CN115168653A - Cross-domain recommendation method and system based on inter-domain and intra-domain relationships - Google Patents
Cross-domain recommendation method and system based on inter-domain and intra-domain relationships Download PDFInfo
- Publication number
- CN115168653A CN115168653A CN202210547462.XA CN202210547462A CN115168653A CN 115168653 A CN115168653 A CN 115168653A CN 202210547462 A CN202210547462 A CN 202210547462A CN 115168653 A CN115168653 A CN 115168653A
- Authority
- CN
- China
- Prior art keywords
- domain
- cross
- user
- users
- intra
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/903—Querying
- G06F16/9035—Filtering based on additional data, e.g. user or group profiles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
A cross-domain recommendation method and a system based on inter-domain and intra-domain relations are disclosed, wherein cross-domain heteromorphic graphs are constructed and obtained according to the fact that comment information among users or articles in different domains is converted into semantic relations among the users or articles, and single-domain heteromorphic graphs are constructed and obtained by utilizing the semantic relations among the users or articles in each domain; and designing a cross-domain recommendation system model based on inter-domain and intra-domain relations on the basis of the two different composition graphs, wherein the cross-domain recommendation system model is used for synchronously learning the cross-domain invariance and single-domain specificity of the users or articles, and generating an individual recommendation result for each user through a gating fusion mechanism. According to the method, the semantic relation among the domains is converted into the cross-domain abnormal composition, the semantic relation in the domains is converted into the single-domain abnormal composition, meanwhile, the inter-domain and intra-domain relation systems are reserved, the high-order inter-domain and intra-domain relation is modeled uniformly, and the double-target cross-domain recommendation is realized.
Description
Technical Field
The invention relates to a technology in the field of information processing, in particular to a cross-domain recommendation method and a cross-domain recommendation system based on inter-domain and intra-domain relationships.
Background
The graph convolution technology mainly uses convolution technology to mine local information on a graph structure, and the core idea is to iteratively aggregate neighbor nodes through a propagation mechanism to update the characteristics of a target node. The graph convolution technique is mainly divided into two categories: user-item interaction graph convolution techniques and heterogeneous graph convolution techniques. For the former, the interaction data of the user and the item is converted into bipartite graph structure data, in which: the user and the item are considered nodes and the interaction between the user and the item is considered edges. However, this single user-item interaction only records the interaction between the user and the item, and ignores the other interactions that exist.
In order to explore the multi-type interaction relationship, researchers put forward modeling on social information of users, attribute information of articles, interaction sequence information and the like, and a heteromorphic graph is constructed. For example, a user-user interaction graph is constructed by using social information of a user, and researchers can explore social relationships of the user; an article-attribute interaction diagram is constructed by using a knowledge graph of an article, and researchers can explore the attribute relationship of the article; a user-user or article-article interaction graph is constructed by utilizing comment information of a user on an article, and researchers can explore semantic relations of the user or the article.
Disclosure of Invention
The invention provides a cross-domain recommendation method and a cross-domain recommendation system based on inter-domain and intra-domain relations, aiming at the defects that the existing cross-domain recommendation system only utilizes intra-domain relations and does not consider the relations between different users or objects between domains.
The invention is realized by the following technical scheme:
the invention relates to a cross-domain recommendation method based on inter-domain and intra-domain relations, which comprises the steps of converting comment information of users or articles in different domains into semantic relations of the users or the articles, constructing to obtain cross-domain heteromorphic graphs, and constructing to obtain single-domain heteromorphic graphs by utilizing the semantic relations of the users or the articles in each domain; and designing a cross-domain recommendation system model based on inter-domain and intra-domain relations on the basis of the two different composition graphs, wherein the cross-domain recommendation system model is used for synchronously learning the cross-domain invariance and single-domain specificity of the users or articles, and generating an individual recommendation result for each user through a gating fusion mechanism.
The invention relates to a system for realizing the method, which comprises the following steps: preprocessing unit, heterogeneous graph building unit, embedding layer, graph convolution layer, gate-controlled fusion layer and output layer, wherein: the preprocessing unit carries out primary processing on the original scoring and comment data to obtain input data; the heterogeneous graph building unit generates a relation matrix reflecting multi-dimensional relations among users, articles and users, and builds two single-field heterogeneous graphs and one cross-field heterogeneous graph on the basis of the relation matrix; the embedding layer converts the high-dimensional heterogeneous graph into a vector of a low-dimensional space; the graph convolution layer considers different types of relations between nodes, and utilizes the heterogeneous graph neural network technology to convert vectors into low-dimensional space for modeling so as to obtain relation-aware feature vectors; the gating fusion layer carries out deep fusion on the feature vectors and further extracts key features; and the output layer calculates the feature vector expression of the final user and the article according to the key features, and uses an output function as the interactive relation between the user and the article.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of the system of the present invention.
