be designed manually. But these two methods are still the cornerstone of subsequent work such as matrix decomposition. Benefiting from the advantages of high readability of meta-path and different weights of different meta-paths, the method based on meta-path also has the advantage of solid interpretability.

[0005] Although there are now a lot of work using meta-path, and to a certain extent, improve the recommendation performance, but so far there are still several challenges to be solved. First, most of the existing path-based methods are designed by professionals manually meta-path mode and then find the path instance according to the mode. Such methods are heavily dependent on artificial input a series of task-related meta-path, which is difficult for users without professional knowledge. More importantly, the meta-path designed by artificial often can not cover all possible path instances, which greatly hinders the improvement of recommendation quality. In addition, the existing recommendation algorithm based on heterogeneous information network, all in as much as possible using multi-element information for recommendation, but rarely in the recommendation algorithm to learn the explicit representation of path or meta-path, can not intuitively represent the recommendation process. In addition, although MCRec learns the context explicit representation of meta-path, it does not consider the user's preference for each path instance when generating the meta-path context embedding. SUMMARY

[0006] The present application aims to at least solve the technical problems existing in the prior art, and particularly innovatively provides a heterogeneous information network recommendation method for automatically extracting meta-path.

[0007] In order to achieve the above purpose of the present application, the present application provides a heterogeneous information network recommendation method for automatically extracting meta-path, comprising the following steps:

[0008] S1, automatically extracting meta-path: first, mining the path instances between entities with different semantics, making full use of the entity relationship in HIN. Then, according to the entity type passed by the path, all path instances are classified and arranged, so as to obtain the automatically mined meta-path;

[0009] S2, inputting the meta-path into the hierarchical attention network to obtain the user's item recommendation list.

[0010] Further, the prominent path is selected from the meta-path extracted in S1:

[0011] The path length is less than the set threshold; there may be several paths between each pair of entities (only considering the user and the project), which have different lengths, different orders and different semantic information. Among these paths, the relatively shorter path is sufficient to model the entity relationship, while the longer path may introduce remote neighbors and blur the semantic meaning, thereby introducing a lot of noise.

[0012] and / or according to the priority, the path with high priority is reserved. After the path instances are classified, there may be several paths under each meta-path, and the quality of these paths may be different. Reserving all the paths will introduce unnecessary semantic information, and randomly selecting a path may reserve a path with low quality. Calculating the priority of the path can effectively reduce the impact of noise on the recommendation.

[0013] Further, the pre-trained model is used to learn the latent vector of each node with user-item interaction record, and then the cosine similarity of adjacent nodes in the path instance is calculated, and the average value is taken as the priority of the path. Such priority score directly reflects the degree of association between two nodes.

[0014] Further, when calculating the cosine similarity of adjacent nodes in the path instance, for each path instance, only the similarity of the first few pairs of nodes is calculated, and the similarity of the last pair of nodes is not calculated. Because after the order of the previous nodes is determined, the order of the subsequent nodes is also determined accordingly, that is, it can be considered that in a path instance, the previous nodes play a decisive role.

[0015] Further, the hierarchical attention network comprises a node aggregation layer, a path instance attention layer, a meta-path attention layer, and a prediction layer,

[0016] The node aggregation layer is composed of several convolutional neural networks, and the path instance attention layer and the meta-path attention layer both use a fully connected network to realize the attention mechanism.

[0017] Further, the S2 comprises:

[0018] S2-1, obtain the embedding of the path instance in the node aggregation layer, and the embedding of the path instance p under the meta-path φ is defined as follows:

[0019]

[0020] wherein represents the embedding matrix of the path instance p;

[0021] Conv1D represents convolution operation;

[0022] Θ is all related parameters in Conv1D;

[0023] S2-2, adopt the path instance attention layer to automatically learn the importance of different path instances:

[0024] The embedding of K path instances under the meta-path φ is represented as The importance of each path instance to the meta-path to which it belongs is learned through a fully connected network, then normalized, and then the weight coefficient is obtained through the softmax function

[0025]

[0026] This represents the weight coefficient of the path instance p under the metapath φ;

[0027] Among them W p and W k These are the weight matrices for path instance p and path instance k, respectively.

