CN116932923B - Project recommendation method combining behavior characteristics and triangular collaboration metrics - Google Patents

Project recommendation method combining behavior characteristics and triangular collaboration metrics Download PDF

Info

Publication number
CN116932923B
CN116932923B CN202311208260.3A CN202311208260A CN116932923B CN 116932923 B CN116932923 B CN 116932923B CN 202311208260 A CN202311208260 A CN 202311208260A CN 116932923 B CN116932923 B CN 116932923B
Authority
CN
China
Prior art keywords
user
representing
node
item
representation
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
Application number
CN202311208260.3A
Other languages
Chinese (zh)
Other versions
CN116932923A (en
Inventor
钱忠胜
张丁
王亚惠
俞情媛
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jiangxi University of Finance and Economics
Original Assignee
Jiangxi University of Finance and Economics
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Jiangxi University of Finance and Economics filed Critical Jiangxi University of Finance and Economics
Priority to CN202311208260.3A priority Critical patent/CN116932923B/en
Publication of CN116932923A publication Critical patent/CN116932923A/en
Application granted granted Critical
Publication of CN116932923B publication Critical patent/CN116932923B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • 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
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G06Q10/00—Administration; Management
    • G06Q10/10—Office automation; Time management
    • G06Q10/103—Workflow collaboration or project management
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Business, Economics & Management (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Quality & Reliability (AREA)
  • Operations Research (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a project recommendation method combining behavior characteristics and triangular collaboration measurement, which comprises the steps of firstly taking a user-project as a core, generating an isomerised user behavior characteristic representation and a project representation through an embedded propagation mechanism, a neighbor routing mechanism and an isomerism graph neural network, and avoiding subsequent user preference singleization; secondly, obtaining trust relations among users through an improved Bhattacharyya similarity distance induction strategy, and further constructing a triangular collaboration measurement model based on user-project-user, so that inaccurate scoring of a predicted project is avoided; and finally, forming triples by the predictive scoring items, the positive scoring items and the negative scoring items, and obtaining the latest dynamic preference of the user through condition similarity learning so as to more accurately recommend the items for the user.

Description

Project recommendation method combining behavior characteristics and triangular collaboration metrics
Technical Field
The invention relates to the technical field of data processing, in particular to a project recommendation method combining behavior characteristics and triangular collaboration metrics.
Background
Collaborative metric learning (Collaboration Metric Learning, CML) is a popular method in current recommendation systems, which predicts item scores through triangle inequality and similarity propagation, can mine implicit information such as user-item interactions, and has good effect on applications.
In the real world, the scoring of items by a user depends on the user's diverse, dynamic preferences. However, existing CML models use only a single user vector representation when modeling a user, resulting in an inability to accurately express the diversity of user preferences when the user has multiple opposite pairs of interests; moreover, most models only establish user-project relationship, and the effect of trust relationship among users on project scoring is not considered, so that the predicted project scoring is inaccurate; in addition, ignoring the impact of item scoring on the predicted user results in the resulting user preferences being static.
Disclosure of Invention
Therefore, the embodiment of the invention provides a project recommendation method combining behavior characteristics and triangular collaboration metrics, which aims to solve the problems that the prior art cannot accurately express the diversity of user preferences, the predicted project scoring is inaccurate, the obtained user preferences are static, and the like.
According to an embodiment of the invention, a project recommendation method combining behavior characteristics and triangular collaboration metrics comprises the following steps:
step 1, learning adjacent module information through a neighbor routing mechanism and an embedded propagation mechanism on the basis of the existing user-project interaction, iteratively updating to obtain diversified behavior characteristics of a user, introducing a heterogeneous graph neural network to realize information transfer among the multiple behavior characteristics, and finally obtaining heterogeneous user behavior characteristic representation and project representation;
step 2, based on the isomerised user behavior characteristic representation and the project representation, an improved Bhattacharyya similarity distance induction strategy is established, a trust relationship between users is established through similar neighbors of the target users, and a triangular collaboration measurement model based on the user-project-user is further established;
and 3, calculating triples consisting of predictive scoring items, active scoring items and passive scoring items by utilizing a triangular collaboration measurement model and combining condition similarity learning, obtaining the latest dynamic preferences of the user through the condition similarity learning of the triples, and recommending the items to the user based on the latest dynamic preferences of the user.
According to the project recommendation method combining the behavior characteristics and the triangular collaboration metrics, firstly, modeling is carried out on the user behavior characteristics by utilizing an isomerization graph neural network, and the user behavior characteristics are iteratively updated by utilizing a neighbor routing mechanism and an embedded propagation mechanism, so that the problem of behavior characteristic singleization caused by unified modeling of a user is solved, and the user behavior characteristics are diversified; secondly, a Bhattacharyya similarity distance induction strategy is introduced to capture the potential relation of a user-user, and the potential relation of the user-project and the project-project are combined to construct a stable triangular collaboration measurement relation, so that the problem of inaccurate scoring of the project predicted by the existing model is solved; finally, the predictive scoring items, the positive scoring items and the negative scoring items are formed into a ternary antagonism relationship through the condition similarity, so that the dynamic recommendation of the user preference is implemented, the problem of the user preference statism of the traditional CML model is solved, and particularly, compared with the prior art, the method has the following beneficial effects:
1) In order to solve the problem that a single user vector cannot capture the diversified preferences of users in the conventional collaborative metric learning, the invention adopts multi-vector user representation with behavior characteristics. In reality, the interests of users are diversified, but in traditional collaborative metric learning, user preferences are captured by using a single user vector representation, which does not enable users to approach two dissimilar items at the same time. Under the limitation, the heterogeneity of the user preference can not be satisfactorily processed, but the multi-vector user behavior characteristic is adopted to express that the problem of limitation of the user preference caused by a single user vector in the original collaborative metric learning can be relieved, so that the finally obtained user preference is diversified, and the recommendation accuracy is further improved;
2) In the traditional collaborative metric learning, similarity calculation only depends on a user-item relationship, so that the final top-N recommendation accuracy is low. The trust relationship between users is calculated through a Bhattacharyya method, and a stable triangular collaboration measurement relationship is established by integrating user-project and project-project, so that the prediction accuracy of recommendation can be effectively improved;
3) The existing collaborative metric learning only predicts the items possibly interacted by the user by considering the simple representation of the user-items from the neighborhood angle, the influence of the predicted scoring items, the triples formed by the positive scoring items and the negative scoring items on the predicted user preference is not considered, the final user preference acquired by the collaborative metric learning is static, and the invention utilizes the condition similarity to read the countermeasure information existing in the implicit data, thereby acquiring the user dynamic preference, and being beneficial to improving the recommendation performance of the model.
Drawings
The foregoing and/or additional aspects and advantages of embodiments of the invention will be apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:
fig. 1 is a flowchart of an item recommendation method combining behavioral characteristics and triangular collaboration metrics according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Referring to fig. 1, a project recommendation method combining behavioral characteristics and triangular collaboration metrics includes steps 1-3:
and step 1, learning adjacent module information through a neighbor routing mechanism and an embedded propagation mechanism on the basis of the existing user-project interaction, iteratively updating to obtain diversified behavior characteristics of a user, introducing a heterogeneous graph neural network to realize information transfer among the multiple behavior characteristics, and finally obtaining heterogeneous user behavior characteristic representation and project representation.
Wherein, step 1 specifically includes step 1.1 to step 1.3:
step 1.1, mapping nodes of different categories to a shared feature space based on existing user-project interactions.
Each item node has different roles in neighbor node path embedding and generates different influences, so that before information from different neighbor nodes is aggregated for each item node, the influence of node-level attention learning neighbor node paths on heterogeneous user behavior characteristics is introduced, and neighbor node representations with common interaction information are aggregated to form node embedding. Because the dimensions of user-project nodes are different, node features need to be mapped to the same space, and the expression for mapping the nodes of different categories to the shared feature space is as follows:
；
wherein,representing item node->Node category of->Is->Corresponding weight matrix, < >>For project node->Corresponding initial node representation,/>For project node->Corresponding feature mapped node representation, +.>Representation->Bias vector of>Representing the dimension.
Through the processing, the characteristics of the user and the project node can be better fused, and unified modeling and learning can be performed in the heterogeneous graph neural network.
And 1.2, evaluating the attention weight between each node and the adjacent nodes, and carrying out the isomerization behavior feature separation of the single user based on the attention weight between each node and the adjacent nodes to obtain the user behavior feature.
After feature mapping of all nodes is obtained, importance among project nodes needs to be measured on the basis of existing user-project interaction in order to separate behavior features of users. This may be achieved by an attention network with which the attention weight between each node and its neighbour is calculated, in particular in step 1.2 the attention weight between each node and its neighbour is calculated with the following expression:
；
wherein,representing item node->And project node->Attention weight of->For project node->Corresponding feature mapped node representation, +.>For project node->Corresponding feature mapped node representation, +.>Representation->And->Vectors formed by joining together ∈ ->Representation->And->The vectors formed by the connection together are,Wis a learnable parameter matrix, and the LeakyReLU is an activation function, which can enhance the sparsity and robustness of the network; />Is item node->Neighbor set of->Representing an index.
Such attention weight computing mechanisms are widely used in graph neural networks for learning important relationships and weights between nodes. It should be noted that the present invention adopts heterogeneous graph neural network, and there is no symmetry between nodes, i.e. project nodesFor project nodejAnd project node->Project node->The impact of (c) may vary considerably.
Through the attention calculation, the neighbor nodes of each node can be weighted, resulting in a more meaningful node representation.
Furthermore, to ensure independence and fine granularity of each behavioral characteristic, an iso-composition is employed to model each user's behavioral characteristics individually. First, after the information acquisition of the first-hop neighbor is completed, the information of the higher-hop neighbor is further stacked to update the representation of the user behavior characteristics and enrich the information thereof. The calculation formula of the attention weight obtains the attention weight by calculating a linear combination of node characteristics and using the LeakyReLU activation function. Because the nodes in the heterogeneous graph neural network do not have symmetry, the invention uses the following steps ofsoftmaxAttention weight of function to all neighbor nodesNormalizing to obtain project node->And its neighbor node->Normalized attention weight between +.>Thereby yielding an embedded representation of the node with respect to its path. Through the mode, the characteristics of the behavior characteristics of the user can be better captured, and the independence and the richness of the user can be kept.
The ultimate goal of the above is to achieve isomerization behavior feature separation for individual users. In acquiring higher order neighbors of a userAfter information, setting the perception and update times of the behavior characteristics as followsIn->After a number of iterations, a propagation mechanism is applied to update the user behavior feature distribution.
Specifically, step 1.2 also satisfies the following conditional expression:
；
；
；
wherein,representation->Normalized value, ++>Representing item node->Corresponding node path, ++>Representing user +.>First->Jumping the item set of the neighbor; />Is indicated at +.>After a number of iterations, user->Collecting the user behavior characteristics after information updating from the neighbor set; />Is user->And project node->In the Graph Laplacian matrix, which is used to define Graph convolution and Graph attention operations, has many important properties, such as representation learning and analysis of Graph structure and node characteristics.
Such an update process helps to better capture the user's behavioral characteristics information so that each user's behavioral characteristics can be effectively separated and updated in the heterogeneous map.
And 1.3, based on the user behavior characteristics, iteratively updating the obtained single-user isomerization behavior characteristic block representation through a neural network, and iteratively updating the single-user isomerization behavior characteristic block representation through a heterogeneous graph neural network based on a neighbor routing mechanism and an embedded propagation mechanism to obtain an isomerization user behavior characteristic representation and a project representation.
Each time a user is updated, it starts from the user-centric interactive subgraph. In this sub-graph, interactive items driven by the same behavioral characteristics will have similar partitioned representations. Therefore, the weights among the nodes can be adjusted according to the interaction set and the neighbor set of the user, and in particular, the following conditional expression is satisfied in the step 1.3:
，/> ；
；
wherein,indicate->Item representation obtained after layer iteration update, +.>Item representation function representing first layer acquisition of neural network,/->For project node->Is represented by the kth block of +.>Representation->And->The degree of intimacy between the two,SELUthe self-normalized activation function can keep the mean value and variance of input data stable in the deep neural network, and is beneficial to relieving the problems of gradient disappearance and gradient explosion; />Representing user +.>Collected and behavioural characteristics from the x-1 th hop neighbor set +.>Related information->Indicate->User after layer iterative update>In->Representation vector on individual behavior features, +.>Representing an isomerized user intent representation function, +.>Is a weight matrix that can be learned, +.>Indicate->User after layer iterative update>In->Representation vector on individual behavior features, +.>Indicate->Item node obtained after layer iteration update>In->Representation vector on individual behavior features, +.>Representing user +.>Is->And (5) jumping the neighbor set.
The method and the system can better utilize the similarity among the users to update the behavior characteristic representation of the users, and iterate the updating process for a plurality of times, so that the behavior characteristic information of the users is enriched gradually. Such a multi-tiered update mechanism is able to capture more advanced features of the user behavior features and effectively integrate these features into the final user representation. In this way, the interests and demands of the user can be more accurately grasped, and therefore the performance and accuracy of the recommendation model are improved.
And obtaining an isomerism user behavior characteristic representation and a project representation by utilizing a isomerism graph neural network of a neighbor routing mechanism and an embedded propagation mechanism, then providing an improved Bhattacharyya similarity distance induction strategy for the isomerism user behavior characteristic and the project representation, establishing a trust relationship between users through similar neighbors of a target user, and further constructing a triangular collaboration measurement model based on the users, the projects and the users.
And 2, establishing an improved Bhattacharyya similarity distance induction strategy based on the heterogeneous user behavior characteristic representation and the project representation, establishing a trust relationship between users through similar neighbors of the target users, and further establishing a triangular collaboration measurement model based on the user-project-user.
Wherein, step 2 specifically includes step 2.1 to step 2.3:
and 2.1, defining Bhattacharyya similarity through isomerism user behavior characteristic representation, and obtaining an improved Bhattacharyya similarity distance induction strategy for the next trust relationship calculation.
Wherein the Bhattacharyya similarity distance induction strategy can calculate the similarity of two sets of samples by integrating overlapping parts. Referring to the definition of the Bhattacharyya similarity, a Bhattacharyya distance mode of calculating a sample is given, the smaller the value is, the higher the similarity is, and in particular, the following conditional expression is satisfied in step 2.1:
；
wherein,、/>two random samples of different sets, < >>Representation sample->And sample->The Bhattacharyya distance between,Tindicating the operation of the transpose,A、Brespectively represent sample->、/>A corresponding two-dimensional random variable covariance matrix,detrepresenting determinant calculations, ++>Representing a logarithmic calculation.
According to the invention, by improving the Bhattacharyya similarity distance induction strategy, the similarity between users and between projects under different vector characterization can be learned, so that different similarity results under different user behavior characteristics are ensured.
Related researches show that the trust relationship is favorable for obtaining the similarity between users and between projects, and the invention mainly models the potential relationship between users. The trust relationship comprises direct trust and indirect trust, the direct trust relationship is processed into a multidimensional matrix vector, the number of neighbors of a user is set to be super-parameter G, certain relativity exists among items due to the attribute such as grading, the relationship between the user and the relationship between the items can be defined by using a Jaccard formula, and in particular, the following conditional expression is further satisfied in step 2.1:
；
；
wherein,representing user +.>And user->Association between->、/>Respectively represent user +>、/>Neighbor trust set,/>Representing item node->And project node->Association between->、/>Representing item nodes +.>、/>Neighbor trust set,/>Representing intersection +.>Representing a union.
And 2.2, weighing direct trust relationship and indirect trust relationship existing between users, projects and between users and projects by taking the isomerised user behavior characteristic representation as a user representation.
For the indirect relation among users, the improved Bhattacharyya strategy is utilized to acquire the similarity of any two items of the users; then calculating local similarity among users based on the project; and finally, obtaining a final trust relationship of the user by combining the direct trust relationship of the user, wherein the step 2.2 specifically satisfies the following conditional expression:
；
；
；
wherein,representing item node->And->Is used for the degree of similarity of (c) to (c),Hrepresenting the maximum value of the item score,/->Representing +.>Score->Is about the number of users and project node>Ratio of total number of users scored +.>Representing nodes for itemsScore->Is about the number of users and project node>The ratio of the total number of users scored; />Representing user +.>And->At item node->And->Lower->Local similarity of individual user behavior features, +.>Representing user +.>Project node->Is (are) overall score value, ">Representing user +.>Project node->Is (are) overall score value, ">Is the median of the scoring values,/->、/>Item nodes->、/>In user behavior feature->The lower score represents;Kfor the total number of user behavior characteristics>Representing user similarity obtained with improved Bhattacharyya similarity distance induction strategy,/I>、/>Respectively represent users、/>Is a set of interactive items.
Item relationships in the triangular collaboration metric learning are then obtained based on the user overall trust relationships. Firstly, calculating user similarity through a Bhattacharyya similarity distance induction strategy; then calculating the similarity of local projects; finally, the overall relevance of the project is obtained, specifically, the following conditional expression is also satisfied in step 2.2:
；
；
；
wherein,representing user +.>And->Similarity of->Representing user +.>Score->Item number and user->Ratio of total number of items scored, +.>Representing user +.>Score->Item number and user->The ratio of the total number of items scored; />Representing item node->And->At the user->And->Lower->Local similarity of individual user behavior features, +.>Representing item node->At the user->Score under->Representing item node->At the user->Scoring the lower part;representing user +.>For project node->And project node->Similarity between, wherein->As a target item, +.>As a scoring item;Urepresenting user set->Representing a collection of items.
And 2.3, constructing a triangular collaboration metric model of the user-item-user by using the direct trust relationship and the indirect trust relationship among the user, the item and the direct trust relationship among the user and the item, wherein the triangular collaboration metric model is used for predicting unscored items of the user.
After the user-item-user triangle collaboration metric relationship is obtained, the scoring of the target item is predicted. Firstly, calculating the similarity between the predicted target item and the scored item of the same user; then, selecting the first N items with the highest similarity as similar items of the target items; finally, the user's score for the target item is predicted using a summation function, and specifically, step 2.3 satisfies the following conditional expression:
；
wherein,representing user +.>Project node->Is predictive of (a) score of->、/>Respectively represent +.>、/>Average score of similar items, ++>Is the number of similar items; />Representing user +.>Project node->Is a score of (2).
After determining the predictive scores for all the target items, the top N items may be selected for recommendation to the user.
Calculating user similarity by using Bhattacharyya similarity distance induction strategy, obtaining the score of the user about the target item through the user-item-user triangle cooperation measurement relation, and re-evaluating the predicted score through conditional similarity learning to obtain the user dynamic preference.
And 3, calculating triples consisting of predictive scoring items, active scoring items and passive scoring items by utilizing a triangular collaboration measurement model and combining condition similarity learning, obtaining the latest dynamic preferences of the user through the condition similarity learning of the triples, and recommending the items to the user based on the latest dynamic preferences of the user.
Wherein, on solving the related top-N recommendation, the standard CML learns metric space coding and represents users and items as low-dimensional vectors. In such a space, user-item similarity is defined by Euclidean distanceL2 regularized measurement of (2), wherein>And->Is the user +.>And project node->The respective vectors represent. The core goal of CML is to pull up matching user-item pairs and push away unmatched pairs until the margin +.>。
As can be seen from the above, a disadvantage of conventional CMLs is that the resulting user preferences are static. The invention calculates the similarity between users by an improved Bhattacharyya method, and obtains the trust relationship between users based on the similarity, thereby predicting the score of the initial project of the user by utilizing the constructed user-project-user triangle cooperation measurement relationship, and learning the dynamic preference of the user by combining condition similarity learning (Conditional Similarity Learning, CSL).
Specifically, in step 3, the predictive scoring item is usedUser preference vectorXTaking the positive scoring item as a vectorYTaking the negative scoring item as a vectorZRepresented as triplets；
Step 3 satisfies the following conditional expression:
；
；
；
wherein,representing feature extractor->Indicating pass->For a pair ofXExtracting the characteristic vector->Indicating pass->For a pair ofYExtracting the characteristic vector->Indicating pass->For a pair ofZThe feature vector obtained after the extraction is used for extracting the feature vector,representing predictionsXAnd (3) withYIs the distance between (2)Leave, go up>Representing predictionsXAnd (3) withZIs used for the distance of (a),Lthe projection matrix is represented by a matrix of projections,Tindicating the operation of the transpose, the term L2 is used to denote L2 norms, ">Expressed in triplet->Calculating the combinationX, Z) And%X, Y) Distance deviation of (2).
And after the triplet distance is acquired, the CSL is utilized to adjust the predicted value. Different user behavior characteristics may exhibit different results when looking for the user's latest preferences. Based on the embedded layer, stacking a plurality of self-attention blocks, extracting information from different subspaces by adopting a multi-head self-attention mechanism, and finally obtaining final prediction item scoring expression under different user behavior characteristics, wherein the following conditional expression is further satisfied in the step 3:
；
wherein,representing predicted values +.>Representing dynamic preference influencing parameters,/->Is used for detecting->Parameter of validity of individual behavior characteristics to user preferences, < ->Is thatLVectors obtained after expansion, < >>Is->Vectors obtained after expansion, < >>Representing the matrix arithmetic symbols, < >>In order to be a multi-task loss function,Sfor the parameter (0 if repeated, or 1 otherwise) for detecting the repeatability of the user behavior feature, ->Is edge distance, is->Indicate->Personal behavior characteristics,/->Is the distance difference [] + Representing standard hinge loss.
The improved cross entropy loss function is used for sequencing the acquired user preferences again through calculation and verifying the accuracy of prediction according to the improved cross entropy loss functionlossThe expression of (2) is:
；
wherein,representing the true value +_>For regular term->Is a weight of (2).
And fitting the predicted value with the true value, and finally obtaining the dynamic preference of the user and recommending according to the dynamic preference.
In step 3, performing item recommendation on the user based on the latest dynamic preference of the user specifically includes:
based on the latest dynamic preference of the user, obtaining the item scores of the possible next interaction of the user;
and recommending the item with the item score higher than the threshold value to the user.
In summary, the project recommendation method combining the behavior characteristics and the triangular collaboration metrics provided by the invention has the following beneficial effects:
1) In order to solve the problem that a single user vector cannot capture the diversified preferences of users in the conventional collaborative metric learning, the invention adopts multi-vector user representation with behavior characteristics. In reality, the interests of users are diversified, but in traditional collaborative metric learning, user preferences are captured by using a single user vector representation, which does not enable users to approach two dissimilar items at the same time. Under the limitation, the heterogeneity of the user preference can not be satisfactorily processed, but the multi-vector user behavior characteristic is adopted to express that the problem of limitation of the user preference caused by a single user vector in the original collaborative metric learning can be relieved, so that the finally obtained user preference is diversified, and the recommendation accuracy is further improved;
2) In the traditional collaborative metric learning, similarity calculation only depends on a user-item relationship, so that the final top-N recommendation accuracy is low. The trust relationship between users is calculated through a Bhattacharyya method, and a stable triangular collaboration measurement relationship is established by integrating user-project and project-project, so that the prediction accuracy of recommendation can be effectively improved;
3) The existing collaborative metric learning only predicts the items possibly interacted by the user by considering the simple representation of the user-items from the neighborhood angle, the influence of the predicted scoring items, the triples formed by the positive scoring items and the negative scoring items on the predicted user preference is not considered, the final user preference acquired by the collaborative metric learning is static, and the invention utilizes the condition similarity to read the countermeasure information existing in the implicit data, thereby acquiring the user dynamic preference, and being beneficial to improving the recommendation performance of the model.
In the description of the present specification, a description referring to terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
While embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that: many changes, modifications, substitutions and variations may be made to the embodiments without departing from the spirit and principles of the invention, the scope of which is defined by the claims and their equivalents.

Claims (3)

1. An item recommendation method combining behavioral characteristics and triangular collaboration metrics, the method comprising the steps of:
step 1, learning adjacent module information through a neighbor routing mechanism and an embedded propagation mechanism on the basis of the existing user-project interaction, iteratively updating to obtain diversified behavior characteristics of a user, introducing a heterogeneous graph neural network to realize information transfer among the multiple behavior characteristics, and finally obtaining heterogeneous user behavior characteristic representation and project representation;
step 2, based on the isomerised user behavior characteristic representation and the project representation, an improved Bhattacharyya similarity distance induction strategy is established, a trust relationship between users is established through similar neighbors of the target users, and a triangular collaboration measurement model based on the user-project-user is further established;
step 3, calculating triples consisting of predictive scoring items, active scoring items and passive scoring items by utilizing a triangular collaboration measurement model and combining condition similarity learning, obtaining latest dynamic preferences of users by learning the condition similarity of the triples, and recommending items to the users based on the latest dynamic preferences of the users;
the step 1 specifically comprises the following steps:
step 1.1, mapping nodes of different categories to a shared feature space based on existing user-project interactions;
step 1.2, evaluating the attention weight between each node and the adjacent node, and carrying out the isomerization behavior feature separation of a single user based on the attention weight between each node and the adjacent node to obtain the behavior feature of the user;
step 1.3, based on user behavior characteristics, iteratively updating the obtained single-user isomerization behavior characteristic block representation through a neural network, and iteratively updating the single-user isomerization behavior characteristic block representation through a heterogeneous graph neural network based on a neighbor routing mechanism and an embedded propagation mechanism to obtain an isomerization user behavior characteristic representation and a project representation;
in step 1.1, the expressions for mapping the nodes of different classes to the shared feature space are:
；
wherein,representing item node->Node category of->Is->Corresponding weight matrix, < >>For project node->Corresponding initial node representation,/>For project node->Corresponding feature mapped node representation, +.>Representation->Bias vector of>Representing dimensions;
in step 1.2, the attention weight between each node and its neighbor node is calculated using the attention network, and the expression is as follows:
；
wherein,representing item node->And project node->Attention weight of->For project node->Corresponding feature mapped node representation, +.>For project node->Corresponding feature mapped node representation, +.>Representation->And->Vectors formed by joining together ∈ ->Representation->And->The vectors formed by the connection together are,Wis a learnable parameter matrix, leakyReLU is an activation function,/->Is item node->Neighbor set of->Representing an index;
step 1.2 also satisfies the following conditional expression:
；
；
；
wherein,representation->Normalized value, ++>Representing item node->Corresponding node path, ++>Representing a userFirst->Jumping the item set of the neighbor; />Is indicated at +.>After a number of iterations, user->Collecting the user behavior characteristics after information updating from the neighbor set; />Is user->And project node->Is a Graph Laplacian matrix of (C);
step 1.3 satisfies the following conditional expression:
，/> ；
；
wherein,indicate->Item representation obtained after layer iteration update, +.>Item representation function representing first layer acquisition of neural network,/->For project node->Is represented by the kth block of (c)，/>Representation->And->The degree of intimacy between the two,SELUis a self-normalizing activation function,/->Representing user +.>Collected and behavioural characteristics from the x-1 th hop neighbor set +.>Related information->Indicate->User after layer iterative update>In->Representation vector on individual behavior features, +.>Representing an isomerized user intent representation function, +.>Is a weight matrix that can be learned, +.>Indicate->User after layer iterative update>In->Representation vector on individual behavior features, +.>Indicate->Item node obtained after layer iteration update>In->Representation vector on individual behavior features, +.>Representing user +.>Is->A jump neighbor set;
the step 2 specifically comprises the following steps:
step 2.1, defining Bhattacharyya similarity through isomerism user behavior characteristic representation, and obtaining an improved Bhattacharyya similarity distance induction strategy for the next trust relationship calculation;
2.2, weighing direct trust relationship and indirect trust relationship existing between users, projects and between users and projects by taking the isomerised user behavior characteristic representation as a user representation;
2.3, constructing a triangular collaboration measurement model of the user-item-user by using the direct trust relationship and the indirect trust relationship among the user, the item and the direct trust relationship among the user and the item, wherein the triangular collaboration measurement model is used for predicting unscored items of the user;
step 2.1 satisfies the following conditional expression:
；
wherein,、/>two random samples of different sets, < >>Representation sample->And sample->The Bhattacharyya distance between,Tindicating the operation of the transpose,A、Brespectively represent sample->、/>Corresponding two-dimensional random variable covariance matrix, +.>Representing determinant calculations, ++>Representing a logarithmic calculation;
step 2.1 also satisfies the following conditional expression:
；
；
wherein,representing user +.>And user->Association between->、/>Respectively represent user +>、/>Neighbor trust set,/>Representing item node->And project node->Association between->、/>Representing item nodes +.>、/>Neighbor trust set,/>Representing intersection +.>Representing a union;
step 2.2 satisfies the following conditional expression:
；
；
；
wherein,representing item node->And->Is used for the degree of similarity of (c) to (c),Hrepresenting the maximum value of the item score,/->Representing +.>Score->Is about the number of users and project node>Ratio of total number of users scored +.>Representing +.>Score->Is about the number of users and project node>The ratio of the total number of users scored; />Representing user +.>And->At item node->And->Lower->Local similarity of individual user behavior features, +.>Representing user +.>Project node->Is (are) overall score value, ">Representing user +.>Project node->Is (are) overall score value, ">Is the median of the scoring values,/->、/>Item nodes->、/>In user behavior feature->The lower score represents;Kfor the total number of user behavior characteristics>Representing the use of improved Bhattacharyya similarity distance mutagenesisUser similarity acquired by guide strategy>、/>Respectively represent user +>、/>Is a set of interactive items;
step 2.2 also satisfies the following conditional expression:
；
；
；
wherein,representing user +.>And->Similarity of->Representing user +.>Score->The number of items and the number of items to the userRatio of total number of items scored, +.>Representing user +.>Score->Item number and user->The ratio of the total number of items scored; />Representing item node->And->At the user->And->Lower->Local similarity of individual user behavior features, +.>Representing item node->At the user->Score under->Representing item node->At the user->Scoring the lower part;representing user +.>For project node->And project node->Similarity between, wherein->As a target item, +.>As a scoring item;Urepresenting user set->Representing a collection of items;
step 2.3 satisfies the following conditional expression:
；
wherein,representing user +.>Project node->Is predictive of (a) score of->、/>Respectively represent +.>、/>Average score of similar items, ++>Is the number of similar items; />Representing user +.>Project node->Is a score of (2).
2. The method for recommending items by combining behavioral characteristics and trigonometric cooperation metrics according to claim 1, wherein in step 3, recommending items to a user based on the latest dynamic preference of the user specifically comprises:
based on the latest dynamic preference of the user, obtaining the item scores of the possible next interaction of the user;
and recommending the item with the item score higher than the threshold value to the user.
3. The method for item recommendation combining behavioral characteristics and trigonometric cooperation metrics according to claim 1, wherein in step 3, a predictive scoring item is used as a user preference vectorXTaking the positive scoring item as a vectorYTaking the negative scoring item as a vectorZRepresented as triplets；
Step 3 satisfies the following conditional expression:
；
；
；
；
wherein,representing feature extractor->Indicating pass->For a pair ofXExtracting the characteristic vector->Indicating pass->For a pair ofYExtracting the characteristic vector->Indicating pass->For a pair ofZThe feature vector obtained after the extraction is used for extracting the feature vector,representing predictionsXAnd (3) withYDistance of->Representing predictionsXAnd (3) withZIs used for the distance of (a),Lthe projection matrix is represented by a matrix of projections,Tindicating the operation of the transpose, the term L2 is used to denote L2 norms, ">Expressed in triplet->Calculating the combinationX, Z) And%X, Y) Distance deviation of (2); />Representing predicted values +.>Representing dynamic preference influencing parameters,/->Is used for detecting->Parameter of validity of individual behavior characteristics to user preferences, < ->Is thatLVectors obtained after expansion, < >>Is->Vectors obtained after expansion, < >>Representing the matrix arithmetic symbols, < >>In order to be a multi-task loss function,Sfor the parameter for detecting the repeatability of the user behavior feature +.>Is edge distance, is->Indicate->Personal behavior characteristics,/->Is the distance difference [] + Representing standard hinge loss;
verifying the accuracy of the prediction in step 3 based on the improved cross entropy loss functionThe expression of (2) is:
；
wherein,representing the true value +_>For regular term->Is a weight of (2).
CN202311208260.3A 2023-09-19 2023-09-19 Project recommendation method combining behavior characteristics and triangular collaboration metrics Active CN116932923B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311208260.3A CN116932923B (en) 2023-09-19 2023-09-19 Project recommendation method combining behavior characteristics and triangular collaboration metrics

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311208260.3A CN116932923B (en) 2023-09-19 2023-09-19 Project recommendation method combining behavior characteristics and triangular collaboration metrics

Publications (2)

Publication Number Publication Date
CN116932923A CN116932923A (en) 2023-10-24
CN116932923B true CN116932923B (en) 2023-12-08

Family

ID=88390146

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311208260.3A Active CN116932923B (en) 2023-09-19 2023-09-19 Project recommendation method combining behavior characteristics and triangular collaboration metrics

Country Status (1)

Country Link
CN (1) CN116932923B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117851687B (en) * 2024-03-04 2024-05-14 江西财经大学 Multi-behavior graph contrast learning recommendation method and system integrating self-attention mechanism

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2017027480A (en) * 2015-07-24 2017-02-02 大学共同利用機関法人情報・システム研究機構 Item recommendation system and item recommendation method
CN110955829A (en) * 2019-11-19 2020-04-03 江西财经大学 Interest point recommendation method and system fusing credibility and measurement factor matrix decomposition
CN111241423A (en) * 2020-01-17 2020-06-05 江西财经大学 Deep recommendation method and system integrating trust distrust relation and attention mechanism
AU2020101885A4 (en) * 2020-08-19 2020-09-24 Xinjiang University A Novel Tensor Factorization Using Trust and Rating for Recommendation, system and method thereof
CN115885297A (en) * 2020-06-25 2023-03-31 微软技术许可有限责任公司 Differentiable user-item collaborative clustering
CN116108246A (en) * 2022-12-06 2023-05-12 重庆邮电大学 Semantic asymmetric similarity recommendation system considering consistency of user preferences
CN116340648A (en) * 2023-03-27 2023-06-27 广东工业大学 Knowledge graph attention network recommendation method based on graph collaborative filtering
CN116595157A (en) * 2023-07-17 2023-08-15 江西财经大学 Dynamic interest transfer type session recommendation method and system based on user intention fusion

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017043244A1 (en) * 2015-09-07 2017-03-16 富士フイルム株式会社 Search device, search system, and search method
CN113222700B (en) * 2021-05-17 2023-04-18 中国人民解放军国防科技大学 Session-based recommendation method and device
CN113254803B (en) * 2021-06-24 2021-10-22 暨南大学 Social recommendation method based on multi-feature heterogeneous graph neural network

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2017027480A (en) * 2015-07-24 2017-02-02 大学共同利用機関法人情報・システム研究機構 Item recommendation system and item recommendation method
CN110955829A (en) * 2019-11-19 2020-04-03 江西财经大学 Interest point recommendation method and system fusing credibility and measurement factor matrix decomposition
CN111241423A (en) * 2020-01-17 2020-06-05 江西财经大学 Deep recommendation method and system integrating trust distrust relation and attention mechanism
CN115885297A (en) * 2020-06-25 2023-03-31 微软技术许可有限责任公司 Differentiable user-item collaborative clustering
AU2020101885A4 (en) * 2020-08-19 2020-09-24 Xinjiang University A Novel Tensor Factorization Using Trust and Rating for Recommendation, system and method thereof
CN116108246A (en) * 2022-12-06 2023-05-12 重庆邮电大学 Semantic asymmetric similarity recommendation system considering consistency of user preferences
CN116340648A (en) * 2023-03-27 2023-06-27 广东工业大学 Knowledge graph attention network recommendation method based on graph collaborative filtering
CN116595157A (en) * 2023-07-17 2023-08-15 江西财经大学 Dynamic interest transfer type session recommendation method and system based on user intention fusion

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于信任和项目偏好的协调过滤算法;郑洁;钱育蓉;杨兴耀;黄兰;马婉贞;;计算机应用(第10期);144-148、158 *

Also Published As

Publication number Publication date
CN116932923A (en) 2023-10-24

Similar Documents

Publication Publication Date Title
CN111523029B (en) Personalized recommendation method based on knowledge graph representation learning
CN110929161B (en) Large-scale user-oriented personalized teaching resource recommendation method
CN110674407A (en) Hybrid recommendation method based on graph convolution neural network
CN109743196B (en) Network characterization method based on cross-double-layer network random walk
CN107545471B (en) Big data intelligent recommendation method based on Gaussian mixture
Bin et al. Collaborative filtering recommendation algorithm based on multi-relationship social network
CN116932923B (en) Project recommendation method combining behavior characteristics and triangular collaboration metrics
CN109947987B (en) Cross collaborative filtering recommendation method
CN110677284A (en) Heterogeneous network link prediction method based on meta path
CN106708953A (en) Discrete particle swarm optimization based local community detection collaborative filtering recommendation method
CN111382283A (en) Resource category label labeling method and device, computer equipment and storage medium
CN113918832A (en) Graph convolution collaborative filtering recommendation system based on social relationship
Huang et al. On the improvement of reinforcement active learning with the involvement of cross entropy to address one-shot learning problem
Vaswani et al. Horde of bandits using gaussian markov random fields
CN112364245B (en) Top-K movie recommendation method based on heterogeneous information network embedding
CN109857928A (en) User preference prediction technique based on polynary credit evaluation
CN111738514B (en) Social network community discovery method using local distance and node rank optimization function
Chang et al. Community detection with attributed random walk via seed replacement
CN114896514B (en) Web API label recommendation method based on graph neural network
CN114936890A (en) Counter-fact fairness recommendation method based on inverse tendency weighting method
Jing Application of Improved K‐Means Algorithm in Collaborative Recommendation System
CN111125541B (en) Method for acquiring sustainable multi-cloud service combination for multiple users
CN111428741B (en) Network community discovery method and device, electronic equipment and readable storage medium
CN113239265B (en) Personalized recommendation method and system based on connection matrix
Yu et al. ADPE: Adaptive dynamic projected embedding on heterogeneous information networks

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