Detailed Description
As shown in fig. 1, this embodiment relates to a cross-domain recommendation method based on inter-domain and intra-domain relationships, which includes:
1.1 remove noise present in < user, item, raiting, review > interaction records, e.g. delete records with no score or comment, delete records with too short a comment;
1.2 aiming at the user and article interaction data in two different fields, selecting users in the two fields which are completely the same, and removing the interaction records of the different users between the two fields;
1.3 after data preprocessing, the user sets U in the two fields are completely the same, and the article sets I and J in the two fields are completely different.
2.1, basic definition: the common user set between the two fields is U, the different item sets of the two fields are I and J respectively, and the number | U | = M of the common user set; the number of different articles in the areas A and B is I | = N A And | J | = N B Wherein: \8230Theuser-item relationships (including the user-item relationships of domains A and B) are established using scoring data, and all the user-item relationships in domain A form a matrixAll user-item relationships in Domain B form a matrixUser-user relationships (including cross-domain user-user relationships and single-domain user-user relationships) and item-item relationships (including cross-domain item-item relationships and single-domain item-item relationships) are established using the review data.
2.2 construction of Cross-DomainPatterning by different patterns: first, user-user relationships between domains are established. For a certain user u, all comments of the user are connected to generate a user document d u . Converting text documents to fixed-size vector representation D using a BERT model u (ii) a Probability of user-user relation obtained by calculating cosine similarityWherein: ψ (x) = max (0, x) is a ReLU function; calculating cosine similarity between all different users in different fields to obtain a cross-field user-user relation matrixSimilarly, a cross-domain item-item relationship matrix may be obtained Composition domain A user-item relationship composition matrix R A Domain B user-item relationships form a matrix R B Cross-domain user-user relationship matrix C U Cross domain item-item relationship matrixA cross-domain metamorphic pattern can be formed, namely:wherein: are each R A ,R B ,C I The transposed matrix of (2).
2.3, constructing a single-field abnormal graph: obtaining a domain A user-user relationship matrix by establishing user-user relationships in domain ABy establishing an item-item relationship in Domain A, a Domain A item-item relationship matrix may be obtainedFinally, a domain a differential map is constructed:wherein:similarly, a domain B differential map is constructed:wherein:
2.4, relation definition: in the isomerous graph G = G C ∪G A ∪G B In (1), there are eight different relationships, for a particular relationshipr =1222 823028, wherein:for the user-user relationship between the domains,for inter-domain item-to-item relationships,for the domain a user-user relationship,for the domain B user-user relationship,for the domain a article-to-article relationship,for the domain B article-to-article relationship,for domain a user-item relationships,is a domain B user-item relationship. In particular, it is possible to use, for example,
step 3, constructing a cross-domain recommendation model based on inter-domain and intra-domain relations according to the two heterogeneous graphs obtained in the step 2, learning a relation-aware vector expression for each user or article, and fusing cross-domain invariance and single-domain specificity through a gating fusion mechanism, wherein the model comprises the following steps: embedding layer, relation perception map convolution layer, gate fusion layer and output layer, wherein: the embedding layer adopts a one-hot vector X as a common user U, a field A article I and a field B article J, and converts the high-dimensional one-hot vector into low-dimensional expression; the relation sensing graph convolution layer considers the relations of different types among the nodes, simultaneously considers the relation types among the nodes and captures the heterogeneous connectivity between users and object objects in different graphs when aggregating the neighbor nodes of a target node, and then splices the vector expressions of different layers to form the final vector expression; gate-controlled fusion layer generating cross-domain invariance of users and items on cross-domain heteromorphic graphs using neural network models: (And) And single domain specificity: (And) Two gate control units sharing part of parameters are utilized to realize a dual-domain target recommendation task; output layer generating vector representations for end usersAndand vector representation of the final articleAndthe predicted probability between user u and item i or j is:
3.1, setting an embedding layer: converting the high-dimensional one-hot vector into low-dimensional expression, specifically comprising the following steps: h = XP, wherein:is a transformation matrix, d 1 <<d 0 . For each row h of the output matrix, there is a low-dimensional representation of a node (user or article), where:respectively a certain user u in a heterogeneous graph G C 、G A And G B The expression of the protein in the medium and low dimensions,respectively, a region A of a certain user i in a heterogeneous graph G C And G A The expression of the medium and low dimension is realized, respectively an object j in the field B in an isomeric diagram G c And G B Low dimensional expression of (b), h r Is in a relationship ofLow dimensional expression of (2).