[0028] The embedding of the path instance p of the metapath φ;

[0029] b p and b k These are the bias vectors of path instance p and path instance k, respectively.

[0030] σ(W) is the ReLU function;

[0031] K represents the number of path instances under the metapath φ;

[0032] Therefore, the embedding representation of the metapath φ is c φ It is obtained by aggregating the path instances it contains with their corresponding coefficients:

[0033]

[0034] S2-3, through the meta-path attention layer, uses user-item pair information from the interaction as auxiliary information to automatically learn the importance of different meta-paths and accordingly aggregates the embedding representations of the meta-paths in S2-2 to learn context-based meta-path representations:

[0035] Embedding of user, item, and metapath φ obtained from the path instance attention layer. φ As input, a fully connected layer is used to transform semantically specific embeddings, and the importance of each meta-path is represented as follows:

[0036] s u,i,φ =σ(W u ·e u +W i ·e i +W φ ·c φ +b) (4)

[0037] s u,i,φ Indicate the importance of the metapath φ;

[0038] W u W i W φ eu , e i , c φ weight matrix of the meta-path;

[0039] e u and e i are the low-dimensional embeddings of users and items (products) respectively;

[0040] c φ denotes the embedding representation of the meta-path φ;

[0041] b is a bias vector;

[0042] σ() is a ReLU function;

[0043] The importance of all the meta-paths above is then normalized by using a softmax function to obtain the weight representation of the meta-path α u,i,φ :

[0044]

[0045] s u,i,φ denotes the importance of each meta-path φ based on the node pair (u, i);

[0046] s u,i,m denotes the importance of each meta-path m based on the node pair (u, i);

[0047] The node pair (u, i) is just a user-item pair, and u and i represent the user and the item respectively;

[0048] m represents a certain meta-path in the meta-path set;

[0049] M u→i denotes the meta-path set;

[0050] The weight α u,i,φ of the meta-path can explain the contribution of the meta-path to the interaction between u and i. It is obvious that the higher the α u,i,φ , the more important the meta-path.

[0051] The weight α u,i,φ of the meta-path is taken as a coefficient to fuse the embedding c φ of the meta-path φ to obtain the final meta-path context embedding, as follows:

[0052]

[0053] where c u→i denotes the meta-path-based context embedding;

[0054] φ denotes the meta-path φ;

[0055] c φembedding of a metapath φ;

[0056] In summary, the metapath attention layer establishes a personalized preference model for metapaths by introducing a preference vector for each metapath, and the learned metapath context embedding can be specific to interactions and capture different interaction contexts.

[0057] S2-4, the one-hot representation of the user and the item is converted into a low-dimensional dense vector through the lookup layer in the prediction layer, and then a triplet formed by the user, the metapath context and the item is modeled to obtain the embedding y of the three-party interaction u,i :

[0058] y u,i = (e u ⊙e t )⊙σ(c u→i ) (7)

[0059] where e u , e i and c u→i are the low-dimensional embedding of the user, the low-dimensional embedding of the item and the context embedding based on the metapath, respectively;

[0060] ⊙ is a pair-wise multiplication operator (also known as Hadamard product operator);

[0061] σ(·) is a sigmoid activation function;

[0062] Then the embedding of the three-party interaction is sent into a two-layer multi-layer perception (MLP) to generate the probability of recommending the item i to the user u:

[0063]

[0064] The probability of recommending the item i to the user u;

[0065] σ1(·), σ2(·) are activation functions;

[0066] W1, W2 are weight matrices;

[0067] b1, b2 are bias vectors;

[0068] y u,i is the embedding representation of the interaction.

[0069] In summary, due to the adoption of the above technical solutions, the present application has the following advantages:

[0070] (1) By designing the method of automatically mining metapaths, meaningful path instances are automatically mined from the interaction scenarios in the heterogeneous information network, and the path instances can represent the recommendation process and the recommendation reason.

