CN117556148A - Personalized cross-domain recommendation method based on network data driving - Google Patents

Abstract

The invention discloses a personalized cross-domain recommendation method based on network data driving, which comprises the following steps: constructing user-object heterograms of a source domain and a target domain respectively; negative sampling of user-object related edges is carried out on the user-object heterogeneous graphs of the source domain and the target domain, a training set and a testing set are divided according to common users, and a meta-path is set; performing representation learning on nodes in the user-object heterograms through a TAHIN model to respectively obtain user node embedding and object node embedding; embedding source domain user nodes and target domain object nodes into an input meta-network to generate personalized bridging functions; the heterogeneous graph neural network model of the source domain and the target domain, and the meta-network are trained. According to the method and the system for recommending the commodity in the target domain, the data sparseness problem and the cold start problem can be effectively relieved during the cross-domain recommendation, more comprehensive and accurate commodity recommendation is provided for the user, and the recommendation of the user in the target domain is more personalized.

Description

Personalized cross-domain recommendation method based on network data driving

Technical Field

The invention belongs to the technical field of artificial intelligence, and particularly relates to a personalized cross-domain recommendation method based on network data driving.

Background

Recommendation systems play a key role in discovering user interests and in buffering information overload in the process of curating. In recent years, the application of deep neural architecture to recommendation systems has achieved great success. However, deep learning recommendation models essentially require learning a large number of parameters, which may be overfitted when their training data (i.e. user-project interactions) is insufficient, and which cannot be well generalized in practice. When a new user comes, the new user has no interactive information, and the deep learning recommendation model cannot recommend the new user. Therefore, the cold start problem and the data sparseness problem become main performance bottlenecks of the current deep learning recommendation model.

To solve the data sparseness problem, cross-domain recommendations (Cross-Domain Recommendation, CDR) were proposed. In the cross-domain recommendation, one domain is regarded as a source domain and the other domain is regarded as a target domain, and the single-target cross-domain recommendation can utilize a domain (source domain) with relatively rich information to improve the recommendation precision of other domains (target domains). For example, the bean paste may be recommended for bean paste books according to the user's film evaluation, because common users in different fields may have similar tastes; the amazon platform may make recommendations of amazon movies according to the type of music that the user is listening to. Cross-domain recommendation based on embedding and mapping is the most important part of cross-domain recommendation, and there are three main steps, namely potential embedding modeling, potential spatial mapping and cross-domain recommendation. In the latent embedding modeling process, the goal is to generate a latent embedding of users and items in each domain . The purpose of the potential spatial mapping is to train the mapping function +.>，/>To establish a relationship of potential space of the source domain and the target domain. In this approach, the most significant challenge is the potential embedded learning and mapping of potential space.

Most current methods do not take full advantage of ancillary information in cross-domain recommendations, such as item tag information, brand information, and classification information, which limits the modeling capabilities of item features. Such ancillary information may provide a model with richer and more accurate features, thereby improving recommendation performance. Modeling such ancillary information also presents challenges, and modeling methods of conventional recommendation systems often lack the ability to model such complex relationships. In recent years, graphic neural networks (Graph Neural Network, GNN) have been widely used in recommendation systems, because most of the information in recommendation systems is graphic in nature and GNN has advantages in terms of graphic representation learning. For example, interaction data in a recommendation application may be represented by a bipartite graph between users and project nodes, with observed interactions represented by edges. However, when there is heterogeneous assistance data, the bipartite graph also lacks modeling capabilities, and in order to handle heterogeneous information, many advanced heterogeneous graph neural network (Heterogeneous Graph Neural Network, HGNN) models have been proposed and proved to be effective. By constructing the heterograms and using the graph convolution network and the graph annotation network as powerful graph data deep representation learning methods, excellent performance is exhibited in terms of recommendation. The benefits of formulating recommendations as graph tasks become particularly apparent when structured auxiliary information is incorporated, such as social relationship knowledge graphs between users and knowledge graphs related to items.

However, most of the cross-domain recommendation methods based on the graph neural network at present form a unified composition of a source domain and a target domain. However, in cross-domain recommendations, the source and target domains often represent different markets or domains, and thus the source and target domains often have different rating patterns and item sets, which results in a problem of insufficient accuracy of user embedded learning.

In practice, the complex relationship between the user preferences of the source domain and the target domain varies from user to user due to individual differences. Thus, it is difficult for a single bridge to capture such complex and diverse relationships, which may degrade the performance of these CDR approaches.

Disclosure of Invention

The technical problems to be solved are as follows: aiming at the technical problem that complex relation between user preferences of a source domain and a target domain is difficult to capture in the prior art, the invention discloses a personalized cross-domain recommendation method based on network data driving, which can effectively relieve the data sparseness problem and the cold start problem during cross-domain recommendation; in addition, compared with the traditional recommendation method which only carries out recommendation according to the interaction information of the user and the article, the method has the advantages that the limitation of personal preference and the problem of information isolation exist, the auxiliary information of the commodity is added, and the similar relationship or the complementary relationship between the commodities can be fully utilized by adopting the heterogeneous graph model, so that more comprehensive and accurate commodity recommendation can be provided for the user; by using the meta-network as a bridging function, the recommendation of the user in the target domain can be more personalized.

The technical scheme is as follows:

a personalized cross-domain recommendation method based on network data driving, the personalized cross-domain recommendation method comprising the steps of:

step 1, data preprocessing is carried out on collected user data and object data of a source domain and a target domain, user-object heterograms of the source domain and the target domain are respectively constructed according to the preprocessed data, and a source domain heterogram and a target domain heterogram are generated;

step 2, carrying out negative sampling on the user-object related edges of the user-object heterogeneous graphs of the source domain and the target domain, dividing a training set and a testing set according to common users, and setting a meta-path;

step 3, performing representation learning on nodes in the user-object heterograms through the heterograms neural network model to respectively obtain user node embedding and object node embedding;

step 4, embedding source domain user nodes and target domain object nodes into an input meta-network to generate personalized bridging functions;

training the heterogeneous graph neural network model of the source domain by utilizing the user-article heterograph of the source domain and the scoring of the article by the source domain user, training the heterogeneous graph neural network model of the target domain by utilizing the user-article heterograph and the scoring of the target domain to remove the test set, training the meta-network by utilizing the common user, and storing the trained model;

And 6, inputting a source domain user and preset articles to obtain Top-K articles for recommendation.

Further, in step 1, the process of constructing the user-object iso-composition of the source domain and the target domain, and generating the source domain iso-composition and the target domain iso-composition a includes the following steps:

step 1.1, respectively acquiring user data and article data from a source domain system and a target domain system, wherein the article data comprises brands of articles, classification of the articles, historical browsing before browsing the articles, and scoring matrixes of users in the source domain and the target domain for the articles;

step 1.2, carrying out structuring treatment on the collected user data and article data to obtain structured data after treatment;

step 1.3, mapping the character strings in the processed structured data into numerical values, and storing the mapping relations as a user dictionary and an article dictionary;

and 1.4, respectively constructing a source domain iso-composition and a target domain iso-composition based on the structured data, the user dictionary and the object dictionary, and designing node types and names of the source domain iso-composition and the target domain iso-composition and types and names of the edges of the heterogeneous graphs.

Further, the node types of the user-item heterograms include a user node u, an item node i, a brand node b, a classification node c and a history browsing item node y;

Types of edges of the user-object iso-graph include: the sides ui and iu between the user and the item represent the relationship that the user u purchases the item i and the item i is purchased by the user u, the sides ib and bi between the item and the brand represent that the item i belongs to the brand b and issues the item i, the sides ic and ci between the item and the category represent that the item i is classified into the category c and contains the item i, and the sides iy and yi between the item and the historically browsed item represent the item y browsed before the item i is purchased and the item i that is purchased after the item y is browsed.

Further, in step 2, the process of performing negative sampling on the user-object of the source domain and the user-object heterogeneous map of the target domain with respect to the user-object of the edge, dividing the training set and the testing set according to the common user, and setting the meta-path includes the following steps:

step 2.1, representing the edges of the user-object of the source domain heterograms and the target domain heterograms in a triplet mode;

step 2.2, randomly constructing a plurality of negative sides, combining the positive sides and the negative sides, and completing negative sampling of the sides of the source domain iso-graph and the target domain iso-graph about the user-object; the sorting of the triples is disordered, and for common users of a source domain and a target domain, the triples after being disordered are divided into a training set and a testing set according to a preset proportion;

Step 2.3, deleting the edge containing the test set user from the target domain heterogeneous graph to obtain a processed target domain heterogeneous graph; storing the processed target domain heterogeneous graph, training set and test set;

step 2.4, respectively setting element paths of a source domain and a target domain required by heterogeneous graph neural network training; meta pathIs defined as +.>For describing node->And node->Compound relation betweenWherein->Representing the composite operator on the relationship.

Further, in step S3, the process of performing representation learning on the nodes in the user-object heterogram through the TAHIN model to obtain user node embedding and object node embedding respectively includes the following steps:

step 3.1, initializing a source domain isomerism mapAnd target Domain isomerism map->Embedding of different types of nodes, V is node set, < ->Is an edge set, different node types correspond to different dimensions, nodesxIs embedded as +.>The initialization of node w is embedded as +.>The initialization of node j is embedded as +.>；

Step 3.2, using projection matrixProjecting different types of neighbor nodes of the target node to the same dimension of the target node;

step 3.3, calculating the attention weight for the target node and other neighbors, respectively, and calculating the node xAnd attention weight between nodes wThe method comprises the following steps: />Wherein->Is a nodexIs a set of all one-hop neighbors +.>Is a deep neural network representing a relationship note between two nodes,is an exponential function;

step 3.4, according to the calculated attention weightPerforming one-hop neighbor aggregation of target nodes and nodesxOne-hop neighbor aggregation information of->Wherein->Is an activation function;

step 3.5, calculating the attention weight of the neighbor based on the meta-path, in whichOn the nodexAnd the attention weight between nodes w +.>，/>For the meta-path->Upper nodexIs>Is a deep neural network representing node level attention;

step 3.6, employing an attention mechanism to aggregate the meta-path basedNeighbor information of node(s)xIn the meta-path->Aggregation information of all neighbor nodes on +.>；

Step 3.7, let the set of meta-paths beCalculating the attention score of each element path,/-for each element path>Wherein->Is the attention score of the J-th element path, W is the weight matrix, b is the bias vector, q is the semantic level attention vector, and V is all node sets; />Is a nodexIn the meta-path->Aggregation information of all neighbor nodes on +. >Representing hyperbolic tangent function, ">Representing a transpose of the vector q;

step 3.8, obtaining the meta-paths by normalizing the attention scores of all meta-paths using a softmax functionAttention weight +.>，/>；

Step 3.9, learning the meta-pathIs used as a coefficient, the semantically specific embedding is fused to obtain the node +.>Based on the meta-wayFinal embedding of diameter->，/>；

Step 3.10, the initial stage in step 3.1Node learned in step 3.4xIs embedded with one-hop aggregation information>And the node learned in step 3.9xFinal embedding based on meta-paths>Aggregation to nodesxIn (3) obtaining a nodexFinal embedding:

wherein the method comprises the steps ofAnd->Is a weight matrix, < >>And->Is a bias vector; />The function is used for splicing a plurality of tensors along the appointed dimension to generate a new tensor; reLU is an activation function;

step 3.11, let the loss function be:

wherein,representing the number of scores>Representing the embedding of the ith user, +.>Indicating the embedding of the jth item,representing the score of the ith user on the real data set for the jth item;

step 3.12, after updating the heterograph node to embed, obtaining the user embedding of the source domainAnd article insert- >And user embedding of the target domain +.>And article insert->。

Further, in step 4, the metadata is used for the metadataGenerating a personalized bridge function for each user>Obtaining the initializing embedding of the source domain user in the target domain>Wherein->Is a network of elements which are connected in series,its parameter is->，/>Is the embedding of the user in the source domain,/-, for example>Is a parameter of a bridge function generated by means of a meta-network,/->Is a linear layer.

Further, in step 5, training the heterogeneous graph neural network model of the source domain by using the user-item heterograph of the source domain and the scoring of the item by the user of the source domain, training the heterogeneous graph neural network model of the target domain by using the user-item heterograph and the scoring of the target domain to remove the test set, training the meta network by using the common user, and storing the trained model comprises the following steps:

training the meta-network by using the cross entropy loss function, and continuously changing all weight parameters of the heterograph neural network model so as to minimize the cross entropy loss function, wherein the cross entropy loss function is as follows:

wherein,representing interactions of overlapping users in the target domain +.>Representing a common user of both domains, +.>Item set representing target domain, +.>Representing the ith user's relationship to the jth article Score, ->Representing user +.>Embedding in source domain->Parameters representing a bridge function generated via a meta-network, < >>Representing a userPersonalized bridge function of->Representing the embedding of the jth item;

and storing the trained heterogeneous graph neural network model of the source domain, the heterogeneous graph neural network model of the target domain and the meta-network.

Further, in step 6, the process of inputting the source domain user and the preset items to obtain Top-K items for recommendation includes the following steps:

step 6.1, calculating the source domain user u and all preset articlesPredictive score between->Representing the set of all preset items, obtaining the user embedding by the heterogeneous graph neural network model of the source domain>Generating a personalized bridge function of the user via the meta-network>；

Step 6.2, obtaining the converted user embedding according to the personalized bridging function；

Step 6.3, obtaining object embedding through the heterogeneous graph neural network model of the target domainAnd obtaining a predicted score of the user on the article by utilizing the user embedding and the article embedding: />；

And 6.4, sequencing all the obtained prediction scores to obtain Top-K articles for recommendation.

The beneficial effects are that:

firstly, according to the personalized cross-domain recommendation method based on network data driving, the heterogeneous graph neural network is applied to personalized cross-domain recommendation, and auxiliary information is introduced to construct a user-article heterogram, so that a specific relationship between a user and an article is revealed.

Secondly, according to the personalized cross-domain recommendation method based on network data driving, aiming at the complexity of the introduced auxiliary information, the three-level attention aggregation heterograph neural network model is adopted to generate the user embedding and the article embedding, so that the relevance between the user and the article is known more accurately, and the recommendation quality is improved.

Third, in the personalized migration, the personalized cross-domain recommendation method based on network data driving converts and adapts to the source domain user embedding by utilizing the meta-network so as to adapt to the characteristics of the target domain and the user preference, and can better realize cross-domain recommendation and improve the accuracy of recommendation results by generating a specific bridging function for each user.

Drawings

FIG. 1 is an exemplary diagram of a "user-item iso-composition" designed based on a network data driven personalized cross-domain recommendation method of an embodiment of the invention;

FIG. 2 is a meta-path exemplary diagram of a personalized cross-domain recommendation method based on network data driving in accordance with an embodiment of the present invention;

FIG. 3 is a pre-preparation flow chart of a personalized cross-domain recommendation method based on network data driving according to an embodiment of the invention;

FIG. 4 is a diagram showing an example of details of a TAHIN model of an embodiment of the present invention;

FIG. 5 is a flow chart of an implementation of a personalized cross-domain recommendation method based on network data driving in accordance with an embodiment of the present invention;

FIG. 6 is a schematic diagram of extracted key information;

FIG. 7 is a converted "item-brand" schematic;

FIG. 8 is a schematic diagram of a converted "article-sort";

FIG. 9 is a schematic diagram of an "item-historically browsed item" after data conversion;

FIG. 10 is a schematic diagram of "user-item-scoring" after data conversion;

FIG. 11 is a mapped "item-brand" schematic;

FIG. 12 is a schematic view of the mapped "item-class";

FIG. 13 is a schematic diagram of a mapped "item-historically browsed item";

fig. 14 is a schematic diagram of the mapped "user-item-score".

Detailed Description

The following examples will provide those skilled in the art with a more complete understanding of the invention, but are not intended to limit the invention in any way.

Referring to fig. 3, the embodiment discloses a personalized cross-domain recommendation method based on network data driving, which comprises the following steps:

step 1, data preprocessing is carried out on collected user data and object data of a source domain and a target domain, user-object heterograms of the source domain and the target domain are respectively constructed according to the preprocessed data, and a source domain heterogram and a target domain heterogram are generated;

Step 2, carrying out negative sampling on the user-object related edges of the user-object heterogeneous graphs of the source domain and the target domain, dividing a training set and a testing set according to common users, and setting a meta-path;

step 3, performing representation learning on nodes in the user-object heterograph through a heterograph neural network model (TAHIN model) to respectively obtain user node embedding and object node embedding;

step 4, embedding source domain user nodes and target domain object nodes into an input meta-network to generate personalized bridging functions;

training the heterogeneous graph neural network model of the source domain by utilizing the user-article heterograph of the source domain and the scoring of the article by the source domain user, training the heterogeneous graph neural network model of the target domain by utilizing the user-article heterograph and the scoring of the target domain to remove the test set, training the meta-network by utilizing the common user, and storing the trained model;

and 6, inputting a source domain user and preset articles to obtain Top-K articles for recommendation.

The method and the device for recommending the goods in the target domain of the platform are capable of recommending the goods optimally for a new user from massive goods in the target domain of the platform, and can achieve personalized goods recommendation in the target domain of the platform according to interaction information of the user on the source domain of the platform.

Specifically, the embodiment firstly utilizes a data preprocessing technology to structure data based on personal interaction information of a user in a source domain and commodity information (such as scoring of the user on an article, relationship between the article and brands, and the like) of the source domain so as to firstly establish a user-article heterogram of the source domain; the source domain heterogeneous graph comprises a large number of nodes and edges, wherein the nodes and the edges can be flexibly designed according to different collected data, and in the heterogeneous graph model, the complex relationship between a user and an article can be deeply understood through characteristic learning and representation learning on the nodes and the edges; and generating a personalized bridging function for each user by using a meta-network, converting the embedding of the user in the source domain into the target domain through the personalized bridging function, and taking the converted embedding of the user as the initial embedding of the cold start user in the target domain. The user of the target domain is utilized to initialize the embedding and the embedding of the target domain object, and the recommendation of the target domain commodity can be carried out for the user.

Compared with the traditional recommendation system, the recommendation method of the embodiment has the advantages that the cross-domain recommendation can effectively relieve the data sparseness problem and the cold start problem; in addition, compared with the traditional recommendation method only according to the interaction information of the user and the article, the problems of limitation of personal preference and information isolation exist, the scheme adds the auxiliary information of the article, and by adopting the heterogeneous graph model, the system can fully utilize the similarity relationship or the complementary relationship between the articles, so that more comprehensive and accurate article recommendation can be provided for the user; the meta-network is used as a bridging function, so that the recommendation of the user in the target domain can be more personalized; in order to make the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.

The specific process of the step 1 is as follows:

step 1.1, acquiring user data and item data from a source domain system and a target domain system respectively, wherein the item data comprises brands of items, classification of the items and historical browsing before browsing the items, and scoring matrixes of the items by users in the source domain and the target domain.

Step 1.2, presetting the articles to include all the collected articles in the target domain, which is recorded as。

And 1.3, carrying out structuring processing on the collected user and article data.

And step 1.4, mapping the character strings in the processed structured data into numerical values, and storing the mapping relation as a user dictionary and an article dictionary.

And 1.5, respectively constructing a source domain iso-graph and a target domain iso-graph according to the processed data, and designing the node type and name of the iso-graph and the type and name of the heterogeneous graph edge.

Structuring refers to the organization, and transformation of data into a specific structure and format for better analysis, storage, and use. The user and article interaction information or article information which can be collected is obtained from the website in a crawling way, so that the data cannot be directly used, and the data can be used only by structured processing. Common structuring treatments include the following:

(1) Data cleaning: data cleansing refers to detecting and correcting errors, deletions, duplicates, or inconsistencies in data. This includes operations to delete duplicate data, fill in missing values, repair erroneous data, etc., to ensure accuracy and consistency of the data.

(2) Data conversion: data conversion includes the operation of converting data from one representation to another. For example, converting unstructured data to structured data, or converting data from one data source to the format of another data source. This may involve data format conversion, data type conversion, unit conversion, etc.

(3) Data integration: data integration is the process of merging and integrating data from different sources. In the data integration process, the problems of inconsistent data modes, key mismatch, data redundancy and the like need to be solved, so that the integrated data has a uniform structure.

(4) Data normalization: data normalization refers to mapping data to a range or format of standards to facilitate comparison and analysis. Common data normalization methods include normalization (scaling the data to the 0-1 range) and normalization (conforming to a normal distribution).

(5) Data aggregation: data aggregation is the process of collecting and statistically analyzing data. By aggregating the data, useful information and metrics can be extracted from a large number of details to aid in decision making and analysis.

For example, in a CDs_and_Vinyl system, the collected user information and item information may be in JSON format, as follows, the first is the item information: { 'asin': '0764476270', 'protected', 'also_boight', 'B00BV9QRP6', 'B003VHSXIU', 'B004NNIV10', 'B002CNIUNW', 'B001 FJN', 'B000BVD3PM', '1470707748', '0764476254', '0764449990', 'B000VYP O', '555861837X', 'B0006V75H4', 'B004NNJTR0', '0159024684', 'B0026XVACQ', 'B00358RQI8', 'B001 qqfucy', 'B00020 jqqv 009', 'B004O0H7V2', 'B000MAB1PG', 'B08 UPI', 'B XIH FO', 'B000_ toga', 'B003 sound', 'B003 NNJTR0', '0159024684', 'B0026X', 'B0026 nnjv 8', 'B0026xv 8', 'B001 qxic Everything Is Possible with God (Mark 10: 27)) @, 18.0,' salesRank ': {' Music ':52601},' imUrl ': http:// ecx.images-amazon.com/images/I/41TNs1at1zL _, SY300_, jpg', 'brand': group Publishing ',' category ': [ (CDs & Vinyl', 'Christian', 'child's ], 'description' ], 'Up & Away Sing & Play Music0' }; the second is the user's evaluation and scoring of the item: { "review ID": A31GBCW6YPY OW "," asin ": 0764476270", "review name": "Dave child," "hellpful": [0,0], "review text": "ok I guess a little over 2 hours was notenough for some but I thought the dvd was better then the vhs was", "review": 5.0 "," review ":" great late 90' sconcert "," unixreview time ": 1378857600", "review time": "09 11, 2013" }.

Key information needed for constructing the heterogeneous map is selected from the two json data, for example, the ID of a user, the ID of an article, the brand of the article, the classification of the article, the historical browsing set of the article and the grading of the article by the user are extracted, and the extracted content is shown in fig. 6. The json format is then converted to csv or dat format required to construct the iso-graph by data conversion, as shown in fig. 7-10.

Some character strings needing to be processed can be mapped into numerical values for operation. This mapping process may be referred to as encoding or indexing. Common character string encoding methods include word encoding, single-hot encoding, hash encoding, and the like. Word encoding is the mapping of strings into integers, each different string corresponding to a unique integer value.

For example, a vocabulary may be created, with each string mapped to an index in the vocabulary. Creating a user dictionary, wherein the keys of the dictionary are the extracted user IDs, the values of the dictionary are index sequences, and the user IDs can be mapped into index values through the dictionary. Similarly, the object ID, brand, and category may all be mapped, and the resulting mapped csv file is shown in fig. 11-14.

The specific process of the step 2 is as follows:

step 2.1, the user-object side of the heterogeneous graphs of the source domain and the target domain is represented in a triplet mode.

And 2.2, carrying out negative sampling on the generated source domain and target domain heterogeneous graphs, namely randomly constructing a plurality of negative sides, combining the positive sides and the negative sides, disturbing the ordering of the triples, and dividing the common users of the source domain and the target domain into a training set and a testing set according to a certain proportion.

Training a link prediction model involves comparing the difference in score between two connected nodes with the score between any pair of nodes. For example, given an edge connecting u and v, a good model expects a score between u and v that is higher than the score between u and the node v' sampled from an arbitrary noise distribution, a method called negative sampling.

The specific way of performing negative sampling on the ui (user-item) side is as follows: the edge ui is expressed in the form of triplets (user ID, item ID, user's score for item), and if user's score, the score data collected in json is given, the triplets are expressed as (user ID, item ID, 5), false edges for some users and unrelated items are randomly generated, the scores of these edges are expressed as 0, and the triplets are expressed as (user ID, item ID, 0).

Step 2.3, for the heterograms of the source domain, the division of the training set and the verification set is not carried out, and the data of the source domain are all used as training; deleting the edge containing the test set user from the target domain iso-graph to obtain a processed target domain iso-graph, and storing the processed iso-graph, the training set and the test set;

and 2.4, respectively setting element paths of a source domain and a target domain required by heterogeneous graph neural network training.

Hypothetical meta-pathsIs defined as +.>(abbreviated as->) It describes the object->And->A complex relationship between->Wherein->Representing the composite operator on the relationship.

As shown in fig. 2, two items may be connected by multiple meta-paths, such as item-user-item, item-brand-item, and so on. Different meta-paths may reveal different semantics. For example, the item-user-item path, which means that both items are purchased by the same user, may be complementary or may have the same style. The item-brand-item element path indicates that the two items belong to the same brand. In addition, the meta-path may also connect different node types, such as user-item-brand, indicating that the user may be more inclined to purchase items published by the brand, the user being contemplating the brand.

The specific process of the step 3 is as follows:

step 3.1, initializing a source domain isomerism mapAnd target Domain isomerism map->Embedding of different types of nodes, V is node set, < ->Is an edge set, different node types correspond to different dimensionsNodexIs embedded as +.>The initialization of node w is embedded as +.>The initialization of node j is embedded as +.>。

An iso-pattern may be expressed as，/>Is a node set, ++>Is an edge set, is also a mapping function with node type +.>And an edge type mapping function ++>Associated, A and R represent a set of predefined node types and edge types, wherein +.>. In the invention, the designed user-object different composition is shown in figure 1, and the node types of the user-object different composition are 5, namely a user node u, an object node i, a brand node b, a classification node c and a history browsing object node y; the types of the edges are 8, namely: edges ui and iu between the user and the item represent the relationship between the user purchasing the item and the item purchased by the user, edges ib and bi between the item and the brand represent that the item belongs to the brand and that the brand has issued the item, edges ic and ci between the item and the category represent that the item is categorized into the category and that the item is contained in the category, edges iy and yi between the item and the historically browsed item represent purchases The items that the item would have previously viewed and the items that the item would have purchased after viewing.

Step 3.2, using projection matrixDifferent types of neighbor nodes of the target node are projected to the same dimension of the target node.

The projection matrix P is a special matrix that can project one vector onto the subspace in which the other vector is located. Assuming a projection matrix P, its dimensions areWhere m is the feature dimension after projection and n is the dimension of the original vector. To project an original vector v into a space with a feature dimension m, the following formula can be used: projection vector->Wherein (1)>Representing a matrix multiplication operation.

Step 3.3, calculating attention weights for the target node and other neighbors thereof, respectively, such as calculating the attention weight between the node x and the node w:wherein->Is all one-hop neighbors of node x.

In the heterogeneous graph, one-hop neighbors refer to all nodes directly connected to a given node. An iso-graph is a graph made up of different types of nodes and edges, each node and edge having its own type. Thus, a one-hop neighbor includes nodes that are directly connected to a given node by any type of edge.

Specifically, given a node in an iso-patternxIts one-hop neighbor comprises the following steps: finding all edges directly connected to node eThe method comprises the steps of carrying out a first treatment on the surface of the For each edge, find the node at the other end, i.e. with the nodexConnected neighbor nodes; adding these neighbor nodes to a one-hop neighbor setIs a kind of medium.

In models such as graph neural networks, attention mechanisms are widely used to weight aggregate information of nodes or edges.

Assume that there is one target nodexAnd its one-hop neighbor node setAn attention mechanism may be used to calculate an attention weight between the target node and each neighbor node. The following steps may be used to achieve this: according to the target nodexAnd the characteristics of the neighbor nodes, calculating the attention score between the characteristics; the attention score is converted into attention weights, so that the sum of the attention weights is 1, and the relative importance of the attention weights corresponding to different neighbor nodes can be ensured. The features of the neighbor nodes are weighted aggregated using the attention weights. The target node can be obtained by multiplying the attention weight by the features of the one-hop neighbor nodes and then summing all the neighbor nodes xIs a polymeric representation of (a).

Step 3.4, performing one-hop neighbor aggregation of the target node according to the calculated attention weight,wherein->Is an activation function.

Step 3.5, two nodes can also be connected by a meta-path, since different neighbors based on the meta-path imply different information, the attention weight of the neighbors based on the meta-path is calculated, in the meta-pathThe upper part of the upper part is provided with a plurality of grooves,，/>for the meta-path->Upper nodexIs a set of all neighbor nodes.

Nodes in a given iso-patternxSum element pathNode is connected withxIs based on the neighbors of the meta-path->Defined as passing meta-path->And nodexA set of connected nodes. The neighbors of a node include itself. Taking fig. 2 as an example, given a meta-path IBI (item-brand-item), the meta-path based neighbors of item 2 include item 2 (itself), item 1 and item 3, it is apparent that the meta-path based neighbors may take advantage of different aspects of the structural information in the iso-graph, and may be derived by the product of the adjacency matrix sequences.

Step 3.6, employing an attention mechanism to aggregate the meta-path basedIs a function of the information of the neighbors of (a),wherein->Is an activation function.

Step 3.7, the set of meta-paths isThe attention score for each meta-path is calculated, Where W is the weight matrix, b is the bias vector, q is the semantic level attention vector, and V is the set of all nodes.

The attention score is a score used in the attention mechanism to measure the relevance of a query to a key. It is used to calculate an attention weight indicating how important the query is to each key. In the attention mechanism, the attention score can be obtained by different calculation methods, and the following two are the most common:

(1) Dot product attention (Dot Product Attention): the dot product between the query vector and the key vector is used as the attention score. The specific calculation mode is as follows:。

(2) Additive attention (Additive Attention): a linear combination between the query vector and the key vector is used as the attention score. The specific calculation mode is as follows:wherein, the method comprises the steps of, wherein,and->Is a matrix of parameters that can be learned to map query vectors and key vectors onto the same dimension.

Step 3.8, after obtaining the attention score of each meta-path, normalizing them by the softmax function, meta-pathsIs expressed as +.>The above attention weights of all meta-paths can be normalized by using a softmax function to obtain +. >。

Attention weighting is one way to measure the relevance or importance between different elements in some machine learning and deep learning models. After deriving the attention score, the score may be converted to an attention weight using a softmax function. The softmax function may normalize a set of scores to a probability distribution, ensuring that the sum of all attention weights is 1.

Specifically, assume that there is a set of attention scoresWherein->Representing the attention score of the ith key and query. Attention weight->The method can be obtained by the following calculation:

in the above-mentioned formula(s),is a natural exponential function, ++>Representing a summation operation.

After converting the attention score into an attention weight by calculating the softmax function, a set of normalized weight vectors is obtained. These weights may reflect the importance of each key to the query or the degree of attention allocation. In subsequent operations, these attention weights may be used to weight aggregate or otherwise operate on the corresponding values to achieve a more accurate model representation or task processing.

Step 3.9, taking the learned weight as a coefficient, fusing the semantic specific embedments to obtain the final embedment of the node x based on the meta-path, 。/>

Step 3.10, the initial stage in step 3.1The +.f learned in step 3.4>And learned in step 3.9Aggregating into node x to obtain the final embedding of node x:

in the method, in the process of the invention,and->Is a weight matrix, < >>And->Is a bias vector.

Step 3.11, the loss function is:wherein->Representing the number of scores>Representing the embedding of the ith user, +.>Representing the embedding of the jth item +.>Representing the score of the jth item by the ith user on the real dataset.

Step 3.12, after updating the heterograph node embedding, the user and object embedding of the source domain and the target domain can be obtained, which are respectively expressed as、/>、/>And->。

The specific process of the step 4 is as follows:

step 4.1, via a meta-networkGenerating a personalized bridging function for each userObtaining the initializing embedding of the source domain user in the target domain>。

The specific process of the step 5 is as follows:

step 5.1, training a meta-network by using a cross entropy function, wherein the cross entropy loss function is as follows:

wherein,representing interactions of overlapping users in the target domain;

step 5.2, updating the ownership parameter of the model;

step 5.3, continuously correcting the weight parameters in the iterative process of the algorithm so as to obtain the minimum loss function;

And 5.4, saving the trained source domain and target domain heterogeneous graph neural network model and the meta-network.

The specific process of the step 6 is as follows:

step 6.1, calculating the source domain user u (assuming that the user has interactions with the items in the source domain, i.e. the user has a history of behavior in the source domain, and the user is a cold-start user in the target domain, and has no interactions with the items in the target domain) and all preset itemsThe prediction score between the two is obtained by a source domain heterogeneous graph neural network model to obtain user embedding ++>Generating a personalized bridge function of the user via the meta-network>；

Step 6.2, obtaining the converted user embedding according to the personalized bridging function；

Step 6.3, obtaining object embedding through the heterogeneous graph neural network model of the target domainAnd obtaining a predicted score of the user on the article by utilizing the user embedding and the article embedding: />；

And 6.4, sequencing all the obtained prediction scores to obtain Top-K articles for recommendation.

Taking samples of two different fields of A (source domain) and B (target domain) of a certain platform as examples, the invention recommends an article for a certain cold start user in the field of B of the platform, and presumes that the user has interactive records in the field A of the platform. A pre-preparation flow chart of the present invention is shown in fig. 3. The specific operation steps are as follows:

Step 1: obtaining information of users and articles from the platform A domain and the platform B domain, wherein the information comprises IDs of the users and the articles, brands of the articles, classifications of the articles, historically browsed articles and the like, and the historically browsed articles refer to articles which are browsed before browsing the articles, and the articles possibly have similar brands or similar classifications; the preset articles comprise all articles in the B domain; carrying out structuring processing and character string mapping on collected data, wherein the data is collected from a webpage and is usually expressed in json format, and simultaneously contains a large number of character strings, and extracting required key information from the json data to form csv format data or dat format data; creating dictionaries such as a user dictionary, an article brand dictionary, an article classification and the like, and mapping the character strings into numerical forms which can be processed by machine learning and deep learning through word coding; and constructing a user-object iso-composition of the A domain and the B domain respectively by using the mapped csv or dat data.

Step 2: it is also critical to design a suitable heterogram that preferably contains as much information as the user and the item are related to, but at the same time avoids noise information as much as possible, so as not to reduce accuracy and waste computing resources. In the method of the invention, an iso-composition shown in figure 1 can be designed, 5 nodes are provided, including users, articles, brands, classifications and historical browsing articles, and the 5 types of nodes can summarize the positioning and attribute information of the articles and the key information and personal preference of the users in each field as much as possible; the types of edges are 8, including user items, item-user, item-brand, brand-item, item-class, class-item, item-history browsing item, and by using these 8 types of edges, the potential relationships between these nodes can be summarized as much as possible; in order to further mine the semantic level relationships between these nodes, it is necessary to set some meta-paths that need to be manually designed based on experience, for example, to design a meta-path like "user-item-user" that can mine that both users have purchased the same item.

Step 3: as a recommendation model designed for cold start users, the task of the model is to recommend the adapted items for the cold start users as accurately as possible; and dividing overlapped users into a training set and a testing set according to a proportion, wherein scores of all users and articles in the A domain are used as the training set, and users except the testing set in the B domain are used as the training set, and constructing heterogeneous graphs of a source domain and a target domain and setting meta paths according to the divided training set.

Step 4: putting the processed iso-graph and the set meta-path into a TAHIN (Three-level Attention Aggregation Heterogenous Information Network) model for training, wherein the details of the model are shown in figure 4; firstly, initializing nodes of various types; when the target node is subjected to embedded calculation, a one-hop neighbor set of the target node is collected, the one-hop neighbors comprise multiple types of nodes, and all the nodes are projected to the dimension of the target node through a projection matrix so as to carry out subsequent calculation; firstly calculating the attention weight of each one-hop neighbor by using an attention mechanism, and obtaining the embedding of the relation neighbor aggregation of the target node by weighting aggregation; on the set meta-path, a neighbor based on the meta-path of the target node is found, and the attention weight of the neighbor based on the meta-path is calculated firstly by using an attention mechanism, and then weighted aggregation is carried out, so that node level attention aggregation based on the meta-path is obtained; then, the attention weight of each element path is calculated, the learned weight is used as a coefficient, and the semantic specific embedments are fused to obtain the final embedment of the target node based on the element path; finally, initializing embedding, embedding of relation neighbor aggregation and embedding aggregation based on meta-paths, wherein the aggregation can adopt the simplest vector connection operation.

Step 5: after obtaining the user embedding and the article embedding of the A domain and the B domain, training the meta-network by utilizing task-oriented optimization, namely training the meta-network by overlapping the article embedding of the A domain and the B domain and the scoring matrix of the article by the users, calculating a loss function by utilizing a prediction score and a true data set score, then calculating a gradient by back propagation, updating model parameters, continuously correcting weight parameters so as to obtain the minimum loss function, and storing a training model.

Step 6: the model was tested using a test set, and the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of the model were calculated using them as metrics.

Step 7: as shown in FIG. 5, when a cross-domain recommendation is made, firstly, inputting a domain A user (the user is a cold start user relative to a domain B) and a preset article (article of the domain B), loading a trained domain A TAHIN model to obtain the embedding of the user, loading a trained domain B TAHIN model to obtain the embedding of the preset article, inputting the user embedding of the domain A by using a trained meta-network to generate a personalized bridging function of the user, obtaining the initialized embedding of the user in the domain B by the bridging function, performing dot product by the initialized user embedding and article embedding, and calculating to obtain a prediction score between the user and the preset article to obtain the prediction score of the cold start user on the article of the domain B.

The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above examples, and all technical solutions belonging to the concept of the present invention belong to the protection scope of the present invention. It should be noted that modifications and adaptations to the invention without departing from the principles thereof are intended to be within the scope of the invention as set forth in the following claims.

Claims (8)

1. The personalized cross-domain recommendation method based on the network data driving is characterized by comprising the following steps of:

step 1, data preprocessing is carried out on collected user data and object data of a source domain and a target domain, user-object heterograms of the source domain and the target domain are respectively constructed according to the preprocessed data, and a source domain heterogram and a target domain heterogram are generated;

step 2, carrying out negative sampling on the user-object related edges of the user-object heterogeneous graphs of the source domain and the target domain, dividing a training set and a testing set according to common users, and setting a meta-path;

step 3, performing representation learning on nodes in the user-object heterograms through the heterograms neural network model to respectively obtain user node embedding and object node embedding;

Step 4, embedding source domain user nodes and target domain object nodes into an input meta-network to generate personalized bridging functions;

training the heterogeneous graph neural network model of the source domain by utilizing the user-article heterograph of the source domain and the scoring of the article by the source domain user, training the heterogeneous graph neural network model of the target domain by utilizing the user-article heterograph and the scoring of the target domain to remove the test set, training the meta-network by utilizing the common user, and storing the trained model;

and 6, inputting a source domain user and preset articles to obtain Top-K articles for recommendation.

2. The personalized cross-domain recommendation method based on network data driving according to claim 1, wherein in step 1, the process of constructing user-object iso-composition of source domain and target domain, and generating source domain iso-composition and target domain iso-composition a comprises the following steps:

step 1.1, respectively acquiring user data and article data from a source domain system and a target domain system, wherein the article data comprises brands of articles, classification of the articles, historical browsing before browsing the articles, and scoring matrixes of users in the source domain and the target domain for the articles;

step 1.2, carrying out structuring treatment on the collected user data and article data to obtain structured data after treatment;

Step 1.3, mapping the character strings in the processed structured data into numerical values, and storing the mapping relations as a user dictionary and an article dictionary;

and 1.4, respectively constructing a source domain iso-composition and a target domain iso-composition based on the structured data, the user dictionary and the object dictionary, and designing node types and names of the source domain iso-composition and the target domain iso-composition and types and names of the edges of the heterogeneous graphs.

3. The network data driven personalized cross-domain recommendation method according to claim 2, wherein the node types of the user-item iso-graph comprise user node u, item node i, brand node b, category node c and history browsing item node y;

types of edges of the user-object iso-graph include: the sides ui and iu between the user and the item represent the relationship that the user u purchases the item i and the item i is purchased by the user u, the sides ib and bi between the item and the brand represent that the item i belongs to the brand b and issues the item i, the sides ic and ci between the item and the category represent that the item i is classified into the category c and contains the item i, and the sides iy and yi between the item and the historically browsed item represent the item y browsed before the item i is purchased and the item i that is purchased after the item y is browsed.

4. The personalized cross-domain recommendation method based on network data driving according to claim 1, wherein in step 2, the process of performing negative sampling on user-object of the edges on the user-object heterogeneous graphs of the source domain and the target domain, dividing the training set and the test set according to the common user, and setting the meta path comprises the following steps:

step 2.1, representing the edges of the user-object of the source domain heterograms and the target domain heterograms in a triplet mode;

step 2.2, randomly constructing a plurality of negative sides, combining the positive sides and the negative sides, and completing negative sampling of the sides of the source domain iso-graph and the target domain iso-graph about the user-object; the sorting of the triples is disordered, and for common users of a source domain and a target domain, the triples after being disordered are divided into a training set and a testing set according to a preset proportion;

step 2.3, deleting the edge containing the test set user from the target domain heterogeneous graph to obtain a processed target domain heterogeneous graph; storing the processed target domain heterogeneous graph, training set and test set;

step 2.4, respectively setting element paths of a source domain and a target domain required by heterogeneous graph neural network training; meta pathIs defined as +. >For describing node->And node->Composite relationship between->Wherein->Representing the composite operator on the relationship.

5. The personalized cross-domain recommendation method based on network data driving according to claim 1, wherein in step S3, the process of performing representation learning on the nodes in the user-object heterogram through the TAHIN model to obtain user node embedding and object node embedding respectively comprises the following steps:

step 3.1, initializing a source domain isomerism mapAnd target Domain isomerism map->Embedding of different types of nodes, V is node set, < ->Is an edge set, different node types correspond to different dimensions, nodesxIs embedded as +.>Initial of node wIs embedded as->The initialization of node j is embedded as +.>；

Step 3.2, using projection matrixProjecting different types of neighbor nodes of the target node to the same dimension of the target node;

step 3.3, calculating the attention weight for the target node and other neighbors, respectively, and calculating the nodexAnd attention weight between nodes wThe method comprises the following steps: />Wherein->Is a nodexIs a set of all one-hop neighbors +.>Is a deep neural network representing the relationship between two nodes note, < > >Is an exponential function;

step 3.4, according to the calculated attention weightPerforming one-hop neighbor aggregation of target nodes and nodesxOne-hop neighbor aggregation information of->Wherein->Is an activation function;

step 3.5, calculating the attention weight of the neighbor based on the meta-path, in whichOn the nodexAnd the attention weight between nodes w +.>，/>For the meta-path->Upper nodexIs>Is a deep neural network representing node level attention;

step 3.6, employing an attention mechanism to aggregate the meta-path basedNeighbor information of node(s)xIn the meta-path->Aggregation information of all neighbor nodes on +.>；

Step 3.7, let the set of meta-paths beThe attention score for each meta-path is calculated,wherein->Is the attention score of the J-th element path, W is the weight matrix, b is the bias vector, q is the semantic level attention vector, and V is all node sets; />Is a nodexIn the meta-path->Aggregation information of all neighbor nodes on +.>Representing hyperbolic tangent function, ">Representing a transpose of the vector q;

step 3.8, obtaining the meta-paths by normalizing the attention scores of all meta-paths using a softmax functionAttention weight +. >，/>；

Step 3.9, learning the meta-pathIs used as a coefficient, the semantically specific embedding is fused to obtain the node +.>Final embedding based on meta-paths>，/>；

Step 3.10, the initial stage in step 3.1Node learned in step 3.4xIs embedded with one-hop aggregation information>And the node learned in step 3.9xFinal embedding based on meta-paths>Aggregation to nodesxIn (3) obtaining a nodexFinal embedding:

wherein the method comprises the steps ofAnd->Is a weight matrix, < >>And->Is a bias vector; />The function is used for splicing a plurality of tensors along the appointed dimension to generate a new tensor; reLU is an activation function;

step 3.11, let the loss function be:

wherein,representing the number of scores>Representing the embedding of the ith user, +.>Representing the embedding of the jth item +.>Representing the score of the ith user on the real data set for the jth item;

step 3.12, after updating the heterograph node to embed, obtaining the user embedding of the source domainAnd article insert->And user embedding of the target domain +.>And article insert->。

6. The network data driven personalized cross-domain recommendation method according to claim 1, wherein in step 4, the personalized cross-domain recommendation method is performed through a meta-network Generating a personalized bridge function for each user>Obtaining the initializing embedding of the source domain user in the target domain>Wherein->Is a meta-network with a parameter of +.>，/>Is the embedding of the user in the source domain,/-, for example>Is a parameter of a bridge function generated by means of a meta-network,/->Is a linear layer.

7. The personalized cross-domain recommendation method based on network data driving of claim 1, wherein in step 5, the heterogeneous graph neural network model of the source domain is trained by using the user-item iso-graph of the source domain and the scoring of the item by the user of the source domain, the heterogeneous graph neural network model of the target domain is trained by using the user-item iso-graph and the scoring of the target domain to remove the test set, the meta-network is trained by using the common user, and the process of storing the trained model comprises the following steps:

training the meta-network by using the cross entropy loss function, and continuously changing all weight parameters of the heterograph neural network model so as to minimize the cross entropy loss function, wherein the cross entropy loss function is as follows:

wherein,representing interactions of overlapping users in the target domain +.>Representing a common user of both domains, +.>Item set representing target domain, +.>Indicating the i-th user's score for the j-th item,/-th item >Representing user +.>Embedding in source domain->Parameters representing a bridge function generated via a meta-network, < >>Representing user +.>Personalized bridge function of->Representing the embedding of the jth item;

and storing the trained heterogeneous graph neural network model of the source domain, the heterogeneous graph neural network model of the target domain and the meta-network.

8. The personalized cross-domain recommendation method based on network data driving according to claim 1, wherein in step 6, the process of inputting source domain users and preset articles to obtain Top-K articles for recommendation comprises the following steps:

step 6.1, calculating the source domain user u and all preset articlesPredictive score between->Representing the set of all preset items, obtaining the user embedding by the heterogeneous graph neural network model of the source domain>Generating a personalized bridge function of the user via the meta-network>；

Step 6.2, obtaining the converted user embedding according to the personalized bridging function；

Step 6.3, obtaining object embedding through the heterogeneous graph neural network model of the target domainAnd obtaining a predicted score of the user on the article by utilizing the user embedding and the article embedding: />；

And 6.4, sequencing all the obtained prediction scores to obtain Top-K articles for recommendation.