3.2, relation perception graph convolution layer setting: in a graph convolution propagation mechanism, a vector expression of a relationship type is added to distinguish the heterogeneity of the relationship between different nodes, and the relationship perception of a user u for a heterogeneous graph convolution layer is specifically as follows: wherein:andare respectively node v r And the vector expression of the relation r after the k-th layer, sigma is a nonlinear activation function, W (k) As a weight matrix, N u Andrespectively node u and node v r Of the node(s) of (a) is,is a regularization constant term; vector representation of relationship typesThe updating method specifically comprises the following steps:wherein:the relationship and node vectors are mapped to the same space as the weight matrix. Note that H is defined as the initial vector representation of the graph-convolution network, i.e., E (0) H, in particular:the heterogeneous graph convolution layer may also describe a relationship-aware heterogeneous graph convolution in a matrix form, specifically: wherein:andare respectively node v r And the vector representation of the relation r after the k-th layer.Is a symmetric matrix normalization matrix, D is a diagonal matrix, and G is an adjacency matrix of the heteromorphic graph. Finally, the vector expressions of the different layers are spliced to form the final vector expression:wherein: and | is a join operation.For the final vector representation after the heterogeneous graph convolution operation, d 2 The size of the vector after the K-layer propagation and join operations is defined. By usingAfter the relation-aware heterogeneous graph convolution operation and the splicing operation, the vector expression of the user on different heterogeneous graphs can be output, namely G C Vector representation of upper user uAt G A Vector representation of upper user uAt G B Vector representation of upper user uSimilarly, can be found in G C Vector representation of item iAt G A Vector representation of item iAt G C Vector representation of item jAt G B Vector representation of item j
3.3, setting of a gated fusion layer: two gating cells sharing part of the parameters were designed to fuse these two properties: wherein:the laminated layer being two doorsThe control units share a weight matrix and,andis a matrix of weights for each of the two gating cells,for the fused vector representation of user u in domain a,fusion vector expression of user u in the field B; a universal gate control unit is adopted to fuse the cross-domain invariance and the single-domain specificity of the A-domain article i, and specifically the following steps are adopted:wherein:andis the respective weight matrix 2 of the two gating cellsTable field a fusion vector representation of item i. Similar toThe generation process of (2) can obtain the fusion vector expression of the cross-domain invariance and the single-domain specificity of the domain B article j
3.4, the output layer refers to: in order to explore the nonlinear high-order characteristics of user vector expression and article vector expression, the output layer adopts a multilayer perceptron (MLP) to model the interaction information of the user and the article, and the method specifically comprises the following steps: φ L =a L (S L φ L-1 +b L ),wherein: s l And b l Weight matrix and bias term, a, of the l-th layer, respectively l For activating the function, f (-) is the mapping function, and L mapping to predicted values for Domain ASimilarly, a prediction value for Domain B may be generated
Step 4, model training, specifically comprising: using cross entropy loss function Wherein:andloss functions for domain a and domain B respectively,the actual interactive data of the user-article can be used as a positive sample training set; for each positive sample, a plurality of non-interacted articles are sampled to be used as negative samples to form a negative sample training set
Since the objective of this embodiment is to optimize the recommended tasks of both domains as accurately as possible, the joint loss function is needed to be used when training the systemAndin combination, the joint loss function is: wherein:is a regularization term.
Step 5, generating a recommendation result: after the model is trained in a gradient descent mode, the prediction probability between all the items of the two fields is calculated for each userAndand then according to the estimated probability, sequencing the articles, wherein the articles with the sequencing result in front can be used as the result of personalized recommendation for the user, so that a task of recommending in two fields (dual-target recommendation) is realized.
To verify the model effect, the present embodiment establishes three cross-domain recommended tasks on amazon data sets to verify the effect of the present embodiment, which specifically includes:
5.1 detailed parameter settings during the experiment: selecting overlapped users from each pair of cross-domain data sets and filtering out non-overlapped users; the experiment was then performed using a tensflow framework. When a training set is constructed, four negative samples are randomly collected while each positive sample is selected. In the model training phase, adam is selected as an optimizer to update model parameters, and the learning rate is set to be 0.001. The amount of training data for each batch is set to 512 and the user and item embedding vector dimension is set to 32. Furthermore, using the dropout technique prevents the overfitting problem and sets the dropout size to 0.1; the number of layers of the final graph convolution is uniformly set to 4.
5.2 evaluation method. The embodiment adopts a cross-validation method (leave-one-out method) for evaluation, and compares the evaluation with the existing single-field method, the cross-field method based on the multilayer perceptron and the cross-field method based on the graph neural network. When the effect of the cross-domain recommendation system is verified, the Hit rate (abbreviated as HR) and the Normalized distributed additive discount Gain (abbreviated as NDCG) are used as evaluation indexes, and the performance of the model above HR @10 and NDCG @10 is specifically given. For each user, 99 items that have not interacted with the user are randomly drawn and combined with the positive sample that has interacted with the user to form a ranked candidate list in a ranking process.
5.3 evaluation results. Compared with the prior related work, the method is improved by 12.16 percent and 18.15 percent on average in three cross-domain recommended tasks.
The foregoing embodiments may be modified in many different ways by one skilled in the art without departing from the spirit and scope of the invention, which is defined by the appended claims and not by the preceding embodiments, and all embodiments within their scope are intended to be limited by the scope of the invention.
Claims (4)
1. A cross-domain recommendation method based on inter-domain and intra-domain relationships is characterized in that cross-domain heteromorphic graphs are constructed and obtained according to the fact that comment information among users or articles in different domains is converted into semantic relationships among the users or articles, and single-domain heteromorphic graphs are constructed and obtained by the aid of the semantic relationships among the users or articles in each domain; and designing a cross-domain recommendation system model based on inter-domain and intra-domain relations on the basis of the two different composition graphs, wherein the cross-domain invariance and single-domain particularity of the users or the articles are synchronously learned, and personalized recommendation results for each user are generated through a gating fusion mechanism.
2. The inter-domain and intra-domain relationship-based cross-domain recommendation method of claim 1, specifically comprising:
step 1, data preprocessing: the data of the interaction between the user and the goods, including the scoring and commenting records, namely < user, item, raiting, review >, is primarily processed,
step 2, constructing an isomeric diagram: given two fields A and B, the embodiment establishes a relationship and a relationship matrix by using the scoring and comment data of a user on an article, and constructs a heteromorphic graph on the basis of the relationship and the relationship matrix;
step 3, constructing a cross-domain recommendation model based on inter-domain and intra-domain relations according to the two heterogeneous graphs obtained in the step 2, learning a relation-aware vector expression for each user or article, fusing cross-domain invariance and single-domain specificity through a gating fusion mechanism,
step 4, training a cross-domain recommendation model based on inter-domain and intra-domain relations by adopting a cross entropy loss function;
and 5, calculating the prediction probability between each user and all the articles in the two fields through the trained model, sequencing the articles based on the prediction probabilities, and obtaining the personalized recommendation result for the user, thereby realizing the dual-target recommendation in the two fields.
3. The inter-domain and intra-domain relationship-based cross-domain recommendation method of claim 1, wherein the inter-domain and intra-domain relationship-based cross-domain recommendation model comprises: embedding layer, relation perception map convolution layer, gate-controlled fusion layer and output layer, wherein: the embedding layer adopts a one-hot vector X as a common user U, a field A article I and a field B article J, and converts the high-dimensional one-hot vector into low-dimensional expression; the relation perception graph convolution layer considers different types of relations between nodes and aggregates neighbor nodes of a target nodeMeanwhile, the relation types among the nodes are considered, heterogeneous connectivity between users and object objects in different graphs is captured, and then vector expressions of different layers are spliced to form final vector expressions; gate-controlled fusion layer generating cross-domain invariance of users and items on cross-domain heteromorphic graphs using neural network models: (And) And single domain specificity: ( And) Two gate control units sharing part of parameters are utilized to realize a dual-domain target recommendation task; output layer generating vector representations for end usersAndand vector representation of the final articleAndthe predicted probability between user u and item i or j is:
4. a cross-domain recommendation system based on inter-domain and intra-domain relationships implementing the method of any one of claims 1 to 3, comprising: preprocessing unit, heterogeneous graph building unit, embedding layer, graph convolution layer, gate-controlled fusion layer and output layer, wherein: the preprocessing unit carries out primary processing on the original scoring and comment data to obtain input data; the heterogeneous graph building unit generates a relation matrix reflecting multi-dimensional relations among users, articles and users, and builds two single-field heterogeneous graphs and one cross-field heterogeneous graph on the basis of the relation matrix; the embedding layer converts the high-dimensional heterogeneous graph into a vector of a low-dimensional space; the graph convolution layer considers the relations of different types among the nodes, and utilizes the heterogeneous graph neural network technology to convert the vectors into the vectors of the low-dimensional space for modeling so as to obtain the feature vectors of relation perception; the gating fusion layer carries out deep fusion on the feature vectors and further extracts key features; and the output layer calculates the feature vector expression of the final user and the article according to the key features, and uses an output function as the interactive relation between the user and the article.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210547462.XA CN115168653A (en) | 2022-05-18 | 2022-05-18 | Cross-domain recommendation method and system based on inter-domain and intra-domain relationships |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210547462.XA CN115168653A (en) | 2022-05-18 | 2022-05-18 | Cross-domain recommendation method and system based on inter-domain and intra-domain relationships |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115168653A true CN115168653A (en) | 2022-10-11 |
Family
ID=83483300
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210547462.XA Pending CN115168653A (en) | 2022-05-18 | 2022-05-18 | Cross-domain recommendation method and system based on inter-domain and intra-domain relationships |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115168653A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117648493A (en) * | 2023-12-13 | 2024-03-05 | 南京航空航天大学 | Cross-domain recommendation method based on graph learning |
-
2022
- 2022-05-18 CN CN202210547462.XA patent/CN115168653A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117648493A (en) * | 2023-12-13 | 2024-03-05 | 南京航空航天大学 | Cross-domain recommendation method based on graph learning |
CN117648493B (en) * | 2023-12-13 | 2024-05-31 | 南京航空航天大学 | Cross-domain recommendation method based on graph learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111881350B (en) | Recommendation method and system based on mixed graph structured modeling | |
CN109190030B (en) | Implicit feedback recommendation method fusing node2vec and deep neural network | |
CN109255586B (en) | Online personalized recommendation method for e-government affairs handling | |
CN111737535B (en) | Network characterization learning method based on element structure and graph neural network | |
WO2023065859A1 (en) | Item recommendation method and apparatus, and storage medium | |
CN113378047B (en) | Multi-aspect enhancement-based graph neural network recommendation method | |
CN113378048B (en) | Individualized recommendation method based on multi-view knowledge graph attention network | |
US20220253722A1 (en) | Recommendation system with adaptive thresholds for neighborhood selection | |
CN113918832B (en) | Graph convolution collaborative filtering recommendation system based on social relationship | |
CN116204729B (en) | Cross-domain group intelligent recommendation method based on hypergraph neural network | |
CN113918834B (en) | Graph convolution collaborative filtering recommendation method fusing social relations | |
CN112364257B (en) | Knowledge graph-based user group recommendation system and method | |
CN113434756A (en) | Personalized comment recommendation method based on graph bidirectional aggregation network link prediction model | |
CN112861006A (en) | Recommendation method and system fusing meta-path semantics | |
CN115510319A (en) | Recommendation method and system based on potential interest multi-view fusion | |
CN116501956A (en) | Knowledge perception multi-domain recommendation method and system based on hierarchical graph comparison learning | |
Yang et al. | An improved genetic algorithm and its application in neural network adversarial attack | |
CN114579852A (en) | Recommendation method based on graph convolution network and attention mechanism | |
CN115168653A (en) | Cross-domain recommendation method and system based on inter-domain and intra-domain relationships | |
CN114997476A (en) | Commodity prediction method fusing commodity incidence relation | |
CN114925268A (en) | Recommendation method and system based on graph neural network, electronic device and computer readable medium | |
CN112256918B (en) | Short video click rate prediction method based on multi-mode dynamic routing | |
Chen et al. | Graph enhanced neural interaction model for recommendation | |
CN111815410B (en) | Commodity recommendation method based on selective neighborhood information | |
CN115599990A (en) | Knowledge perception and deep reinforcement learning combined cross-domain recommendation method and system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |