CN115545833A - Recommendation method and system based on user social information - Google Patents
Recommendation method and system based on user social information Download PDFInfo
- Publication number
- CN115545833A CN115545833A CN202211233278.4A CN202211233278A CN115545833A CN 115545833 A CN115545833 A CN 115545833A CN 202211233278 A CN202211233278 A CN 202211233278A CN 115545833 A CN115545833 A CN 115545833A
- Authority
- CN
- China
- Prior art keywords
- user
- interaction
- embedded representation
- representation
- type
- 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
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- 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
- 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
- G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
Abstract
The invention discloses a recommendation method and a system based on user social information, the method includes integrating user representation and item representation with time perception information into a relation perception dependency network of multi-interaction type user-item interaction, modeling a multi-type interaction behavior mode by using the relation perception dependency network to obtain user embedded representation and item embedded representation containing high-order heterogeneous multi-type dynamic interaction, extracting high-order multi-interaction perception collaborative semantic information, meanwhile, enhancing the user embedded representation by using a graph attention machine system and a mutual information learning mode based on a user social graph to obtain user enhanced embedded representation containing the user social information, thereby realizing fine-grained modeling of a recommendation model based on the user social information, and finally, performing dot product operation on the user embedded representation and the item embedded representation after the user embedded representation and the user enhanced embedded representation are integrated to formulate a recommendation list. According to the scheme, the recommendation performance and the recommendation accuracy are improved.
Description
Technical Field
The invention relates to the technical field of natural language processing and deep learning, in particular to a recommendation method and a recommendation system based on user social information.
Background
The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.
The recommendation system is a simulation of certain behaviors of people, analyzes and processes specific data information through a recommendation algorithm, and then recommends the processed result to users with related requirements, and is a new research field formed by combining multiple disciplines such as data mining, prediction algorithm, machine learning and the like. Current conventional recommendation systems include content filtering based recommendations, collaborative filtering based recommendations, and hybrid recommendations. The recommendation system based on content filtering is characterized in that a user history selection record or a preference record is used as a reference recommendation, and an item with high relevance to the reference recommendation in other unknown records is mined to serve as a system recommendation content, and the recommendation system mainly comprises a data mining processing part and an adaptive recommendation part, however, the recommendation system based on content filtering is easy to ignore the typicality of a recommendation object and is poor in safety; recommendation based on collaborative filtering usually obtains the dependency relationship between users and items through the scoring of the users on the items, further predicts the association relationship between new users and items, and mainly comprises recommendation based on a memory and recommendation based on a model, however, recommendation based on collaborative filtering has a cold start problem and cannot process recommendation with complex operation, and is lack of interpretability; the hybrid recommendation technology is a recommendation mode which retains the advantages of different recommendation technologies and avoids the disadvantages of the different recommendation technologies, different algorithms are integrated into a recommendation system, namely hybrid recommendation, however, the hybrid recommendation lacks an efficient hybrid mode, and the recommendation process is complex.
Therefore, in the prior art, a deep learning technology is applied to a recommendation system, and the deep learning technology not only can find potential feature representations hidden in user behavior records, but also can capture interactive features of nonlinear relations between users, between users and items, between items, and between items, so that more opportunities are brought to performance (such as recall rate, precision and the like) improvement of the system, and more accurate recommendation can be realized.
On the basis, with the increasing popularization of the network community service platform, considering that the interest and preference of the user are influenced by people in the social circle, such as parents, classmates, colleagues and the like, the rich social network information of the user is applied to the recommendation system, and the recommendation performance can be improved.
However, the inventor finds that although a model utilizing social information of a user is already applied to a recommendation system, the existing recommendation system is difficult to overcome the problems of data sparsity and cold start, and the task of completing recommendation by using social information of the user still has the problems of not fully utilizing various types of interaction information, time series information, global social context and the like of the user, so that the recommendation performance of the final recommendation system is poor.
Disclosure of Invention
In order to solve the defects of the prior art, the invention provides a recommendation method and a recommendation system based on user social information, which are characterized in that user representation and item representation with time perception information are merged into a multi-interaction type user-item interaction relation perception dependency network to generate user embedded representation and item embedded representation containing high-order heterogeneous multi-type dynamic interaction, high-order multi-interaction perception collaborative semantic information is extracted, heterogeneous advantages of user social relation are obtained by using an attention mechanism, and local-global context semantic information of a user is obtained by using an inter-information structural model, so that fine-grained modeling of a recommendation model based on the user social information is realized, and recommendation performance and recommendation accuracy are improved.
In a first aspect, the present disclosure provides a recommendation method based on user social information, including the following steps:
acquiring time perception information when multi-type interaction is carried out between a user and a project based on a multi-interaction type user-project interaction diagram;
modeling a multi-type interaction behavior mode by utilizing a multi-interaction type user-item interaction relation perception dependency network according to user representation, item representation and time perception information during user-item interaction to obtain user embedded representation and item embedded representation containing high-order heterogeneous multi-type dynamic interaction;
based on the user social graph, enhancing the user embedded representation by using a graph attention machine mechanism and a mutual information learning mode to obtain the user enhanced embedded representation containing the user social information;
and fusing the user embedded representation and the user enhanced embedded representation, and then performing dot product operation on the fused user embedded representation and the project embedded representation, and formulating a recommendation list according to a score matrix of the user to the project generated after dot product.
According to a further technical scheme, the obtaining of the time perception information when the user and the project are subjected to multi-type interaction specifically comprises:
let user u j And item i s The interaction relationship between the users is the t type relationship, and the user u is used for j And item i s Time stamping for t-type interactions
Is mapped as
Coding the generated time sequence based on a position coding method of a Transformer framework to obtain a time embedded expression
For representing users u at different times j And item i s Each t-th type of interaction between them.
According to the further technical scheme, the relation perception dependency network is utilized to model a multi-type interaction behavior mode, and the modeling specifically comprises a message propagation process and a message aggregation process;
acquiring perception information of different interaction types from the interactive users and projects by describing a message propagation process of the users and the projects including time perception information in the interaction;
and realizing the message aggregation process of users and projects through a message aggregation network, and aggregating the perception information of different interaction types.
According to a further technical scheme, the polymerization process comprises the following steps:
obtaining an explicit relevance score for an embedded representation of a particular interaction type based on a message aggregation network;
aligning the embedded representation of the specific interaction type by using attention operation according to the explicit relevance score, and obtaining the updated embedded representation of the specific interaction type by splicing operation;
and aggregating the embedded representations of a plurality of specific interaction types in the message propagation process to obtain the embedded representation fusing a plurality of interaction behaviors.
The further technical scheme also comprises the following steps:
calculating an importance score of the embedded representation based on the embedded representation fusing the plurality of interactive behaviors;
according to the importance scores, describing the embedded representation of the high-order heterogeneous multi-type interaction;
and aggregating the embedded representations of the high-order heterogeneous multi-type interaction to obtain an embedded representation containing the high-order heterogeneous multi-type dynamic interaction, wherein the embedded representation comprises a user embedded representation and an item embedded representation.
According to the further technical scheme, the user embedded representation is enhanced by using an attention mechanism and a mutual information learning mode, and the user enhanced embedded representation containing the user social information is obtained, and the method specifically comprises the following steps:
taking the obtained user embedded representation as a node-level user embedded representation, updating by using a graph attention machine mechanism, and obtaining the updated node-level user embedded representation which is transmitted through the connection edge of the user and the user;
generating a fused graph-level embedded representation of the user social graph based on the user social graph;
according to the sectionBuilding a context-aware discriminator by combining a mutual information structure paradigm with a point-level user-embedded representation and the fused graph-level embedded representation to construct a context-aware discriminator
As a positive user instance, i.e., a positive sample, to the evaluator
Inputting the negative user example, namely a negative sample, into the discriminator so as to train the discriminator; wherein the content of the first and second substances,
embedding representations, R, for node level users u Embedding representations for graph-level users;
based on the trained discriminator, it is determined whether the input user is a positive or negative case by determining a probability score that the input user belongs to the user's social graph.
The further technical scheme also comprises the following steps:
defining a corresponding loss function by taking the unobserved user-item interaction as a negative sample and the observed user-item interaction as a positive sample, training the whole recommendation model, and obtaining a user embedded representation with enhanced characteristics, namely the user enhanced embedded representation containing the social information of the user, through the trained recommendation model.
In a second aspect, the present disclosure provides a recommendation system based on user social information, including:
the time perception information processing module is used for acquiring time perception information when multi-type interaction is carried out between a user and a project based on a multi-interaction type user-project interaction diagram;
the multi-type user-project interaction module is used for modeling a multi-type interaction behavior mode by utilizing a multi-interaction type user-project interaction relation perception dependency network according to user representation, project representation and time perception information during user-project interaction to obtain user embedded representation and project embedded representation containing high-order heterogeneous multi-type dynamic interaction;
the global relationship enhanced social dependency module is used for enhancing the user embedded representation by utilizing a graph attention mechanism and a mutual information learning mode based on the user social graph to obtain the user enhanced embedded representation containing the user social information;
and the recommendation list generation module is used for fusing the user embedded representation and the user enhanced embedded representation and then performing dot product operation on the fused user embedded representation and the item embedded representation, and formulating a recommendation list according to a score matrix of the user to the item generated after dot product.
In a third aspect, the present disclosure also provides an electronic device comprising a memory and a processor, and computer instructions stored on the memory and executed on the processor, wherein the computer instructions, when executed by the processor, perform the steps of the method of the first aspect.
In a fourth aspect, the present disclosure also provides a computer-readable storage medium for storing computer instructions which, when executed by a processor, perform the steps of the method of the first aspect.
The above one or more technical solutions have the following beneficial effects:
the invention provides a recommendation method and a recommendation system based on user social information, which are characterized in that user representation and project representation with time perception information are integrated into a multi-interaction type user-project interaction relationship perception dependency network, user embedded representation and project embedded representation containing high-order heterogeneous multi-type dynamic interaction are generated, high-order multi-interaction perception collaborative semantic information is extracted, heterogeneous advantages of user social relationships are obtained by using a drawing and attention mechanism, and local-global context semantic information of users is obtained by using a mutual information structure model, so that fine-grained modeling of a recommendation model based on the user social information is realized, and recommendation performance and recommendation accuracy are improved.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, are included to provide a further understanding of the invention, and are included to illustrate an exemplary embodiment of the invention and not to limit the invention.
FIG. 1 is a flowchart illustrating the overall steps of a method according to an embodiment of the present invention;
fig. 2 is a flowchart of a method according to an embodiment of the present invention.
Detailed Description
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the invention. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
Most of the existing recommendation methods based on deep learning neglect the characteristic of diversity of the interaction types of the users and the projects, and lack deep analysis and research on the interaction of heterogeneous users. In particular, in many real-life recommendation scenarios, user-item interactions exhibit inherent relationships, such as clicking, commenting, adding to a shopping cart, purchasing, etc., in an online retail platform, where the addition of behavior information to a shopping cart may provide a powerful signal for final purchasing decisions. Therefore, how to obtain high-order collaborative signals and implicit dependency semantic features thereof from heterogeneous multi-type user-item interactions is a great challenge to improve recommendation performance.
Moreover, most of the current recommendation tasks utilizing the user social information do not consider heterogeneous multi-interaction type dynamic coding information, do not deeply research the social similarity of the user from local and global spaces of user social relationship heterogeneity, and do not consider the strength of the social influence of the user, so in order to solve the problems, the invention discloses a recommendation method and a recommendation system based on the user social information.
Example one
The embodiment provides a recommendation method based on user social information, as shown in fig. 1, including the following steps:
acquiring time perception information when multi-type interaction is carried out between a user and a project based on a multi-interaction type user-project interaction diagram;
modeling a multi-type interaction behavior mode by utilizing a multi-interaction type user-item interaction relation perception dependency network according to user representation, item representation and time perception information during user-item interaction to obtain user embedded representation and item embedded representation containing high-order heterogeneous multi-type dynamic interaction;
based on the user social graph, enhancing the user embedded representation by using a graph attention mechanism and a mutual information learning mode to obtain the user enhanced embedded representation containing the user social information;
and fusing the user embedded representation and the user enhanced embedded representation, performing dot product operation on the fused user embedded representation and the project embedded representation, and formulating a recommendation list according to a score matrix of the user to the project generated after dot product.
As shown in the figure 2 of the drawings, in the recommendation method of this embodiment, let U = { U = { (U) } 1 ,u 2 ,,,,u j H and I = { I = 1 ,i 2 ,,,,i s Are used to represent a collection of users and items, respectively, where u j Represents the jth user, j being the number of users, i s Representing the s-th item, s being the number of items, and defining the user-item interaction graph of the multi-interaction type as G on the basis of the s-th item s ＝{U,I,E s For representing different types of interactions between a user and an item, E s Representing a set of interaction behaviors between a user and an item, in this embodiment, E s Has t types of interactive behaviors, the types of interactive behaviors comprise purchase,Browse etc. at G s In, if user u j Under type t with item i s Interaction occurs, then
Otherwise
Meanwhile, a user social graph is defined as G u ＝{U,E u Represents a social relationship between the user and the user, where U = { U = } 1 ,u 2 ,,,,u j Denotes a user node, E u Representing a set of edges connecting users with users, i.e. if two users u j And u j' There is a social connection, then there is an edge between the two users.
Further, in order to make the description clearer, the present embodiment uses the embedded matrix to represent the users and the items, that is, the user embedded representation matrix P and the item embedded representation matrix Q:
wherein the content of the first and second substances,
representing user u j The user embedded representation of
Distinguished from user representation u j User represents u j Is j 1, and the user embedded representation
Dimension j x d; also, in the same manner as above,
represents item i s Is shown embedded. In the above-described embedded representation, "(0)" represents the initial embedded representation, and the embedded representation is performed a plurality of times in consideration of subsequent needsConvolution, and therefore avoids confusing representations for subsequent computations by changing the superscript of the embedded representation.
First, the present embodiment discloses a time-aware information processing method, which allows user-item interaction to be interleaved with different timestamps, and captures time information when multi-type interaction is performed between a user and an item. Specifically, user u is set j And item i s The interaction relationship between them is the t-th type relationship, at this time, the time stamp is used
Is mapped as
Generating a time-embedded representation
For representing users u at different times j And item i s Each t-th type of interaction therebetween, wherein,
representing individual slot functions.
To express the above time vector, i.e. user u j And item i s In this embodiment, the time sequence generated by performing the t-th type of interaction is encoded based on a position encoding method of a transform framework, so that the influence of the sequence on the final recommendation performance in the interaction process of the user and the item can be reflected in the subsequent analysis process, that is, the dynamic interaction characteristics of the user and the item can be reflected in the final recommendation. In this embodiment, the time measurements are encoded using the sinussoid sine function, generating a time-embedded representation. By utilizing the periodicity of the sine function, semantic information can be captured at a specific moment, the scheme can introduce different cycle sizes, and extract the relation between different cycles on different dimensions, so that the semantic information contained in various types of interaction between a user and a project can be better modeled at the stage, and the time embedding is expressed as:
in the above formula, lot represents the position of the coding element, and 2lot and 2lot +1 represent the coding elements of odd number and even number respectively; d represents an embedding size of the code vector; (j, s) represent the actual position of the code.
According to the scheme, time perception information of the user and the project in the interaction process is described, on the basis, a multi-interaction type user-project interaction relationship perception dependency network is constructed, modeling of the dependency relationship between the user and the project with multiple interaction types is achieved through the relationship perception dependency network, and heterogeneous behavior relationships between the user and the project are represented, namely different interaction types between the user and the project are represented.
User-item interaction graph G for expressing multiple interaction types by using relationship-aware dependency network s And the information is propagated, aggregated and converted, learning is carried out based on the embedded representation initialized in the interaction diagram, and finally the user embedded representation and the item embedded representation containing various dynamic interaction behaviors are obtained through learning.
The relationship-aware dependency network described above includes two important stages: message propagation and message aggregation.
First, by describing a message propagation process of users and items in interaction, perception information of a specific interaction type is obtained from the users and items of the interaction. Specifically, the message passing process between the user and the project is as follows: assuming that the interaction type is t, the user message, the time perception information on the interaction edge and the item information are combined into a message propagation formula, which is as follows:
wherein, the first and the second end of the pipe are connected with each other,
and
respectively represent the slave items i s Spread to user u j And a slave user u j Propagate to item i s The superscript of the parameter represents the l-th layer because the GCN is used in the neural network for multi-layer convolution;
user-embedded representation of representation initialization
With corresponding time-embedded representation
The elements in between are added up to each other,
item-embedded representation of representation initialization
With corresponding time-embedded representation
Element addition in between; ω (-) represents a message encoding function that preserves the unique characteristics of each interaction type (i.e., type t).
On the basis, the message propagation process of the user node is based on a user-item interaction graph G of multiple interaction types s Describing implementation in a set of item nodes N (j, t) adjacent to the user node, which is expressed as:
wherein the content of the first and second substances,δ c,t representing the weight of the tth user interaction type learned from the projected c-th potential dimension, under the interaction type t, N (j, t) represents the interaction with the user u j Set of connected neighboring item nodes, ξ (-) represents an activation function similar to ReLU, W 1 And b 1 Are learnable hyper-parameters.
Also, the message passing process of the project node is based on a user-project interaction graph G of multiple interaction types s The description of the adjacent user node set N (s, t) is realized in a similar way to the user node message transmission process, and the description is performed on the project node i under the interactive type t s The process of message propagation from neighboring user nodes is represented as:
the perception information of a specific interaction type (i.e., the t-th interaction type) is obtained through the description of message propagation in the user and project interaction process, and then the perception information of different interaction types is aggregated through a message aggregation network set in the relationship perception dependency network according to the embodiment. In this embodiment, the perception information of the user in the interaction graph is an item, that is, information contained in the item is aggregated into the user; and the perception information of the items in the interaction graph, namely the user, aggregates the information contained by the user into the items.
The message aggregation network is implemented based on a graph attention neural network, and can automatically learn the explicit relevance score of the embedded representation of a specific interaction type
Calculating which types of interaction behaviors are more important, and the formula is as follows:
where Q and K are transformation weight matrices, used to represent the projections embedded between different interaction types t and t'.
Scoring explicit relevance
Applying a softmax function, and performing an embedding projection process by utilizing H subspaces (H belongs to H) so as to realize interactive relation dependence modeling from different hidden dimensions, wherein the formula is as follows:
wherein, V h To transform the weights, att (-) represents the attention operation.
Thereafter, the embedded representations from the different learning subspaces are aggregated using a stitching operation to realign the embedded representations of a particular type by
Obtaining an embedded representation of a particular interaction type during message dissemination, the method
Representing updated user-embedded representations propagated through the connecting edges of the t-th class of interactions,
representing an original embedded representation
And realigned embedded representation
The elements between.
Then, the embedded representation of the specific interaction type in the process of the aggregate message propagation is expressed by the following formula:
wherein Aggregation (-) represents a message Aggregation function. For example, there are 5 kinds of heterogeneous interaction behaviors between the user and the project, and the embedded representation of each specific interaction type in the message dissemination process of the user is obtained through the scheme
And then integrating interactive behavior embedded representations with different embedding strengths by adopting an aggregation method.
Based on user embedding or project embedding fusing multiple interactive behaviors, an importance score is calculated to identify a representation of each user's specific interactive behavior type, taking user embedding as an example, the formula is:
where ξ (. Cndot.) denotes the ReLU activation function, W 3 And W 2 Represents a trainable transformation matrix, an
Where d represents the weight using a custom dimension size.
After obtaining the weight scores, the embedding aggregation process of the (l + 1) th layer of the user and the item is respectively defined, and the formula is as follows:
based on the message transmission and aggregation functions, the user-project interaction graph G of multiple interaction types is realized s And learning the description of the high-order heterogeneous multi-type interaction, specifically by superposing a plurality of message propagation layers. Formally defining the high order polymerization process as:
wherein, the first and the second end of the pipe are connected with each other,
for the updated user embedded representation, # indicates an add operation.
Based on the same scheme, similar operations are applied to the project to obtain the final project embedded representation
Through the scheme, the user embedded representation and the item embedded representation are updated, so that the updated user embedded representation and the updated item embedded representation learn the high-order heterogeneous multi-type interactive relation representation. Then, on the basis of the updated user embedded representation and item embedded representation, the embodiment deeply studies the user social similarity relationship from the local and global spaces of the user social relationship heterogeneity, considers the strength of the user social influence, captures different influences of the user social relationship by using an attention mechanism by considering the heterogeneity of the user social relationship strength, reasonably distributes the weight of the influence of neighbors on the user, and performs adaptive fine-grained modeling on the interaction between the user and the user, so as to further update and enhance the user embedded representation, improve the user-item recommendation performance on the basis and improve the recommendation accuracy.
In the present embodiment, the user social graph G is based on u Let us assume a node-level user-embedded representation
The user-embedded representation preserves the user's social graph G u The updated user embedded representation is used for updating the user social graph G u Let us order
Embedding representations into a user in view of subsequent needsA multi-dimensional operation is performed to enhance the feature representation, so that the updated user-embedded representation obtained as described above is represented here by another character.
Using a similar approach as described above, explicit relevance scores for different user-embedded representations are obtained
The intermediate values are then paired by the softmax function
Normalizing to obtain social attention scores between users
Namely, the specific formula is as follows:
wherein the content of the first and second substances,
and
represents the learnable projection matrix in the H-th subspace (H e H),
representing social attention scores between users, N j,j′ Indicating connection to user u j Of the neighbor user.
Then, the embedded representations from different learning spaces are aggregated by adopting splicing operation to obtain an updated user embedded representation propagated through the connection edges of the users, and the formula is as follows:
then, based on the user social graph G u Generating a user social graph G u The fused graph-level potential representation of (a), the formula of which is:
in the above formula, σ (-) is sigmoid activation function, b j,j′ And a j,j′ Are respectively degree matrix
And adjacent matrix
Of (1).
On the basis, in order to jointly enhance the embedding characteristics of the social relationship dependency of the local and global users, the node-level embedding representation is explored in the embodiment
Corresponding graph-level embedding representation R u Mutual information between them. Constructing and training a context-aware discriminator based on the mutual information structure paradigm from the user's social graph G u The middle part distinguishes positive samples and negative samples (pairwise relationship), and pairs G u The interrelationship between users is encoded while maintaining a connected topology.
In particular, to
As a positive user instance, i.e., a positive sample, to the discriminator
The discriminator is trained by inputting negative user instances, i.e., negative examples, into the discriminator. Wherein, the positive and negative samples are generated randomly through the node shuffling strategy. The identifier finished by the training distinguishes the user social graph G u The positive and negative samples in (1), i.e. input samples to the identifier after training, by judging the input user u j Belonging to a userSocial graph G u To determine whether the input user is a positive or negative example.
Defining a context-aware discriminator function v (·):
wherein, w 4 Is a learnable transformation matrix.
The user social graph G is distinguished by the discriminator u On the basis of the positive example and the negative example in (1), in order to further improve the recommendation performance of the entire recommendation model, in this embodiment, another positive example and another negative example are defined to train the entire model. Specifically, the observed user-item interactions are used as negative samples, and the observed user-item interactions are used as positive samples, and corresponding loss functions are defined, so that training is realized.
Further, a mutual information based loss L is defined α The formula is as follows:
wherein ε (-) is an indicator function, e.g.
And
corresponding to the positive and negative examples of the training phase, where any unobserved user-item interactions are considered negative examples, λ is the balance parameter, E p And E n Representing a user social graph G u Number of positive and negative examples above. Minimizing the above-mentioned losses helps to maximize mutual information and to jointly preserve node-specific user characteristics and global graph-level dependencies.
Moreover, the relative order between observed and unobserved user-item interactions is taken into account, and observed interactions can better reflectThe observed interactions should be given a higher predictive value than the unobserved interactions. Therefore, in the training process, the loss L based on mutual information α On the basis, the embodiment also adopts the BPR loss, and the final training loss formula is:
wherein O = { (j, s) + ,s - )∣(j,s + )∈R + ,(j,s - )∈R - Represents training data, R + represents observed interactions, R-represents unobserved interactions, σ () is a sigmoid function, and Θ represents trainable parameters. In the embodiment, the pre-regularization term | | Θ | | does not exist 2 Combined with the L2 regularization coefficient β to prevent overfitting.
By the scheme, the final user embedded representation with enhanced features is obtained
Wherein, the first and the second end of the pipe are connected with each other,
a set of representations is embedded for a user.
Fusing user-embedded representations generated based on multi-interaction behavior with user-embedded representations containing user social information using additive operations, i.e. using a combination of a plurality of user-embedded representations
Obtaining a final user embedded representation, dot-product the user embedded representation and the item embedded representation to obtain a score
Namely:
in this embodiment, the most common dot product operation in the recommendation algorithm is adopted, a scoring matrix of the user for the items is generated after dot product, TOP-K is set for selection, K items with the highest scoring in the items interacted with the user are selected and recommended to the user, a recommendation list containing the K items is obtained, and item recommendation is achieved.
Example two
The embodiment provides a recommendation system based on user social information, which comprises:
the time perception information processing module is used for acquiring time perception information when multi-type interaction is carried out between a user and a project based on a multi-interaction type user-project interaction diagram;
the multi-type user-project interaction module is used for modeling a multi-type interaction behavior mode by utilizing a multi-interaction type user-project interaction relation perception dependency network according to user representation, project representation and time perception information during user-project interaction to obtain user embedded representation and project embedded representation containing high-order heterogeneous multi-type dynamic interaction;
the global relationship enhanced social dependency module is used for enhancing the user embedded representation by utilizing a graph attention machine mechanism and a mutual information learning mode based on the user social graph to obtain the user enhanced embedded representation containing the user social information;
and the recommendation list generation module is used for fusing the user embedded representation and the user enhanced embedded representation and then performing dot product operation on the fused user embedded representation and the item embedded representation, and formulating a recommendation list according to a score matrix of the user to the item generated after dot product.
EXAMPLE III
The embodiment provides an electronic device, which comprises a memory, a processor and computer instructions stored in the memory and executed on the processor, wherein the computer instructions, when executed by the processor, implement the steps of the recommendation method based on the social information of the user.
Example four
The present embodiment also provides a computer-readable storage medium for storing computer instructions, which when executed by a processor, perform the steps of the recommendation method based on user social information as described above.
The steps involved in the second to fourth embodiments correspond to the first embodiment of the method, and the detailed description thereof can be found in the relevant description of the first embodiment. The term "computer-readable storage medium" should be taken to include a single medium or multiple media containing one or more sets of instructions; it should also be understood to include any medium that is capable of storing, encoding or carrying a set of instructions for execution by a processor and that cause the processor to perform any of the methods of the present invention.
Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using general purpose computer means, or alternatively, they can be implemented using program code that is executable by computing means, such that they are stored in memory means for execution by the computing means, or they are separately fabricated into individual integrated circuit modules, or multiple modules or steps of them are fabricated into a single integrated circuit module. The present invention is not limited to any specific combination of hardware and software.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Although the embodiments of the present invention have been described with reference to the accompanying drawings, it is not intended to limit the scope of the present invention, and it should be understood by those skilled in the art that various modifications and variations can be made without inventive efforts by those skilled in the art based on the technical solution of the present invention.
Claims (10)
1. A recommendation method based on user social information is characterized by comprising the following steps:
acquiring time perception information when multi-type interaction is carried out between a user and a project based on a multi-interaction type user-project interaction diagram;
modeling a multi-type interaction behavior mode by utilizing a multi-interaction type user-item interaction relation perception dependency network according to user representation, item representation and time perception information during user-item interaction to obtain user embedded representation and item embedded representation containing high-order heterogeneous multi-type dynamic interaction;
based on the user social graph, enhancing the user embedded representation by using a graph attention mechanism and a mutual information learning mode to obtain the user enhanced embedded representation containing the user social information;
and fusing the user embedded representation and the user enhanced embedded representation, performing dot product operation on the fused user embedded representation and the project embedded representation, and formulating a recommendation list according to a score matrix of the user to the project generated after dot product.
2. The recommendation method based on the user social information as claimed in claim 1, wherein the obtaining of the time perception information when the user and the item perform multi-type interaction specifically comprises:
let user u j And item i s The interaction relationship between the users is the t type relationship, and the user u is used for j And item i s Time stamping for t-type interactions
Is mapped as
Coding the generated time sequence based on a position coding method of a Transformer framework to obtain a time embedded expression
For representing users u at different times j And item i s Every t-th type of interaction between.
3. The recommendation method based on the user social information as claimed in claim 1, wherein the relationship-aware dependency network is used to model a multi-type interactive behavior pattern, specifically comprising a message propagation process and a message aggregation process;
acquiring perception information of different interaction types from the interactive users and projects by describing a message transmission process that the users and the projects contain time perception information in interaction;
and realizing the message aggregation process of users and projects through a message aggregation network, and aggregating the perception information of different interaction types.
4. The recommendation method based on the user social information as claimed in claim 3, wherein the aggregation process comprises:
obtaining an explicit relevance score for the embedded representation of the particular interaction type based on the message aggregation network;
aligning the embedded representation of the specific interaction type by using attention operation according to the explicit relevance score, and obtaining the updated embedded representation of the specific interaction type through splicing operation;
and aggregating the embedded representations of a plurality of specific interaction types in the message propagation process to obtain the embedded representation fusing a plurality of interaction behaviors.
5. The method as claimed in claim 4, further comprising:
calculating an importance score of the embedded representation based on the embedded representation fusing the multiple interactive behaviors;
according to the importance scores, describing the embedded representation of the high-order heterogeneous multi-type interaction;
and aggregating the embedded representations of the high-order heterogeneous multi-type interactions to obtain an embedded representation containing the high-order heterogeneous multi-type dynamic interactions, wherein the embedded representation comprises a user embedded representation and an item embedded representation.
6. The recommendation method based on the user social information as claimed in claim 1, wherein the enhancing the user embedded representation by using the graph attention mechanism and the mutual information learning manner to obtain the user enhanced embedded representation including the user social information specifically comprises:
taking the obtained user embedded representation as a node-level user embedded representation, updating by using a graph attention mechanism, and obtaining the updated node-level user embedded representation which is propagated through the connection edges of the user and the user;
generating a fused graph-level embedded representation of the user social graph based on the user social graph;
constructing a context-aware discriminator according to the node-level user embedded representation and the fusion graph-level embedded representation in combination with a mutual information structure paradigm to
As a positive user instance, i.e., a positive sample, to the discriminator
Inputting the negative user example, namely a negative sample, into the discriminator so as to train the discriminator; wherein the content of the first and second substances,
embedding representations, R, for node level users u Embedding representations for graph-level users;
based on the trained discriminator, it is determined whether the input user is a positive or negative case by determining a probability score that the input user belongs to the user's social graph.
7. The method as claimed in claim 6, further comprising:
defining a corresponding loss function by taking the unobserved user-item interaction as a negative sample and the observed user-item interaction as a positive sample, training the whole recommendation model, and obtaining a user embedded representation with enhanced characteristics, namely the user enhanced embedded representation containing the social information of the user, through the trained recommendation model.
8. A recommendation system based on user social information is characterized by comprising:
the time perception information processing module is used for acquiring time perception information when multi-type interaction is carried out between a user and a project based on a multi-interaction type user-project interaction diagram;
the multi-type user-project interaction module is used for modeling a multi-type interaction behavior mode by utilizing a multi-interaction type user-project interaction relation perception dependency network according to user representation, project representation and time perception information during user-project interaction to obtain user embedded representation and project embedded representation containing high-order heterogeneous multi-type dynamic interaction;
the global relationship enhanced social dependency module is used for enhancing the user embedded representation by utilizing a graph attention mechanism and a mutual information learning mode based on the user social graph to obtain the user enhanced embedded representation containing the user social information;
and the recommendation list generation module is used for fusing the user embedded representation and the user enhanced embedded representation and then performing dot product operation on the fused user embedded representation and the project embedded representation, and formulating a recommendation list according to a score matrix of the user to the project generated after dot product.
9. An electronic device, characterized by: comprising a memory and a processor and computer instructions stored on the memory and executed on the processor, which when executed by the processor, perform the steps of a method for user social information based recommendation according to any of claims 1-7.
10. A computer-readable storage medium characterized by: for storing computer instructions which, when executed by a processor, perform the steps of a method for user social information based recommendation as claimed in any one of claims 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211233278.4A CN115545833A (en) | 2022-10-10 | 2022-10-10 | Recommendation method and system based on user social information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211233278.4A CN115545833A (en) | 2022-10-10 | 2022-10-10 | Recommendation method and system based on user social information |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115545833A true CN115545833A (en) | 2022-12-30 |
Family
ID=84732643
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211233278.4A Pending CN115545833A (en) | 2022-10-10 | 2022-10-10 | Recommendation method and system based on user social information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115545833A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116662676A (en) * | 2023-06-09 | 2023-08-29 | 北京华品博睿网络技术有限公司 | Online recruitment bidirectional reciprocity recommendation system and method based on multi-behavior modeling |
-
2022
- 2022-10-10 CN CN202211233278.4A patent/CN115545833A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116662676A (en) * | 2023-06-09 | 2023-08-29 | 北京华品博睿网络技术有限公司 | Online recruitment bidirectional reciprocity recommendation system and method based on multi-behavior modeling |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112989064B (en) | Recommendation method for aggregating knowledge graph neural network and self-adaptive attention | |
| Liu et al. | Motif-preserving dynamic attributed network embedding | |
| CN112613602A (en) | Recommendation method and system based on knowledge-aware hypergraph neural network | |
| CN111881350A (en) | Recommendation method and system based on mixed graph structured modeling | |
| CN114265986B (en) | Information pushing method and system fusing knowledge graph structure and path semantics | |
| CN114817663B (en) | Service modeling and recommendation method based on class perception graph neural network | |
| CN113918832B (en) | Graph convolution collaborative filtering recommendation system based on social relationship | |
| CN113918833B (en) | Product recommendation method realized through graph convolution collaborative filtering of social network relationship | |
| CN113918834B (en) | Graph convolution collaborative filtering recommendation method fusing social relations | |
| CN115221413B (en) | Sequence recommendation method and system based on interactive graph attention network | |
| CN115270007B (en) | POI recommendation method and system based on mixed graph neural network | |
| Mukunthu et al. | Practical automated machine learning on Azure: using Azure machine learning to quickly build AI solutions | |
| CN115438169A (en) | Text and video mutual inspection method, device, equipment and storage medium | |
| CN115545833A (en) | Recommendation method and system based on user social information | |
| CN113590976A (en) | Recommendation method of space self-adaptive graph convolution network | |
| CN115631008B (en) | Commodity recommendation method, device, equipment and medium | |
| CN115982480A (en) | Sequence recommendation method and system based on cooperative attention network and comparative learning | |
| CN115599990A (en) | Knowledge perception and deep reinforcement learning combined cross-domain recommendation method and system | |
| CN114168804B (en) | Similar information retrieval method and system based on heterogeneous subgraph neural network | |
| CN115982373A (en) | Knowledge graph recommendation method combining multi-level interactive contrast learning | |
| CN113343100B (en) | Smart city resource recommendation method and system based on knowledge graph | |
| CN116932873A (en) | Video account recommending method, device, equipment, storage medium and program product | |
| CN116594627B (en) | Multi-label learning-based service matching method in group software development | |
| Yang | Item Recommendation Based on Monotonous Behavior Chains | |
| Wu et al. | SGT: Session-based Recommendation with GRU and Transformer |
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 |