CN117648493B - Cross-domain recommendation method based on graph learning - Google Patents
Cross-domain recommendation method based on graph learning Download PDFInfo
- Publication number
- CN117648493B CN117648493B CN202311712911.2A CN202311712911A CN117648493B CN 117648493 B CN117648493 B CN 117648493B CN 202311712911 A CN202311712911 A CN 202311712911A CN 117648493 B CN117648493 B CN 117648493B
- Authority
- CN
- China
- Prior art keywords
- user
- domain
- representation
- learning
- item
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 41
- 230000008569 process Effects 0.000 claims abstract description 13
- 238000012512 characterization method Methods 0.000 claims abstract description 11
- 238000013528 artificial neural network Methods 0.000 claims abstract description 10
- 238000013508 migration Methods 0.000 claims abstract description 5
- 230000005012 migration Effects 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 30
- 230000003993 interaction Effects 0.000 claims description 15
- 238000013461 design Methods 0.000 claims description 8
- 230000004927 fusion Effects 0.000 claims description 8
- 239000011159 matrix material Substances 0.000 claims description 7
- 230000007246 mechanism Effects 0.000 claims description 7
- 230000000052 comparative effect Effects 0.000 claims description 6
- 238000012549 training Methods 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 5
- 230000002708 enhancing effect Effects 0.000 claims description 4
- 230000002452 interceptive effect Effects 0.000 claims description 4
- 230000008901 benefit Effects 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 238000012544 monitoring process Methods 0.000 claims description 3
- 230000008447 perception Effects 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 claims description 2
- 230000004931 aggregating effect Effects 0.000 claims 1
- 230000002195 synergetic effect Effects 0.000 abstract description 3
- 238000005516 engineering process Methods 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000003623 enhancer Substances 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000004806 packaging method and process Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000000844 transformation Methods 0.000 description 1
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
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/09—Supervised learning
-
- 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
- G06N3/096—Transfer learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Databases & Information Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses a cross-domain recommendation method based on graph learning. The method reasonably judges the effectiveness and feasibility of the proposed cross-domain recommendation method by predicting the probability of a given user item pair. Mainly comprises three parts: the first part precisely separates domain-invariant features and domain-specific features of the domain in a spatially mapped manner, and aggregates user-project high-order synergistic relationship information using a lightweight graph convolution neural network to pre-enhance the information content of the disentangled tokens. The second part considers that the information quantity of the characterization after the disentanglement is further assisted and enhanced through the heterogeneous graph neural network by means of the social relationship and the project dependency relationship of the user, and meanwhile, personalized knowledge migration of the user and the project embedding is carried out through the meta network, so that the heterogeneous relationship learning is realized. The third part builds two contrast learning tasks, one is to monitor the disentanglement process, so that the disentanglement process can obtain embedded characterization with low redundancy and light noise to a certain degree in the initial stage, and the other is to enhance the robustness of self-enhanced heterogeneous relation learning. The invention combines the related technologies such as graph learning, contrast learning and the like, effectively utilizes the side information to realize user-project characterization modeling, and improves the performance of personalized cross-domain recommendation.
Description
Technical Field
The invention belongs to the technical field of cross-domain recommendation systems, and particularly relates to a cross-domain recommendation method based on graph learning.
Background
The challenge of data sparsity has been a long-felt problem in Recommendation Systems (RSs). Cross-domain recommendations (CDRs) facilitate seamless migration of knowledge from source domain to target domain to address this challenge and improve personalized recommendations. The existing CDR method mainly focuses on the scene of user overlapping, and realizes knowledge transfer based on shared user representation. However, these approaches take into account the consistency of user interests, ignoring the multi-dimensional differences in user preferences across domains. Based on this, recent studies have attempted to generate domain-invariant features and domain-specific features. Nevertheless, identifying and packaging disparate semantic data in complex data sets presents certain challenges. Thus, the challenge of building a representation with rich multi-dimensional semantics to mitigate data sparsity arises. In reality, the social network association structure between users presents dynamic features. Meanwhile, mutual dependency exists between items. Therefore, integrating heterogeneous information to help enrich the information content of domain-invariant and domain-specific representations, thereby improving the performance of the recommendation system, is a solution worthy of research.
In recent years, graph Neural Networks (GNNs) have achieved significant success in feature extraction in non-euclidean space. This motivated the researchers' search for heterographic neural networks (Hetero-GNNs). Their goal is to smartly integrate rich semantic information from heterogeneous relationships into potential tokens. However, due to the sparsity of data, contemporary Hetero-GNNs have difficulty achieving high quality user/item embedding. Fortunately, contrast Learning (CL) was proposed. It is particularly suitable for alleviating the problem of data scarcity because it can self-enhance unlabeled data. Subsequently, one notable advance is the integration of CL with GNN. This synergy is known as Graph Contrast Learning (GCL). It may enhance the robustness of the representation learning in the graph structure without enough observation of the labels. Therefore, inspired from previous GCL studies, the present invention proposes hetero-relationship contrast learning. The invention aims to enrich the representation information quantity after the entanglement and enhance the individuation performance of cross-domain recommendation.
The invention obtains light noise, low redundancy and high richness depth representation by separating domain invariant features from domain specific features while embedding heterogeneous semantic relationships therein. Based on this, the present invention needs to face the problems that: (1) How to accurately separate domain-invariant features from domain-specific features in complex user-project interactions. (2) How to integrate the heterogeneous semantic relationships to enrich the information content of the disentangled characterization. (3) How to reduce noise and redundancy in the disentangled user and project representations.
Disclosure of Invention
The invention aims at the problems, and designs and invents a cross-domain recommendation method based on graph learning. Specifically, the present invention exploits a feature abstraction component to separate domain-invariant and domain-specific representations while capturing high-level collaboration information from a user-project interaction graph. The invention designs a heterogeneous representation learning component. The component employs a Hetero-GNN as an encoder to preserve rich semantic information related to heterogeneous relationships. Furthermore, a self-supervising presentation enhancement component is proposed that comprises two CL paradigms. The invention aims to realize personalized knowledge transfer through self-adaptive contrast enhancement while supervising the entanglement.
In order to achieve the above purpose, the invention adopts the following technical scheme:
A cross-domain recommendation method based on graph learning comprises the following steps:
Step 1: first, abstracting the user and item representation of each domain to domain-invariant and domain-specific space, initializing the disentangled embedded token. A lightweight graph convolution neural network is used in the interaction graph to aggregate user-project high-order synergistic relationship information to pre-enhance the amount of information represented by the user and project after de-entanglement.
In said step 1, the invention first introduces user-item pairs as input. By sampling from a gaussian normal distribution, two independent embedding matrices are randomly generatedAnd/>To generate initial embedding/>Representing the initial embedded representations of the generated user and item, respectively.
Second, the present invention separates the domain-invariant representation and domain-specific representation of users and items in a spatially mapped manner. In particular, the invention willAnd/>Abstract into two independent subspaces, namely a domain-invariant space x and a domain-specific space y. Each user/>The representation of (2) will be represented by/>Two-part composition, representing the domain-invariant and domain-specific representation of item u, respectively, while item/>And the same is true. Next, it is considered to enrich the de-entangled representation in advance in the interaction map using higher-order collaboration information. The invention fuses the learned characteristics of each layer of graph after convolution, captures multi-order neighbor semantics, and defines a fusion function as follows:
Next, it is considered to enrich the de-entangled representation in advance in the interaction map using higher-order collaboration information. Taking the field-invariant as an example, in the present invention And/>Can be abstracted into:
wherein, Is a mean function,/>Representing the join operation, the enrichment of user and item representations for a particular domain may be performed similarly. The present invention connects enhanced domain-invariant and domain-specific user and item representations to achieve a more comprehensive embedded representation. The specific operations can be summarized as follows:
Where || denotes a connection operation. This enrichment process is the same for D (s) and D (t). And/>As an initial embedded representation of the next-to-be-implemented personalized rich user and item and.
And 2, obtaining embedding for modeling social relationships and project dependency relationships of users from the disentangled domain space, fusing relationship perception semantics into disentangled user and project representations through a heterogeneous graph neural network, and meanwhile, carrying out personalized knowledge migration of the users and the projects through a meta network to realize heterogeneous relationship learning.
In the step2, as heterogeneous information contains rich semantic information, the user and the item representation are further enriched by combining the advantage. An initial embedding for modeling user social relationships and project dependencies is first derived from the above-described embedding.Representing the user-user, item-item, and relationship representation between user-item, respectively, the initial representation derived by this mechanism is heterogeneous in user and item relationship modeling and is related to/>Share the same semantics. The embedded characterization is again learned through the heterographing convolutional neural network.