[0071] (2) The meta-path based context is learned by a hierarchical attention network, considering the user's preference for each path instance, and then incorporated into a heterogeneous information network recommendation model for three-party interaction modeling of the interaction, so that the most suitable information is matched for the user.

[0072] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0073] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the description of the embodiments, given by reference to the following drawings:

[0074] Figure 1 is the overall architecture of the MAERec model of the present application. DETAILED DESCRIPTION

[0075] Embodiments of the present application are described in detail below with reference to examples illustrated in the accompanying drawings, in which the same or similar components have the same or similar designations throughout. The embodiments described below are examples of the present application, and are intended to explain the present application, and should not be understood as limiting the present application.

[0076] 1 Problem definition related to the present application

[0077] Definition 1. Heterogeneous information network. An information network can be defined as a directed graph G(V, E) consisting of a set of entities (nodes) V and a set of relationships (edges) E. An information network is also associated with an entity type mapping function and a relationship type function, where and R represent entity types and relationship types, respectively. If this graph G(V, E) is called a heterogeneous information network (HIN).

[0078] Definition 2. Meta-path. A meta-path φ is essentially defined as a path of the form (abridged as ) that describes a composite relationship R1°R2°…°R l+1 between entities and where ° denotes a composition operator over relationships.

[0079] Definition 3. Meta-path based context. Given a user u and an item (object) i, the meta-path based context is described as an aggregated set of path instances under the meta-path connecting the two nodes on the HIN.

[0080] Definition 4. Top-N Heterogeneous Recommendation. Given the ratings between users u and items i as y u

[0081]

[0082] where (u, i) denotes a pair of user-item;

[0083] Ω denotes all user-item pairs on the HIN,

[0084] In addition, a subset of things involved in the HIN is denoted by where denotes the set of items, denotes all item entities on the network. The goal of Top-N Heterogeneous Information Network is to provide a ranked list of highly likely items for each user by leveraging the associated HIN.

[0085] Our model first leverages a hierarchy to obtain meta-path based contextual embeddings, the most important part of the model. In addition, we use the one-key ID of <user, item> as input, which is sent to the lookup layer to form the initial embedding (i.e., e(u), e(v)). Finally, in order to make more full use of the contextual information between <user, item, meta-path>, the model learns the embedding of three-party interaction by Hadamard product on the embedding of user, meta-path and item.

[0086] 2 The method of the application

[0087] We propose the MAERec model, which mainly includes a meta-path automatic extraction module and a hierarchical attention network module. The meta-path automatic extraction module automatically extracts high-quality meta-paths and path instances from the heterogeneous information network. The hierarchical attention network module learns the explicit representation of the meta-path based context using a hierarchical attention network for recommendation.

[0088] The overall architecture of the MAERec model is shown in Figure 1 , which includes: node aggregation layer, path instance attention layer, meta-path attention layer, prediction layer. The node aggregation layer is composed of several convolutional neural networks, and the path instance attention layer and the meta-path attention layer both use a layer of fully connected network to realize the attention mechanism.

[0089] ​The MAERec model first utilizes a hierarchy to obtain meta-path based context-aware embeddings, the most important part of the model. In addition, we utilize the one-key ID of <user, item> as input, which is sent to the lookup layer to form the initial embedding (i.e., e(u), e(v)). Finally, in order to make more full use of the context information between <user, item, meta-path>, the model learns the embedding of three-party interaction by Hadamard product on the embedding of user, meta-path and item.

[0090] 2.1 Meta-path automatic extraction module

[0091] Most of the existing HIN-based recommendation models utilize meta-paths to obtain rich user and item auxiliary information from HIN. However, the meta-path based method largely relies on hand-crafted features to represent the path semantics, which further relies on domain knowledge. More importantly, the hand-designed features are often incomplete and cannot cover all possible entity relationships, thus hindering the improvement of recommendation quality. To solve this problem, we propose a method that does not require professionals to manually design meta-paths. We first mine path instances with different semantics between entities, making full use of the entity relationships in HIN. Then, we classify all path instances according to the entity types passed by the path, thus obtaining automatically mined meta-paths. Due to the huge number and complexity of HIN, there are a large number of paths connecting entity pairs, and each entity pair may contain various entity types and relationship types in different orders and lengths. In order to improve the efficiency of the model, we therefore designed two strategies to help select prominent paths.

[0092] Strategy 1 Length. We constrain the length of the path, i.e., only select paths with length less than a certain threshold. There may be several paths between each pair of entities (only considering users and items), with different lengths, orders, and semantic information. Among these paths, the relatively shorter paths are sufficient to model the entity relationship, while the longer paths may introduce distant neighbors, blur the semantic meaning, and thus introduce a lot of noise.

[0093] Strategy 2 priority, we filter the path according to the priority, that is, only keep the path with higher priority. After sorting the path instances, there may be several paths under each meta-path, and their quality may vary. Keeping all paths will introduce unnecessary semantic information, and randomly selecting paths may keep low-quality paths. Calculating the priority of the path can effectively reduce the impact of noise on the recommendation. The basic idea of priority calculation is to use the pre-trained model to learn the latent vector of each node with user-item interaction records, and then calculate the cosine similarity of adjacent nodes in the path instance, and take the average value as the priority of the path. Such priority score directly reflects the degree of association between two nodes. In particular, for each path instance, we only calculate the similarity of the first few pairs of nodes, and do not calculate the similarity of the last pair of nodes. Because we have reason to believe that after the order of the previous nodes is determined, the order of the subsequent nodes is also determined, that is, we can believe that in a path instance, the previous nodes play a decisive role.

[0094] 2.2 Hierarchical attention network module

[0095] 1) Node aggregation into layers

[0096] In fact, the path instance between the user and the item is a sequence composed of entity nodes. The node aggregation layer is to obtain the low-dimensional embedding of the node sequence. The basic idea is to process the variable-length node sequence through the convolutional neural network to extract features. The embedding of the path instance p under the meta-path φ is defined as follows:

[0097]

[0098] where represents the embedding matrix of the path instance p with dimension d formed by L node embeddings. Conv1D represents the convolution operation, and Θ is all related parameters in Conv1D.

[0099] 2) Path instance attention layer

[0100] Since there may be multiple path instances under each meta-path, and these path instances imply different semantics, users may also show different preferences for these path instances. In order to better learn the embedding representation of the meta-path, we propose path instance-level attention to automatically learn the importance of different path instances. Given a meta-path φ, let represent the embedding of the K path instances in it. Then we use a fully connected network to represent the importance of different path instances, and after obtaining the importance of each path instance to the meta-path it belongs to, we normalize it to get the weight coefficient through the softmax function:

[0101]

[0102] where W p and W k are weight matrices;

[0103] denotes the weight coefficient of path instance p under meta-path φ;

[0104] denotes the embedding of path instance p from meta-path φ;

[0105] b p and b k are bias vectors;

[0106] σ(·) is the ReLU function;

[0107] K is the number of path instances under meta-path φ.

[0108] The embedding of meta-path φ can then be obtained by aggregating its contained path instances with the corresponding coefficients as follows:

[0109]

[0110] c φ denotes the embedding of meta-path φ;

[0111] 3) Meta-path attention layer

[0112] Generally, there are more than one meta-path in the interaction between users and items, and each meta-path contains different and diverse semantic information. In order to have a more comprehensive understanding of the interaction between users and items, we need to fuse the representations of different types of meta-paths between them, and model the context-based meta-path. In addition, different types of meta-paths may have different semantics in the interaction, and the users and items involved in them will also affect the context of the interaction. For this purpose, we propose a meta-path level attention to automatically learn the importance of different meta-paths with the information of user-item pairs in the interaction as auxiliary information, and accordingly aggregate the meta-path representation of the last step to learn the context-based meta-path representation.

[0113] To learn the importance of each meta-path, we take the user embedding, item embedding, and the embedding c φ of meta-path φ obtained from the path instance attention layer as input, and use a fully connected network to transform the semantics-specific embedding. The importance of each meta-path can be represented as follows:

[0114] s u,i,φ = σ(W u · e u + W i · e i + Wφ ·c φ +b) (4)

[0115] where W u , W i , W φ are weight matrices of e u , e i , c φ respectively;

[0116] b is a bias vector;

[0117] σ(·) is a ReLU function;

[0118] e u and e i are the low-dimensional embedding of users and items respectively; that is, the low-dimensional embedding of users and items after the lookup layer transformation.

[0119] After obtaining the importance of each meta-path, the weight of the meta-path is represented as a u,i,φ , which can be obtained by normalizing the importance of all the above meta-paths using the softmax function.

[0120]

[0121] s u,i,φ represents the importance of each meta-path φ based on the node pair (u, i);

[0122] s u,i,m represents the importance of each meta-path m based on the node pair (u, i);

[0123] The node pair (u, i) is just a user-item pair, and u and i represent the user and the item respectively;

[0124] m represents a certain meta-path in the meta-path set;

[0125] Mu→i represents the meta-path set;

[0126] It can be interpreted as the contribution of the meta-path to the interaction between u and i. Obviously, the higher a u,i,φ , the more important the meta-path. Using the learned weights as coefficients, we can fuse these specific semantic embeddings to obtain the final meta-path context embedding, as follows:

[0127]

[0128] where c u→i represents the meta-path-based context embedding;

[0129] φ represents the meta-path φ;

[0130] c φ Embedding of a metapath φ.

[0131] In summary, the metapath attention layer builds a personalized preference model for metapaths by introducing a preference vector for each metapath, and the learned metapath context embedding can be specific to interactions and capture different interaction contexts.

[0132] 4) Prediction layer

[0133] Through the above steps, we obtain the final metapath context embedding between a given user and an item. In addition, we convert the one-hot representation of users and items into low-dimensional dense vectors by setting the lookup layer. In order to model the triplets formed by users, metapath contexts and items, we generate an embedding for the current interaction by performing a Hardamard product inspired by HIN2Vec, as follows:

[0134] y u,i = (e u ⊙ e t ) ⊙ σ(c u→i ) (7)

[0135] where e u , e i and c u→i are the low-dimensional embedding of users, the low-dimensional embedding of items and the metapath-based context embedding, respectively;

[0136] “⊙” is a pairwise multiplication operator (also known as Hadamard product operator);

[0137] σ(·) is a sigmoid activation function.

[0138] Then, we feed the embedding of the triple interaction into a two-layer MLP to generate the probability of recommending item i for user u

[0139]

[0140] the probability of recommending item i for user u;

[0141] σ1(·), σ2(·) are activation functions;

[0142] W1, W2 are weight matrices;

[0143] b1, b2 are bias vectors;

[0144] y u,i embedding representation of the interaction;

[0145] In addition, we conduct comprehensive experiments on publicly accessible datasets to compare and evaluate the effectiveness of the proposed method.

[0146]

[0147] The experimental results of MAERec model and other benchmark models on the Movielens dataset can be seen from the above table. The accuracy and recall rate of MAERec on the Movielens dataset are better than those of other benchmark models, which shows the high quality of the meta-paths and path instances automatically extracted according to the strategy in the heterogeneous information network, and shows that the meta-path context embedding expression obtained based on the hierarchical attention network and the modeling of three-party interaction can improve the recommendation performance, indicating the effectiveness of the proposed model.

[0148] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A heterogeneous information network recommendation method for automatically extracting meta-paths, characterized in that, The method comprises the following steps: S1, automatically extracting meta-paths: firstly, mining path instances with different semantics between entities, and then classifying and arranging all path instances according to the entity types passed by the paths, so as to obtain meta-paths; S2, inputting the meta-paths into a hierarchical attention network to obtain a project recommendation list of the user, and the S2 comprises: S2-1, the embedding of the path instance in the node aggregation into a layer, the meta-path under the path instance The embedding of the path instance in the node aggregation into a layer, the meta-path is defined as follows: ， wherein represents a path instance of the embedding matrix; denotes a convolution operation; Θ is all relevant parameters in S2-2, using a path instance attention layer to automatically learn the importance of different path instances: Meta-path Common Embedding of individual path instances is denoted as The importance of each path instance to its meta-path is learned by a fully connected network, and then normalized to obtain the weight coefficient by the softmax function : ， representation meta-path under the path instance weight coefficient; wherein and are the weight matrix of the path instance and the weight matrix of the path instance respectively; representation meta-path path instance of the meta-path embedding of the meta-path and are a bias vector of the path instance and a bias vector of the path instance respectively σ(·) is a ReLU function; path instances under the meta-path path instances Thus, the embedding of a meta-path is represented by the aggregation of the instances of the paths it contains with the corresponding coefficients. ， S2-3, through the meta-path attention layer, using the information of the user-item pair in the interaction as auxiliary information, automatically learning the importance of different meta-paths, and correspondingly aggregating the embedding representation of the meta-paths in S2-2 to learn the context-based meta-path representation: user embeddings, item embeddings, and meta-path embeddings obtained from the path instance attention layers For the input, a fully connected network is used to transform the semantic-specific embeddings, the importance of each meta-path is represented as follows:​ ， representing the importance of the meta-path ​ respectively, are weight matrices; respectively, are weight matrices; and are low-dimensional embeddings of users and item items, respectively; representing a meta-path of embedded representations; is a bias vector; σ(·) is a ReLU function; The importance of all the meta-paths above is then normalized by using a softmax function to obtain the weight representation of the meta-paths : ， representing the importance of each meta-path based on node pairs pairs​ representing the importance of each meta-path based on node pairs pairs​ node pair i.e., user-item pair, respectively, user and item; represents a certain meta-path in the meta-path set; represents a set of meta-paths; The weight of the meta-path is calculated as follows: As coefficients, the embeddings of the meta-paths are fused The final meta-path context embedding is obtained as follows:​ ， wherein denotes a meta-path based context embedding; representation meta-path ; representation of a meta-path embedding of a meta-path; S2-4, converts one-hot representation of users and items into low-dimensional dense vectors by lookup layer in the prediction layer, and then models triplets formed by users, meta-path context and items to obtain embedding of three-party interaction : ， wherein, , and are a low-dimensional embedding of users, a low-dimensional embedding of items, and a meta-path-based context embedding, respectively; ⊙ is a pair-wise multiplication operator; σ(·) is a sigmoid activation function; The three-way interaction embedding is then fed into a two-layer multi-layer perception, MLP, to generate a probability that the user will recommend the item . ， user recommended item probability; (⋅), (⋅) is an activation function; is a weight matrix; is a bias vector; Embedded representation of the interaction. 2.The heterogeneous information network recommendation method of automatically extracting meta-path according to claim 1, wherein, Further comprising selecting prominent paths from the meta-paths extracted in S1: Selecting paths with a length less than a set threshold value; And / or screening paths according to priorities, and retaining paths with high priorities. 3.The method of claim 2, wherein, The calculation of the priority comprises: using a pre-trained model to learn a latent vector for each node with a user-item interaction record, then calculating the cosine similarity of adjacent nodes in a path instance, and taking the average value as the priority of the path. 4.The method of claim 3, wherein, When calculating the cosine similarity of adjacent nodes in a path instance, for each path instance, the similarity of the last pair of nodes is not calculated.

5. The method of claim 1, wherein, The hierarchical attention network comprises: a node aggregation layer, a path instance attention layer, a meta-path attention layer and a prediction layer, The node aggregation layer is composed of several convolutional neural networks, and the path instance attention layer and the meta-path attention layer both use a fully connected network to realize the attention mechanism.

Images