Inspired by the soft-element path design, the information in each iteration is aggregated from heterogeneous relationships. Through multiple iterations of heterogeneous message propagation, the high-order embedding preserves the heterogeneous semantics of the multi-hop connection. In particular, the embedding of users and items is updated by custom heterogeneous fusion programs.
Wherein,Is a mean function,/>Representing the connection operation. /(I)Is the end user and project representation after fusing heterogeneous side information, and similar steps are adopted to obtain corresponding/>
To generate a personalized mapping of the auxiliary view to the user-item interaction code of each user and each item, first the meta-knowledge is extractedImportant features of the user and the item are preserved in the auxiliary view and the interactive view. To facilitate user and project embedded personalized transformations, multiple Layer Perceptron (MLP) is used to transform meta-knowledge representations into weight matricesTo obtain a more comprehensive representation, the present invention uses weighted summation to integrate the user and item representations. /(I)Representing the final embedding for recommending the primary task, this series of processes is the same for spaces y and D (t).
And 3, constructing two CL tasks. The MI mutual information maximization mechanism is based on a contrast learning target design, and in order to monitor the disentanglement process, the embedded characterization with low redundancy and light noise can be obtained to a certain degree in the initial stage. A contrast learning objective is designed based on infoNCE penalty functions in order to enhance the robustness of learning with self-enhanced heterogeneous relationships.
In the step 3, two comparison learning tasks are constructed. Under the influence of the maximum Mutual Information (MI), the invention hopes to mitigate the influence of noise and redundancy, and the comparative learning targets based on the design are as follows:
where c (-) is a cosine similarity function and τ is a parameter that controls the smoothness of the softmax curve. The inspiration is obtained from the success of the contrast learning of infoNCE loss functions, enhancing the user-project representation learning in the model. This procedure is as follows:
wherein, Respectively matrix/>Is used to determine the embedded vector of (a). g (-) represents a similarity function,/>Is a super parameter of the softmax temperature. Based on the above method, the comparison learning objectives of the project, the space y and the D (t) can be obtained as well. Overall comparative loss/>Is a weighted sum of all losses.
Step 4, the invention connects the domain-invariant representation and domain-specific representation of the user and item, respectively, to generate a final representation and predict the probability of a given user item pair. Top-k recommendations with implicit feedback for all users in each domain are considered given the two domains D (s) and D (t) to improve the recommendation performance of both domains simultaneously.
In said step 4, the invention predicts the likelihood of user-item interactions in D (s) by dot product. In the optimization stage, the invention follows a classical ordering model, and utilizes Bayesian personalized ordering (BPR) recommended loss as main task loss to further optimize parameters. The present invention proposes a training strategy paradigm to better learn information from different angles and further optimize the model. The total training loss is as follows:
Where Ω is a model parameter that can be learned, η 1 and η 2 are weights for controlling the self-monitoring signal and regularization term.
The cross-domain recommendation method based on graph learning provided by the invention has the beneficial effects that:
(1) The invention provides a cross-domain recommendation characterization learning strategy. It emphasizes the key distinction between domain-invariant features and domain-specific features and integrates heterogeneous relationship semantics into the de-entangled representation. This enhances personalized semantic interactions between the user and the item in the recommendation.
(2) The invention designs two comparison learning tasks. Firstly, generating a supervision signal based on MI (MI) for a disentanglement process; another is to enhance the isopgram comparison learning with a meta-network to enable customized knowledge extraction from users and projects.
(3) The invention provides a cross-domain recommendation solution, which improves the cross-domain recommendation performance by effectively utilizing graph learning knowledge.
Drawings
For a clearer description of embodiments of the present application or technical solutions in the prior art, the drawings that are needed in the embodiments will be briefly described, and it is obvious that the drawings in the following description are only some embodiments described in the present application, and other drawings may be obtained according to these drawings for a person having ordinary skill in the art.
FIG. 1 is a schematic diagram of the overall structure of a model according to the present invention;
FIG. 2 is a schematic diagram of a generic isomerism diagram construction in accordance with the present invention;
FIG. 3 is a schematic diagram of a general comparative learning process according to the present invention.
Detailed Description
In order to better understand the above technical solutions, the following further details of the technical solutions of the present invention are described with reference to the accompanying drawings and examples. The relevant experiments of the present example are based on the programming language Python 3.8 version and Pytorch deep learning framework. All experiments were run on NVIDIAGPU T4. And carrying out parameter learning optimization on the model by utilizing Adam.
The cross-domain recommendation method based on graph learning provided by the invention mainly comprises three parts: domain feature abstraction, heterogeneous representation learning, and self-supervised characterization enhancement. Without loss of generality, the present invention details modeling with domain D (s) as an example.
Step 1, a domain feature abstract module. First, abstracting the user and item representation of each domain to domain-invariant and domain-specific space, initializing the disentangled embedded token. A lightweight graph convolution neural network is used in the interaction graph to aggregate user-project high-order synergistic relationship information to pre-enhance the amount of information represented by the user and project after de-entanglement.
(1) De-entangled representation generator
The module separates the domain-invariant representation from the domain-specific representation. The present invention first introduces user-item pairs (u, i (s)) and (u, i (t)) as inputs. Then, two independent embedding matrices are randomly generatedAnd/>As an initial embedding, by sampling from a gaussian normal distribution:
Where α represents a gaussian normal distribution with a mean value of 0 and a standard deviation of 0.1. h u and Is a single thermal encoding of users and items,/>And/>Representing the initial embedded representations of the generated user and item, respectively.
Based on the embedding, the present invention separates the domain-invariant representation and domain-specific representation of users and items in a spatially mapped manner. In particular, the invention willAnd/>Abstractions are two independent subspaces:
Where x, y represents the dimension space of the shared and specific representation, σ (·) is the activation function, Θ x(～),Θy (-) is the projection space mapping function. Each user The representation of (2) will be represented by/>Two-part composition, representing the domain-invariant and domain-specific representation of item u, respectively, while item/>And the same is true.
(2) Collaboration enhancer
Since recent studies have demonstrated the effectiveness of LightGCN, the present invention contemplates the use of higher order collaboration information in interactive maps to pre-enrich the amount of representation information after de-entanglement. Taking the field-invariant as an example, in the present inventionAnd/>Can be abstracted into:
The invention fuses the learned characteristics of each layer of graph after convolution, captures multi-order neighbor semantics, and defines a fusion function as follows:
wherein, Is a mean function,/>Representing the join operation, the enrichment of user and item representations for a particular domain may be performed similarly. The enhanced domain invariance and domain specific user and item representations are connected to achieve a more comprehensive embedded representation. The specific operations can be summarized as follows:
Where || denotes a connection operation. This enrichment process is the same for D (s) and D (t). And/>As an initial embedded representation of the next-to-be-implemented personalized rich user and item and.
And 2, heterogeneous representation learning module. Embedding for modeling user social relations and project dependency relations is obtained from the disentangled domain space, relation perception semantics are fused into disentangled user and project representations through a heterogeneous graph neural network, personalized knowledge migration of the user and the project is carried out through a meta network, and heterogeneous relation learning is achieved.
(1) Intelligent relational awareness representation
Because heterogeneous information contains rich semantic information, the present invention contemplates further enriching user and project representations in combination with this advantage. The main function of this module is to obtain an embedding for modeling user social relationships and project dependencies from the disentangled domain space. The specific method comprises the following steps:
wherein, Representing element level multiplication, delta represents federation, sig (-) is an activation function.Representing the weight matrix generated by the Xavier normal distribution. /(I)Representing the user-user, item-item, and relationship representation between user-item, respectively, the initial representation derived by this mechanism is heterogeneous in user and item relationship modeling and is related to/>Share the same semantics.
(2) Domain aware heterogeneous information fusion
In combination with the previously generated relational awareness embeddings, the module again utilizes LightGCN to incorporate different information into the domain-invariant and domain-specific representations, further improving the quality of the representation. To be used forThe following are examples:
similarly, the user-user embedding and project-project embedding relationships also iterate in accordance with the same GCN model.
Inspiration is drawn from the soft-element path, and heterogeneous side information is enhanced by using dedicated aggregation functions. To further integrate the different information types, and incorporate coding layer specific representations. In particular, the embedding of users and items is updated by a heterogeneous fusion procedure defined as follows:
wherein, And integrating heterogeneous relation semantics to become the input of the next layer. Phi (·) represents the L2 normalization operation, and f up (-) is an information fusion function. /(I)Is the end user and project representation after fusing heterogeneous side information, and similar steps are adopted to obtain corresponding/>
(3) Knowledge element optimization
And personalized learning is facilitated, and knowledge information is transmitted from the side face. The invention introduces a new method for extracting heterogeneous side information element path knowledge through a meta-network attention mechanism, and the specific formulas are summarized as follows:
wherein, The meta-knowledge representation of the user and the item, respectively, f meta (to) is a meta-knowledge extraction function. The auxiliary representation effectively captures the category dependencies of the items based on the time stamps and trust networks between users.
To facilitate personalized conversion of user and project embeddings, a multi-layer perceptron (MLP) is used to convert the meta-knowledge representation into a weight matrix. The specific procedure can be summarized as follows:
wherein f mlp (-) is a meta-knowledge learner consisting of two linear transformation layers, with PReLU activation function. so (-) is a softmax function. And/>Is embedded for personalizing the converted user and item. To obtain a more comprehensive representation, a weighted sum is used to integrate the user and item representations:
wherein λ, β ε R control the weights between embeddings and the sum is 1. Representing the final embedding for recommending the primary task, this series of processes is the same for spaces y and D (t).
And 3, enhancing self-supervision characterization. Two CL tasks are constructed, a comparison learning target is designed based on an MI mutual information maximization mechanism, and in order to supervise the disentanglement process, the embedded characterization with low redundancy and light noise can be obtained to a certain degree in the initial stage. A contrast learning objective is designed based on infoNCE penalty functions in order to enhance the robustness of learning with self-enhanced heterogeneous relationships.
(1) Mutual information driven contrast learning
The invention creates one positive sample and three negative samples. It is desirable to minimize dissimilarity between positive samples, subject to the greatest mutual information. At the same time, dissimilarity between negative samples is maximized to mitigate the effects of noise and redundancy. The contrast learning targets designed by the invention are as follows:
where c (-) is a cosine similarity function and τ is a parameter that controls the smoothness of the softmax curve.
(2) Heterogeneous relationship contrast learning
And generating two sets of embeddings by adopting heterogeneous graph contrast learning and meta-path personalized conversion. The inspiration is obtained from the recent success of contrast learning based on infoNCE loss functions, enhancing user-project representation learning in the model. This procedure is as follows:
wherein, Respectively matrix/>Is used to determine the embedded vector of (a). g (-) represents a similarity function,/>Is a super parameter of the softmax temperature. Based on the method, the comparison learning targets of the project, the space y and the space D (t) can be obtained. Overall comparative loss/>Is a weighted sum of all contrast losses.
And 4, a joint learning module. With the above modules, it is an object of the present invention to achieve a more comprehensive embedded representation by smartly connecting enhanced domain-invariant representations with domain-specific user and item representations. The specific operations can be summarized as follows:
The model of the present invention predicts the likelihood of user-item interactions in D (s) by dot product: Wherein/> Is an embedded vector in a representative matrix. In the optimization stage, the invention follows a classical ordering model, and uses Bayesian personalized ordering (BPR) recommended loss as main task loss to further optimize parameters:
wherein, For a positive set of interactions with the user in all samples,/>And vice versa. D (t) estimates the corresponding BPR loss as/>
In combination with the above loss function, the present invention proposes a training strategy paradigm to better learn information from different angles and further optimize the model. The total training loss is as follows:
Where Ω is a model parameter that can be learned, η 1 and η 2 are weights for controlling the self-monitoring signal and regularization term.
In a recommended task, the embodiment of the invention evaluates the model by using NDCG@K and HR@K, and for three sets of real comment data sets from the Amazon data set, the invention first keeps a valid sample with a score greater than 4, wherein each entry is marked as 0 or 1, indicating whether the user has scored the item. Then, the present invention preprocesses a data set by selecting users having interactions more than 5 times in each domain, and extracts overlapping users in two domains while limiting each domain to 10000 entries or less. In addition, the invention also performs data filtering to preserve user votes and project timestamps. Experimental results show that the model of the embodiment of the invention has excellent performance on all data sets and indexes, and the recommended accuracy is improved.
The foregoing is merely a preferred embodiment of the present invention and it should be noted that modifications and adaptations to those skilled in the art may be made without departing from the principles of the present invention, which are intended to be comprehended within the scope of the present invention.
Claims (3)
1. A cross-domain recommendation method based on graph learning aims at learning new users and project embedded representations so as to improve recommendation performance of two domains at the same time; the method is characterized by comprising the following steps of:
Step 1, abstracting user and item representations of each field to a field-invariant and field-specific space, and initializing an embedded representation of disentanglement; aggregating user-project high-order cooperative relation information in an interactive graph by using a lightweight graph convolution neural network so as to pre-enhance the information quantity of the user and project representation after disentanglement;
Step 2, obtaining embedding for modeling user social relations and project dependency relations from the disentangled domain space, fusing relation perception semantics into disentangled user and project representations through a heterogeneous graph neural network, and meanwhile, carrying out personalized knowledge migration of the user and the project through a meta network to realize heterogeneous relation learning;
Step 3, constructing two contrast learning CL tasks; the method is characterized in that a comparison learning target is designed based on an MI mutual information maximization mechanism, so that the embedded representation of low redundancy and light noise can be obtained to a certain extent in an initial stage in order to supervise the disentanglement process; designing a contrast learning target based on infoNCE loss functions in order to enhance the robustness of learning with self-enhanced heterogeneous relationships;
Step 4, respectively connecting the domain-invariant representation and the domain-specific representation of the user and the project to generate a final representation, and predicting the probability of a given user item pair; considering Top-k recommendation with implicit feedback for all users in each domain given two domains D (s) and D (t) to improve recommendation performance for both domains simultaneously;
In said step 1, user-item pairs are introduced as input for each domain; by sampling from a gaussian normal distribution, two independent embedding matrices are randomly generated And/>To generate initial embedding/>Initial embedded representations of the generated user and item, respectively;
separating the domain-invariant representation and domain-specific representation of the user and the item in a spatially mapped manner based on the embedding; specifically, it will And/>Respectively abstracting into two independent subspaces, namely a domain-invariant space x and a domain-specific space y; each user/>The representation of (2) will be represented by/>Two-part composition, representing the domain-invariant and domain-specific representation of item u, respectively, while item/>And the same is done; next, consider pre-enriching the de-entangled representation with higher order collaboration information in an interaction graph; fusing the learned characteristics after each layer of graph convolution, capturing multi-order neighbor semantics, and defining a fusion function as follows:
wherein, Is a mean function,/>Representing a join operation, the user of a particular domain and the enrichment of the item representation are performed similarly; connecting the enhanced domain-invariant and domain-specific user and item representations to achieve a more comprehensive embedded representation; the specific operations are summarized as follows:
Wherein || represents a connection operation; this enrichment process is the same for D (s) and D (t); and/> As an initial embedded representation of the personalized rich user and item and to be achieved next;
In the step2, as heterogeneous information contains rich semantic information, users and project representations are further enriched by considering the combination of the advantages; firstly, deriving initial embedding for modeling user social relations and project dependency relations from the embedding; representing user-user, item-item relationship representations,/>, respectively Representing a representation of the relationship between user-items, the initial representation derived by this mechanism has heterogeneity in both user and item relationship modeling, and is consistent with/>Share the same semantics; the embedded characterization is learned again through the heterographing convolutional neural network;
Inspired by soft element path design, the information in each iteration is aggregated from heterogeneous relations; through multiple iterations of heterogeneous message propagation, the high-order embedding maintains heterogeneous semantics of multi-hop connection; in particular, the embedding of users and items is updated by a custom heterogeneous fusion program;
wherein, Is a mean function,/>Representing a join operation; /(I)Is the end user and project representation after fusing heterogeneous side information, and similar steps are adopted to obtain corresponding/>
To generate a personalized mapping of the auxiliary view to the user-item interaction code of each user and each item, first the meta-knowledge is extractedPreserving important features of users and items in the auxiliary view and the interactive view; to facilitate personalized conversion of user and project embedding, the meta-knowledge representation is converted into a weight matrix using a multi-layer perceptron MLPTo obtain a more comprehensive representation, a weighted sum is used to integrate the user and item representations; /(I)Representing the final embedding for recommending the primary task, this series of processes is the same for spaces y and D (t).
2. The cross-domain recommendation method based on graph learning of claim 1, wherein the method comprises the following steps: in the step 3, two comparison learning tasks are constructed; under the influence of the maximum mutual information MI, it is desirable to mitigate the influence of noise and redundancy, and the comparative learning objective based on this design is as follows:
Wherein c (to) is a cosine similarity function, and τ is a parameter for controlling smoothness of a softmax curve; obtaining inspiration from success of contrast learning of infoNCE loss functions, enhancing user-project representation learning in the model; this procedure is as follows:
wherein, Respectively matrix/>Is a vector of embedding; g (-) represents a similarity function,/>Is a super parameter of the softmax temperature; based on the method, the comparison learning targets of the project, the space y and the D (t) are obtained; overall comparative loss/>Is a weighted sum of all losses.
3. The cross-domain recommendation method based on graph learning of claim 1, wherein the method comprises the following steps: in said step 4, predicting the likelihood of user-item interactions in D (s) by dot product; in the optimization stage, following a classical ordering model, and further optimizing parameters by using Bayesian personalized ordering BPR recommendation loss as main task loss; a training strategy paradigm is proposed to better learn information from different angles and further optimize the model; the total training loss is as follows:
Wherein Ω is a model parameter that can be learned, η 1 and η 2 are weights for controlling the self-monitoring signal and regularization term; Indicating the corresponding BPR loss for D (t).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311712911.2A CN117648493B (en) | 2023-12-13 | 2023-12-13 | Cross-domain recommendation method based on graph learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311712911.2A CN117648493B (en) | 2023-12-13 | 2023-12-13 | Cross-domain recommendation method based on graph learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117648493A CN117648493A (en) | 2024-03-05 |
| CN117648493B true CN117648493B (en) | 2024-05-31 |
Family
ID=90049357
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311712911.2A Active CN117648493B (en) | 2023-12-13 | 2023-12-13 | Cross-domain recommendation method based on graph learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117648493B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118094018B (en) * | 2024-04-26 | 2024-07-30 | 安徽思高智能科技有限公司 | Enhanced graph contrast learning recommendation system method, device and storage medium |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1750221A1 (en) * | 2005-07-14 | 2007-02-07 | The Boeing Company | System, method, and computer program to predict the likelihood, the extent, and the time of an event or change occurrence using a combination of cognitive causal models with reasoning and text processing for knowledge driven decision support |
| CN108022161A (en) * | 2017-12-26 | 2018-05-11 | 河北中晟易通科技有限公司 | Clothing matching commending system based on image recognition and big data analysis |
| CN112541132A (en) * | 2020-12-23 | 2021-03-23 | 北京交通大学 | Cross-domain recommendation method based on multi-view knowledge representation |
| CN115168653A (en) * | 2022-05-18 | 2022-10-11 | 上海交通大学 | Cross-domain recommendation method and system based on inter-domain and intra-domain relationships |
| CN115982480A (en) * | 2023-02-13 | 2023-04-18 | 山东师范大学 | Sequence recommendation method and system based on cooperative attention network and comparative learning |
| CN116010708A (en) * | 2023-02-13 | 2023-04-25 | 重庆市知识产权保护中心(重庆摩托车(汽车)知识产权信息中心) | Multi-comparison learning recommendation method based on knowledge graph |
| CN116468720A (en) * | 2023-05-09 | 2023-07-21 | 南京航空航天大学 | Medical image multi-target domain field self-adaption method based on prompt learning |
| CN116501956A (en) * | 2023-04-03 | 2023-07-28 | 清华大学 | Knowledge perception multi-domain recommendation method and system based on hierarchical graph comparison learning |
| CN116610874A (en) * | 2023-06-06 | 2023-08-18 | 哈尔滨理工大学 | Cross-domain recommendation method based on knowledge graph and graph neural network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112001385B (en) * | 2020-08-20 | 2024-02-06 | 长安大学 | Target cross-domain detection and understanding method, system, equipment and storage medium |
-
2023
- 2023-12-13 CN CN202311712911.2A patent/CN117648493B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1750221A1 (en) * | 2005-07-14 | 2007-02-07 | The Boeing Company | System, method, and computer program to predict the likelihood, the extent, and the time of an event or change occurrence using a combination of cognitive causal models with reasoning and text processing for knowledge driven decision support |
| CN108022161A (en) * | 2017-12-26 | 2018-05-11 | 河北中晟易通科技有限公司 | Clothing matching commending system based on image recognition and big data analysis |
| CN112541132A (en) * | 2020-12-23 | 2021-03-23 | 北京交通大学 | Cross-domain recommendation method based on multi-view knowledge representation |
| CN115168653A (en) * | 2022-05-18 | 2022-10-11 | 上海交通大学 | Cross-domain recommendation method and system based on inter-domain and intra-domain relationships |
| CN115982480A (en) * | 2023-02-13 | 2023-04-18 | 山东师范大学 | Sequence recommendation method and system based on cooperative attention network and comparative learning |
| CN116010708A (en) * | 2023-02-13 | 2023-04-25 | 重庆市知识产权保护中心(重庆摩托车(汽车)知识产权信息中心) | Multi-comparison learning recommendation method based on knowledge graph |
| CN116501956A (en) * | 2023-04-03 | 2023-07-28 | 清华大学 | Knowledge perception multi-domain recommendation method and system based on hierarchical graph comparison learning |
| CN116468720A (en) * | 2023-05-09 | 2023-07-21 | 南京航空航天大学 | Medical image multi-target domain field self-adaption method based on prompt learning |
| CN116610874A (en) * | 2023-06-06 | 2023-08-18 | 哈尔滨理工大学 | Cross-domain recommendation method based on knowledge graph and graph neural network |
Non-Patent Citations (8)
| Title |
|---|
| Multi-graph Convolutional Feature Transfer for Cross-domain Recommendation;Yanling Zhang等;2022 International Joint Conference on Neural Networks (IJCNN);20220930;1-8 * |
| Self-Supervised Dynamic Hypergraph Recommendation based on Hyper-Relational Knowledge Graph;Yi Liu等;Birmingham, United Kingdom;20230831;1-10 * |
| 基于图神经网络的推荐模型研究;林伟业;中国优秀硕士学位论文全文数据库信息科技辑;20230215(第2期);I138-3985 * |
| 基于深度学习的域适应方法综述;田青等;数据采集与处理;20220515;512-541 * |
| 基于深度迁移学习的跨领域细粒度情感分析;曹子旋;中国优秀硕士学位论文全文数据库信息科技辑;20200815(第8期);I138-795 * |
| 基于自监督图掩码神经网络的社交推荐模型;臧秀波等;模式识别与人工智能;20231015;第36卷(第10期);942-952 * |
| 海量数据下广义线性模型算法研究;陈少东;中国优秀硕士学位论文全文数据库基础科学辑;20220815(第8期);A002-60 * |
| 融合零样本学习和小样本学习的弱监督学习方法综述;潘崇煜;黄健;郝建国;龚建兴;张中杰;;系统工程与电子技术;20200916(10);104-114 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117648493A (en) | 2024-03-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Darban et al. | GHRS: Graph-based hybrid recommendation system with application to movie recommendation | |
| Hui et al. | Personalized recommendation system based on knowledge embedding and historical behavior | |
| Wang et al. | Knowledge graph embedding via graph attenuated attention networks | |
| CN112732936B (en) | Radio and television program recommendation method based on knowledge graph and user microscopic behaviors | |
| CN117648493B (en) | Cross-domain recommendation method based on graph learning | |
| CN116431914A (en) | Cross-domain recommendation method and system based on personalized preference transfer model | |
| CN114461907A (en) | Knowledge graph-based multi-element environment perception recommendation method and system | |
| Wang et al. | An enhanced multi-modal recommendation based on alternate training with knowledge graph representation | |
| Sakurai et al. | Deep reinforcement learning-based music recommendation with knowledge graph using acoustic features | |
| Xie et al. | Unsupervised learning of paragraph embeddings for context-aware recommendation | |
| Ali et al. | Dynamic context management in context-aware recommender systems | |
| CN116433800B (en) | Image generation method based on social scene user preference and text joint guidance | |
| CN117235375A (en) | User multi-behavior recommendation method based on graphic neural network and element learning | |
| CN117056609A (en) | Session recommendation method based on multi-layer aggregation enhanced contrast learning | |
| Ma et al. | Knowledge graph based recommendation algorithm for educational resource | |
| CN116662601A (en) | Song recommendation method based on graphic neural network and knowledge graph | |
| CN114048396B (en) | NQI service recommendation method and system based on heterogeneous information network and deep learning | |
| Li et al. | Semi-siamese network for content-based video relevance prediction | |
| Wang et al. | A Tri‐Attention Neural Network Model‐BasedRecommendation | |
| Hoxha | Cross-domain recommendations based on semantically-enhanced User Web Behavior | |
| Zeng et al. | CKEN: collaborative knowledge-aware enhanced network for recommender systems | |
| Suo et al. | Entity Recommendation Via Integrating Multiple Types of Implicit Feedback in Heterogeneous Information Network | |
| Wang et al. | Interactive genetic algorithm joining recommender system | |
| Zhang et al. | A Multi-classification Division-aggregation Framework for Fake News Detection | |
| Zheng et al. | Routing User-Interest Markov Tree for Scalable Personalized Knowledge-Aware Recommendation |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |