CN116167812A - Heterogeneous collaborative filtering method for multi-behavior recommendation - Google Patents

Heterogeneous collaborative filtering method for multi-behavior recommendation Download PDF

Info

Publication number
CN116167812A
CN116167812A CN202211540250.5A CN202211540250A CN116167812A CN 116167812 A CN116167812 A CN 116167812A CN 202211540250 A CN202211540250 A CN 202211540250A CN 116167812 A CN116167812 A CN 116167812A
Authority
CN
China
Prior art keywords
behavior
user
recommendation
representation
collaborative filtering
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
Application number
CN202211540250.5A
Other languages
Chinese (zh)
Inventor
文凯
秋锴
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.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
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 Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN202211540250.5A priority Critical patent/CN116167812A/en
Publication of CN116167812A publication Critical patent/CN116167812A/en
Pending legal-status Critical Current

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
    • G06F16/9536—Search customisation based on social or collaborative filtering
    • 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
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Physics & Mathematics (AREA)
  • Economics (AREA)
  • General Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • Development Economics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to a heterogeneous collaborative filtering method for multi-behavior recommendation, which belongs to the field of multi-behavior recommendation, and is based on a heterogeneous collaborative filtering model HCFMR of the multi-behavior recommendation, and comprises the following steps: s1: the improved lightweight graph convolutional network is utilized to obtain user and project interaction characteristics of specific behaviors; s2: multiple behavior dependent semantics of a user bottom layer are obtained through a multi-head attention network; s3: setting a weight of a specific behavior for each user, and distinguishing the contribution degree of different behaviors; s4: and obtaining the high-order embedded representation of the node by aggregating all the convolution layers to obtain the preference of the user, thereby performing multi-behavior recommendation. The invention reduces the complexity of the model; the embedded display of users and projects is enriched, and the recommendation quality is improved; the interpretability of the model is improved.

Description

Heterogeneous collaborative filtering method for multi-behavior recommendation
Technical Field
The invention belongs to the field of multi-behavior recommendation, and relates to a heterogeneous collaborative filtering method for multi-behavior recommendation.
Background
To address the dilemma of information overload, recommendation systems have grown and become an important tool to assist user decision making. Conventional recommendation systems often focus on one type of interaction data (e.g., purchase), but in actual recommendation, multiple types of interaction data (e.g., click-to-browse, join shopping carts, collect, purchase, etc.) can be utilized, thereby building a fine-grained recommendation system and alleviating the data sparseness problem.
The multi-behavior recommendation system improves recommendation effects by utilizing matrix decomposition and a neural network method in early researches based on various interaction data of users, but only utilizes first-order interaction information to ignore a large amount of high-order heterogeneous interactions contained in the multi-behavior data. Some recent methods based on graph-rolling networks (Graph Convolution Network, GCN) make up for the defect that the early methods cannot capture high-order interaction information by modeling user and project interactions on two graphs, but follow a typical collaborative filtering idea, encode only high-order heterogeneous paths of users and projects, and ignore some fine-granularity interaction semantics of the bottoms of the users and the projects.
For example, in the actual e-commerce recommendation scene, various behavior data such as click browsing, collection, shopping cart adding, purchase and the like exist, and each type of behavior can reflect user preferences from different angles. Some users may add items of interest to the shopping cart for comparison prior to purchase, others may purchase after clicking for browsing, and the reasons for the behavioral dependence may be closely related to the user's traits (consumption ability, shopping habits) and item attributes (prices). In modeling user preferences, it is beneficial to incorporate dependencies on different behaviors exhibited by a user into a representation of the user, item, to promote recommendation. Moreover, of the various behaviors of the user, purchasing is obviously a stronger signal than clicking browsing, and if each behavior is treated uniformly is obviously inadequate, it is necessary to represent the strengths of the different behaviors from the data when modeling the behavior recommendation model. Some previous work has always ignored these considerations, making the performance of the recommendation system compromised.
Disclosure of Invention
Accordingly, the present invention is directed to a multi-behavior recommendation model HCFMR (Heterogeneous collaborative filtering for multi-behavior recommendation).
In order to achieve the above purpose, the present invention provides the following technical solutions:
a heterogeneous collaborative filtering method for multi-behavior recommendation based on a heterogeneous collaborative filtering model HCFMR for multi-behavior recommendation comprises the following steps:
s1: the improved lightweight graph convolutional network is utilized to obtain user and project interaction characteristics of specific behaviors;
s2: multiple behavior dependent semantics of a user bottom layer are obtained through a multi-head attention network;
s3: setting a weight of a specific behavior for each user, and distinguishing the contribution degree of different behaviors;
s4: and obtaining the high-order embedded representation of the node by aggregating all the convolution layers to obtain the preference of the user, thereby performing multi-behavior recommendation.
Further, in step S1, obtaining the multi-behavior-dependent semantics of the user bottom layer through the multi-head attention network specifically includes:
based on a lightweight graph convolutional network Light GCN, joint embedding is carried out on nodes and relations to obtain respective embedded representations, the representations of the nodes and the relations are updated in a mutually enhanced mode, and specific representations of users and the relations are shown as formulas (1) and (2):
Figure BDA0003977181060000021
Figure BDA0003977181060000022
in (i, r) k ) E N (u) represents the set of items i interacting with user u under the kth relationship,
Figure BDA0003977181060000023
is a normalization operation; />
Figure BDA0003977181060000024
Is a layer-specific matrix of trainable parameters; />
Figure BDA0003977181060000025
Is a combined function for incorporating the relational embedded representation intoIn the node representation, the node representation is enhanced;
respectively aggregating neighbor features under specific relations, and obtaining neighbor features of different orders of users through L-order propagation of graph rolling network
Figure BDA0003977181060000026
Relational embedding representation comprising interaction features of a specific behavior>
Figure BDA0003977181060000027
Multiple behavior dependent semantics with the bottom layer->
Figure BDA0003977181060000028
Further, in step S2, the obtaining, through the multi-head attention network, the multi-behavior-dependent semantics of the user bottom layer specifically includes:
s21: defining a query, key, value transformation matrix Q, K,
Figure BDA0003977181060000029
For the relationship k and->
Figure BDA00039771810600000210
Embedding projections between representations;
s22: will query vectors
Figure BDA00039771810600000211
And all key vectors->
Figure BDA00039771810600000212
Scaling the dot product to obtain the correlation weight between the relation embeddings>
Figure BDA00039771810600000213
And normalize it to obtain +>
Figure BDA00039771810600000214
Figure BDA00039771810600000215
S23: weighting and splicing value vectors based on the learned correlation weights of specific heads to obtain multi-behavior dependency semantics of users and projects
Figure BDA0003977181060000031
Figure BDA0003977181060000032
Wherein H represents the number of heads of the multi-head attention network, and I represents the splicing operation;
s24: adding the interaction features of the specific behaviors to the underlying multi-behavior-dependent semantics to obtain a final relational embedded representation
Figure BDA0003977181060000033
Figure BDA0003977181060000034
Further, step S3 sets a specific action weight beta for each user uk The following are provided:
Figure BDA0003977181060000035
w in the formula k Is the weight of the user u's K-th behavior, n uk Is the number of items user u interacted with under the kth behavior, Σ m∈N(r) n um Is the total number of items interacted with by user u and Σ m∈N(r) β uk ＝1。
Further, in step S4, a high-order embedded representation of the node is obtained by aggregating the convolution layers, so as to obtain a preference of the user, thereby performing multi-behavior recommendation, which specifically includes:
initial embedding
Figure BDA0003977181060000036
Firstly, converting the one-hot code into a sparse vector, then embedding the sparse vector into a space with smaller dimension, and carrying out multi-layer propagation to enable users, projects and relations to obtain embedded representations of different levels;
obtaining user-embedded representations of each behavior
Figure BDA0003977181060000037
Then according to the behavior weight beta uk For->
Figure BDA0003977181060000038
Weighted summation is performed to obtain the user representation +.>
Figure BDA0003977181060000039
The embedded representation of the item representation of layer L for each behavior is summed to:
Figure BDA00039771810600000310
Figure BDA00039771810600000311
combining the embedded representations of the different layers to obtain a final representation:
Figure BDA00039771810600000312
wherein each layer of representation is provided with a unified super parameter weight
Figure BDA00039771810600000313
Estimating the likelihood of user u having an interaction with item i under the kth behavior, y (K) u,i The method comprises the following steps:
Figure BDA0003977181060000041
further, the non-sampling loss function at the kth behavior is:
Figure BDA0003977181060000042
in the middle of
Figure BDA0003977181060000043
Representing the interactive item of user u under the Kth action, B representing the user of a batch, I representing the item set, the complexity of formula (11) is o ((|B|+|I|) d) 2 +|I k+ |d)；/>
The HCFMR adopts a multitasking training mode, and the objective function is as follows:
Figure BDA0003977181060000044
lambda in k The super parameters are designated for different types of data and are used for controlling the influence of the K-th behavior on the multi-task training; using L 2 Regularization prevents overfitting, μ being the regularization coefficient.
Further, mini-batch Adam was used as an optimizer to optimize the objective function.
Further, a message discarding method and a node discarding method are employed to prevent the neural network from being overfitted.
The invention has the beneficial effects that:
(1) By improving the lightweight graph convolution network to make multi-behavior recommendations, model complexity is reduced relative to previous graph convolution-based multi-behavior recommendations.
(2) By introducing a multi-head attention mechanism to grasp the dependence among different behaviors of users and projects, the embedded display of the users and the projects is enriched, and the recommendation quality is improved.
(3) In the fusion user representation, the importance of the behaviors and the interaction number of the user items are utilized to learn weights under different behaviors for the user, so that the interpretability of the model is improved.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objects and other advantages of the invention may be realized and obtained by means of the instrumentalities and combinations particularly pointed out in the specification.
Drawings
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in the following preferred detail with reference to the accompanying drawings, in which:
FIG. 1 is a diagram of a heterogeneous collaborative filtering model framework based on multi-behavior recommendation according to the present invention;
FIG. 3 is the effect of the number of convolution layers on the model;
FIG. 2 is a schematic diagram of a multi-head attention network architecture;
FIG. 4 is a graph of performance of different graph convolutions;
FIG. 5 is a graph of the impact of auxiliary behavior data;
FIG. 6 is a graph of the impact of user behavior weight and attention;
fig. 3 to 6 (a) and (b) are corresponding experimental graphs of experimental data sets beibeibeibei and Taobao, respectively.
Detailed Description
Other advantages and effects of the present invention will become apparent to those skilled in the art from the following disclosure, which describes the embodiments of the present invention with reference to specific examples. The invention may be practiced or carried out in other embodiments that depart from the specific details, and the details of the present description may be modified or varied from the spirit and scope of the present invention. It should be noted that the illustrations provided in the following embodiments merely illustrate the basic idea of the present invention by way of illustration, and the following embodiments and features in the embodiments may be combined with each other without conflict.
Wherein the drawings are for illustrative purposes only and are shown in schematic, non-physical, and not intended to limit the invention; for the purpose of better illustrating embodiments of the invention, certain elements of the drawings may be omitted, enlarged or reduced and do not represent the size of the actual product; it will be appreciated by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The same or similar reference numbers in the drawings of embodiments of the invention correspond to the same or similar components; in the description of the present invention, it should be understood that, if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc., that indicate an azimuth or a positional relationship based on the azimuth or the positional relationship shown in the drawings, it is only for convenience of describing the present invention and simplifying the description, but not for indicating or suggesting that the referred device or element must have a specific azimuth, be constructed and operated in a specific azimuth, so that the terms describing the positional relationship in the drawings are merely for exemplary illustration and should not be construed as limiting the present invention, and that the specific meaning of the above terms may be understood by those of ordinary skill in the art according to the specific circumstances.
Referring to FIG. 1, the present invention proposes a heterogeneous collaborative filtering model HCFMR (Heterogeneous collaborative filtering for multi-behavior recommendation) based on multi-behavior recommendation, representing the embedding of users (items, relationships) as
Figure BDA0003977181060000051
The interaction matrix of the user and the project is Y= { Y (1) 、Y (2) ...Y (k) K represents the total number of types of interactions, if the user has an interaction y (K) with the item under the K-th action u,i =1, otherwise 0, where y (k) u,i ∈Y (k) . Often, one target behavior (such as purchasing behavior in electronic commerce) most relevant to user preference is set in the multi-behavior recommendation, and the rest of behaviors are set as auxiliary behaviors (such as clicking browsing, adding shopping carts and the like). The task of HCFMR is to estimate the likelihood of user interaction with the item under target behavior +.>
Figure BDA0003977181060000052
Hcmr contains three modules in total: 1) The embedded propagation module is used for obtaining embedded expression vectors of users, items and relations of each order
Figure BDA0003977181060000061
2) A multi-behavior prediction module for estimating the possibility of interaction of user u with item i under the kth behavior +.>
Figure BDA0003977181060000062
3) And the multi-behavior optimization module optimizes model parameters through an advanced non-sampling learning method.
HCFMR is in a multiple relationship diagram
Figure BDA0003977181060000063
Upper modeling, wherein->
Figure BDA0003977181060000064
Representing nodes, edges, and sets of relationships, respectively. The model takes advantage of the Light GCN neighborhood aggregation and makes some improvements, respective embedded representations are obtained by joint embedding of nodes and relations, the representations of the nodes and the relations are updated in a mutual enhancement mode, the updating mode of a user is similar to that of a project, and specific representations of the user and the relations are shown as formulas (1) and (2):
Figure BDA0003977181060000065
Figure BDA0003977181060000066
in (i, r) k ) E N (u) represents the set of items i interacting with user u under the kth relationship,
Figure BDA0003977181060000067
is a normalization operation, and can avoid the increase of the embedding scale with the increase of the number of layers of the picture volume. />
Figure BDA0003977181060000068
Is a layer-specific trainable parameter matrix that projects all relationships to the same embedding space as the node and allows for use in the next GCN layer. />
Figure BDA0003977181060000069
Is a combined function, and the relation embedded representation is integrated into the node representation to strengthen the node representation. The partial derivative calculation may be calculated from an automatic differential packet in Pytorch or TensorFlow. Because the user has various behavior data, the neighbor features under specific relations are respectively aggregated, and the neighbor features of different orders of the user are obtained through L-order propagation of the graph rolling network
Figure BDA00039771810600000610
Relational embedding representation comprising interaction features of a specific behavior>
Figure BDA00039771810600000611
Multiple behavior-dependent semantics with the bottom layer +.>
Figure BDA00039771810600000612
As shown in FIG. 2, since multi-head attention has strong capability of mining data relevance, the multi-head attention is introduced into a multi-behavior recommendation system to capture potential dependency semantics among user and project multi-behaviors. First, a query, key, value transformation matrix Q, K is defined,
Figure BDA00039771810600000613
For the relationship k and->
Figure BDA00039771810600000614
Projection between embedded representations, query vector +.>
Figure BDA00039771810600000615
And key vector->
Figure BDA00039771810600000616
The relevance weight between the relation embeddings is obtained by scaling the dot product of the query vector and all key vectors>
Figure BDA00039771810600000617
Finally, normalization treatment is carried out to the extract to obtain->
Figure BDA00039771810600000618
The specific representation is as follows:
Figure BDA0003977181060000071
then the weighted and spliced value vectors based on the learned correlation weight of the specific head are used for obtaining the multi-behavior dependency semantics of the user and the project
Figure BDA0003977181060000072
Figure BDA0003977181060000073
Wherein H represents the number of heads of the multi-head attention network, I represents the splicing operation, and finally, the interactive features of the specific behaviors and the underlying multi-behavior dependency semantics are added to obtain the final relation embedded representation
Figure BDA0003977181060000074
Figure BDA0003977181060000075
In multi-behavior recommendation, the contribution degree of different behaviors in the user preference prediction needs to be considered, for example, the purchase behavior is obviously the largest contribution degree in the e-commerce recommendation, but other behaviors may play a larger role when the user has few purchase records. The model thus needs to learn automatically from different usersContribution degree of different behaviors, HCFMR distributes different behavior weights beta for each user uk The expression is as follows:
Figure BDA0003977181060000076
w in the formula k Is the weight of the user u's K-th behavior, n uk Is the number of items user u interacted with under the kth behavior, Σ m∈N(r) n um Is the total number of items interacted with by user u and Σ m∈N(r) β uk ＝1。
Obtaining user-embedded representations of each behavior through an embedded propagation layer
Figure BDA0003977181060000077
Then according to the behavior weight beta uk For->
Figure BDA0003977181060000078
Weighted summation is performed to obtain the user representation +.>
Figure BDA0003977181060000079
The embedded representation of the item representation of layer L for each behavior is summed to:
Figure BDA00039771810600000710
Figure BDA00039771810600000711
the multi-behavior prediction module: initial HCF-MR embedding
Figure BDA00039771810600000712
Firstly, the sparse vector is converted into a sparse vector through one-hot coding, and then the sparse vector is embedded into a space with smaller dimension. Through multi-layer propagation, users, items, relationships obtain embedded representations of different levels, which representInformation aggregated from different orders, the embedded representations of the different layers are combined to obtain the final representation: />
Figure BDA0003977181060000081
Wherein each layer of representation is provided with a uniform weight
Figure BDA0003977181060000082
This is a hyper-parameter that enables the model to achieve better performance. Estimating the likelihood of user u having an interaction with item i under the kth behavior, y (K) u,i The method comprises the following steps:
Figure BDA0003977181060000083
the multi-behavior optimization module: the negative sampling training method is high in efficiency and easy to realize, is a preferred optimization method for most of works, and has the characteristics of instability and high randomness. If negative sampling is performed in the multi-feedback data, it samples a negative case for each interaction, the negative case generated in the multi-feedback data is approximately K times (K is the number of behavior types) of single feedback, thus leading to multiple randomness and making model optimization difficult, while the non-sampling method considers all data to avoid the problems and ensure better robustness of the model, and the non-sampling loss function under the K-th behavior is:
Figure BDA0003977181060000084
in the middle of
Figure BDA0003977181060000085
Representing the interactive item of user u under the Kth action, B representing the user of a batch, I representing the item set, the complexity of formula (11) is o ((|B|+|I|) d) 2 +|I k+ I d), in order to learn parameters from the user's multi-line data, hcmr adopts a multitasking training method, whose objective function is:
Figure BDA0003977181060000086
lambda in k Is a hyper-parameter specified for different types of data in order to control the impact of the kth behavior on the multitasking training. To prevent overfitting use L 2 Regularization, μ is a regularization coefficient. For optimizing the objective function, mini-batch Adam is selected as an optimizer, and the main advantage of the optimization method is that the learning rate can be self-adaptive in the training stage. Dropout is an effective solution to prevent overfitting of the neural network, hcmr employs two discarding methods: message discard and node discard.
Experiment verification is carried out by using an experiment data set Beibei and Taobao, wherein the experiment data set comprises three types of data of click browsing, shopping cart adding and purchasing, wherein purchasing is recommended target behavior, and specific data are shown in table 1. The two data sets filter out users and items purchased less than 5 times, the last purchase record of the user is used as test data, the penultimate record is used as verification data, and the rest of the data is used for training.
TABLE 1
Figure BDA0003977181060000091
The model uses two commonly used evaluation criteria Hit rates (Hit Ratio, HR) and normalized cumulative discount gains (Normalized Discounted Cumulative Gain, NDCG). HR is an indicator that emphasizes the accuracy of the recommendation and measures whether the items in the test set are in the user's TOP-K recommendation list. NDCG is an importance that emphasizes the order of hit items in TOP-K recommendation list, i.e. the higher the hit item position in the recommendation list the better the recommendation quality.
To evaluate hcmr performance, the comparative baselines are divided into two categories: the single behavior recommendation model and the multi-behavior recommendation model, wherein the single behavior recommendation model is compared with the multi-behavior recommendation model, and the single behavior recommendation model comprises the following components:
(1) NCF: a neural collaborative filtering model with multiple implementations;
(2) Light GCN: a graph collaborative filtering model that simplifies GNN, has higher performance in a single row of recommendations;
among the recommendation models for multi-behavior are:
(1) NMTR: cascading multiple behaviors by applying multiple NCF modules to a multitasking learning framework and capturing finer information;
(2) SGCNMB: the multi-behavior recommendation is realized by applying a plurality of Light GCN modules to a multi-task learning framework;
(3)EHCF [ : each behavior prediction is associated through a migration learning mode, and a non-sampling mode is adopted to conduct multi-behavior recommendation;
(4)GHCF [7] : the user item embedding and the relation embedding are combined to predict the multi-behavior by improving the traditional GCN, and advanced non-sampling learning is combined to enable the model to achieve higher performance;
in the experiment, the HCFMR model searches the optimal parameters on the verification set by using a grid search method, and evaluates the optimal parameters on the test set, wherein all the baselines are initialized according to the corresponding paper parameters so as to achieve the optimal parameters. epoch is set to 500, batch-Size is set to 256, embedding dimension d is set to 64, learning rate is set to 0.001, and the number of layers of the convolution on the two data sets is from [1,2,3,4,5 ]]In a multi-head attention network, the number of heads H is set to 2, and the length of a recommendation list is [10,50,100 ]]. Parameter lambda in a multitasking loss function 1 、λ 2 、λ 3 From [0,1/6,2/6,3/6,4/6,5/6,1]And lambda is adjusted by 1 +λ 2 +λ 3 =1, two of which are determined to be the other, the optimal combination in HCF-MR is [1/6,4/6,1/6 ]]。
The results of the comparative experiments on the two data sets are shown in table 2, and are specifically analyzed as follows:
TABLE 2
Figure BDA0003977181060000101
Table 2 shows the results of the individual models on two data sets, it can be seen that:
(1) All the metrics of HCF-MR on both datasets are better than the optimal baseline GHCF, with an average improvement of 4.07% and 6.4% on the beibeibei and Taobao datasets, mainly due to the results of hcmr modeling the different behavioral dependencies and contributions of the user. The SGCNMB models the bipartite graph of each type of data, and respectively uses the Light GCN to conduct neighborhood feature aggregation, compared with HCFMR, the SGCNMB lacks explicit modeling of high-order heterogeneous paths of users and projects, and meanwhile does not pay attention to behavior dependence and importance, so that the effect of multi-behavior data is greatly limited, the capability of extracting fine granularity semantics of the users and the projects is insufficient, and the performance on user preference prediction is poor. The recommendation effect of the first-order interaction information of the user and the project is obviously inferior to that of the HCFMR and the GHCF by using the neural network model NMTR and the matrix decomposition model EHCF, which shows that the graph convolution-based method can fully utilize the high-order neighbors of the user and the project, and deep mining of the high-order interaction information is more advantageous for predicting the user preference. The effectiveness of the HCFMR in inducing multiple heads of attention to explore the contribution degree of different behaviors of users by setting different behavior weights and distinguishing the different behaviors is shown.
(2) All the multi-behavior recommendation models have better effects than the single-behavior recommendation models, which shows that the introduction of multi-behavior data helps to improve recommendation performance because different behaviors have complementarity and reflect interaction modes between users and items from different angles. For a user with less target behavior data, the auxiliary behavior data of the user can ensure that the user obtains better recommendation effect. The multi-behavior recommendation models NMTR and SGCNMB are improved from the single-row recommendation models NCF and Light GCN, respectively, which can further illustrate the advantage of introducing multiple rows of data to promote the recommendation effect.
(3) For all evaluation criteria, all models performed better on the beibeibeiei dataset than the Taobao dataset, which is attributed to the greater sparsity of the latter, resulting in limited information extraction capabilities of the models.
In order to evaluate the model complexity of the hcmr, the average single training time consumption of a plurality of multi-behavior recommendation models on two data sets is recorded under the same experimental condition, and as can be seen from the results of table 3, the average single training time consumption of the hcmr is reduced by 9.1% and 16.5% compared with the average single training time consumption of the optimal baseline GHCF on the Beibei and Taobao data sets, and the main reason is that the hcmr performs neighborhood polymerization iteration based on a lightweight graph convolution network, so that the model parameter number is greatly reduced, and the training difficulty is reduced. While also demonstrating the effectiveness of improving the lightweight graph convolutional network. The three models are more time consuming on the Taobao dataset than the Beibei dataset, mainly because of the greater number of users, items of the Taobao dataset.
TABLE 3 Table 3
Figure BDA0003977181060000111
To investigate the effect of the number of layers of the graph convolution on the model, hcmr was set to different depths, and the performance results for hr@100 on both datasets are shown in fig. 3 (a) (b).
From the results of fig. 3, it can be seen that, as the performance of the superimposed hcmr with the convolution layer number also increases, the optimum is achieved when the number of the graph convolution layers increases to layer 3 on the beibeibei data set, which is improved by 2.0% compared with the single-layer convolution performance, and the optimum is achieved when the number of the graph convolution layers increases to layer 4 on the Taobao data set, which is improved by 4.6% compared with the single-layer convolution performance. And then the performance of the model starts to decline along with the increase of the number of convolution layers, which shows that stacking proper convolution layers is helpful for exploring the higher-order neighbor information of the user item, the recommendation effect is improved, and particularly for a sparse data set, excessive overlapped convolution layers bring about certain noise and weaken the performance of the model.
The improvement of HCFMR is based on a lightweight graph convolution network, and meanwhile, the multi-relation graph neural network R-GCN for processing entity and relation embedding in a knowledge graph is used for reference [16] Comp-GCN [17] The relationships are integrated into the embedded representation of the node, and are better applied to multi-behavior recommendation. Therefore, to verify the success of HCAMR improvement, the two methods described above were substituted for the graph rolling network of the present invention, experimental results on two data setsAs in (a) (b) of fig. 4.
From the results of fig. 4, it can be seen that the hcmr improvement is best, with a 27.1% and 33.8% improvement over R-GCN on both data sets, respectively, first. This is because the R-GCN model can top a parameter matrix for each relationship, introducing too many relationship matrices as the number of relationships increases results in the model being unable to train. The percentage of improvement over Comp-GCN on both data sets was 4.0%, 5.0%. The HCFMR adopts an iterative updating mode of mutual enhancement of the node and the relation, and gives consideration to the relation expression when the relation is integrated into the node expression, so that stronger relation expression is required to obtain multi-behavior dependency semantics of a user bottom layer in the multi-head attention network.
To explore the effect of auxiliary behavior data on experiments, several variant experiments were performed on hcmr on two data sets, with hcmr-P containing only purchase behavior data, hcmr-PV containing purchase and click-through data, and hcmr-PC containing purchase and addition shopping cart data, the results on both data sets being shown in fig. 5 (a) (b).
The effect of the variant model using the helper behavior data is significantly better than the model hcmr-P using the target behavior data alone as a whole from fig. 5. The model hcmr using three data improved 22.3% and 31.8% over hcmr-P using only the purchased data on the beibeibei and Taobao datasets, respectively. This suggests that the introduction of auxiliary behavior data may have a positive effect on the model. The effect of adding shopping cart behavior in both data sets is lower than click-through behavior, which may be due to differences in auxiliary behavior data size.
To evaluate the effectiveness of the user behavior weights and attention modules of hcmr, three variant models of hcmr-1 were proposed: only the graph roll-up network module hcmr-2 is reserved: the user uses uniform behavior weight under different behaviors, HCFMR-3: the attention module was removed and the experimental results on both datasets were as in (a) (b) of fig. 6.
From the results of fig. 6, it can be seen that the performance of the model hcmr without ablation on both data sets is superior to the three variant models. The hcmr-1 retaining only the graph convolution module improved the mean 2.8% and 3.3% over the beibeibei and Taobao datasets compared to the model hcmr without ablation, indicating the effectiveness of introducing user behavioral weights and attention. Notably, the performance of HCMR-3 on both datasets is slightly more pronounced compared to the effect of HCMR-1 on HCMR-2, HCMR-3 in both datasets, indicating that the introduction of user behavior weights is more conducive to improving model performance.
Finally, it is noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made thereto without departing from the spirit and scope of the present invention, which is intended to be covered by the claims of the present invention.

Claims (8)

1. A heterogeneous collaborative filtering method for multi-behavior recommendation is characterized in that: heterogeneous collaborative filtering model HCFMR based on multi-behavior recommendation comprises the following steps:
s1: the improved lightweight graph convolutional network is utilized to obtain user and project interaction characteristics of specific behaviors;
s2: multiple behavior dependent semantics of a user bottom layer are obtained through a multi-head attention network;
s3: setting a weight of a specific behavior for each user, and distinguishing the contribution degree of different behaviors;
s4: and obtaining the high-order embedded representation of the node by aggregating all the convolution layers to obtain the preference of the user, thereby performing multi-behavior recommendation.
2. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 1, wherein: the step S1 of obtaining the multi-behavior-dependent semantics of the user bottom layer through the multi-head attention network specifically comprises the following steps:
based on a lightweight graph convolutional network Light GCN, joint embedding is carried out on nodes and relations to obtain respective embedded representations, the representations of the nodes and the relations are updated in a mutually enhanced mode, and specific representations of users and the relations are shown as formulas (1) and (2):
Figure FDA0003977181050000011
Figure FDA0003977181050000012
in (i, r) k ) E N (u) represents the set of items i interacting with user u under the kth relationship,
Figure FDA0003977181050000013
is a normalization operation; />
Figure FDA0003977181050000014
Is a layer-specific matrix of trainable parameters; />
Figure FDA0003977181050000015
Is a combined function for integrating the relational embedding representation into the node representation, enhancing the node representation;
respectively aggregating neighbor features under specific relations, and obtaining neighbor features of different orders of users through L-order propagation of graph rolling network
Figure FDA0003977181050000016
Relational embedded representation contains interactive features of specific behavior
Figure FDA0003977181050000017
Multiple behavior dependent semantics with the bottom layer->
Figure FDA0003977181050000018
3. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 1, wherein: step S2, obtaining multi-behavior-dependent semantics of the user bottom layer through the multi-head attention network specifically includes:
s21: defining a query, key, value transformation matrix Q, K,
Figure FDA0003977181050000019
For the relationship k and->
Figure FDA00039771810500000110
Embedding projections between representations;
s22: will query vectors
Figure FDA00039771810500000111
And all key vectors->
Figure FDA00039771810500000112
Scaling dot product to obtain correlation weight between relation embeddings
Figure FDA00039771810500000113
And normalize it to obtain +>
Figure FDA00039771810500000114
/>
Figure FDA0003977181050000021
S23: weighting and splicing value vectors based on the learned correlation weights of specific heads to obtain multi-behavior dependency semantics of users and projects
Figure FDA0003977181050000022
Figure FDA0003977181050000023
Wherein H represents the number of heads of the multi-head attention network, and I represents the splicing operation;
s24: adding the interaction features of the specific behaviors to the underlying multi-behavior-dependent semantics to obtain a final relational embedded representation
Figure FDA0003977181050000024
Figure FDA0003977181050000025
4. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 1, wherein: step S3, setting a weight beta of a specific behavior for each user uk The following are provided:
Figure FDA0003977181050000026
w in the formula k Is the weight of the user u's K-th behavior, n uk Is the number of items user u interacted with under the kth behavior, Σ m∈N(r) n um Is the total number of items interacted with by user u and Σ m∈N(r) β uk ＝1。
5. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 1, wherein: step S4, obtaining a high-order embedded representation of the node by aggregating the convolution layers, to obtain a preference of the user, thereby performing multi-behavior recommendation, including:
initial embedding
Figure FDA0003977181050000027
Firstly, converting the one-hot code into a sparse vector, then embedding the sparse vector into a space with smaller dimension, and carrying out multi-layer propagation to enable users, projects and relations to obtain embedded representations of different levels;
obtaining user-embedded representations of each behavior
Figure FDA0003977181050000028
Then according to the behavior weight beta uk For->
Figure FDA0003977181050000029
Weighted summation is performed to obtain the user representation +.>
Figure FDA00039771810500000210
The embedded representation of the item representation of layer L for each behavior is summed to:
Figure FDA00039771810500000211
Figure FDA00039771810500000212
combining the embedded representations of the different layers to obtain a final representation:
Figure FDA0003977181050000031
/>
wherein each layer of representation is provided with a unified super parameter weight
Figure FDA0003977181050000032
Estimating the likelihood of user u having an interaction with item i under the kth behavior, y (K) u,i The method comprises the following steps:
Figure FDA0003977181050000033
6. the heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 1, wherein: the non-sampling loss function at the kth behavior is:
Figure FDA0003977181050000034
in the middle of
Figure FDA0003977181050000036
Representing the interactive item of user u under the Kth action, B representing the user of a batch, I representing the item set, the complexity of formula (11) is o ((|B|+|I|) d) 2 +|I k+ |d)；
The HCFMR adopts a multitasking training mode, and the objective function is as follows:
Figure FDA0003977181050000035
lambda in k The super parameters are designated for different types of data and are used for controlling the influence of the K-th behavior on the multi-task training; using L 2 Regularization prevents overfitting, μ being the regularization coefficient.
7. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 6, wherein: the mini-batch Adam was used as an optimizer to optimize the objective function.
8. The heterogeneous collaborative filtering method for multi-behavior recommendation according to claim 7, wherein: a message discarding method and a node discarding method are adopted to prevent the neural network from being over fitted.
CN202211540250.5A 2022-12-02 2022-12-02 Heterogeneous collaborative filtering method for multi-behavior recommendation Pending CN116167812A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211540250.5A CN116167812A (en) 2022-12-02 2022-12-02 Heterogeneous collaborative filtering method for multi-behavior recommendation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211540250.5A CN116167812A (en) 2022-12-02 2022-12-02 Heterogeneous collaborative filtering method for multi-behavior recommendation

Publications (1)

Publication Number Publication Date
CN116167812A true CN116167812A (en) 2023-05-26

Family

ID=86415258

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211540250.5A Pending CN116167812A (en) 2022-12-02 2022-12-02 Heterogeneous collaborative filtering method for multi-behavior recommendation

Country Status (1)

Country Link
CN (1) CN116167812A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116645174A (en) * 2023-07-27 2023-08-25 山东省人工智能研究院 Personalized recommendation method based on decoupling multi-behavior characterization learning
CN117171448A (en) * 2023-08-11 2023-12-05 哈尔滨工业大学 Multi-behavior socialization recommendation method and system based on graph neural network
CN117171448B (en) * 2023-08-11 2024-05-28 哈尔滨工业大学 Multi-behavior socialization recommendation method and system based on graph neural network

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116645174A (en) * 2023-07-27 2023-08-25 山东省人工智能研究院 Personalized recommendation method based on decoupling multi-behavior characterization learning
CN116645174B (en) * 2023-07-27 2023-10-17 山东省人工智能研究院 Personalized recommendation method based on decoupling multi-behavior characterization learning
CN117171448A (en) * 2023-08-11 2023-12-05 哈尔滨工业大学 Multi-behavior socialization recommendation method and system based on graph neural network
CN117171448B (en) * 2023-08-11 2024-05-28 哈尔滨工业大学 Multi-behavior socialization recommendation method and system based on graph neural network

Similar Documents

Publication Publication Date Title
Liu et al. Deep reinforcement learning based recommendation with explicit user-item interactions modeling
CN111310063B (en) Neural network-based article recommendation method for memory perception gated factorization machine
CN110910218B (en) Multi-behavior migration recommendation method based on deep learning
CN108648049A (en) A kind of sequence of recommendation method based on user behavior difference modeling
CN111523047A (en) Multi-relation collaborative filtering algorithm based on graph neural network
CN112364976B (en) User preference prediction method based on session recommendation system
CN108563755A (en) A kind of personalized recommendation system and method based on bidirectional circulating neural network
CN110955826B (en) Recommendation system based on improved cyclic neural network unit
CN112950324B (en) Knowledge graph assisted pairwise sorting personalized merchant recommendation method and system
CN113918833B (en) Product recommendation method realized through graph convolution collaborative filtering of social network relationship
CN113918832B (en) Graph convolution collaborative filtering recommendation system based on social relationship
CN112925977A (en) Recommendation method based on self-supervision graph representation learning
CN113918834B (en) Graph convolution collaborative filtering recommendation method fusing social relations
CN112258262A (en) Conversation recommendation method based on convolution self-attention network
CN110059251B (en) Collaborative filtering recommendation method based on multi-relation implicit feedback confidence
CN116071128A (en) Multitask recommendation method based on multi-behavioral feature extraction and self-supervision learning
CN116167812A (en) Heterogeneous collaborative filtering method for multi-behavior recommendation
CN116401542A (en) Multi-intention multi-behavior decoupling recommendation method and device
Luo et al. BGNN: Behavior-aware graph neural network for heterogeneous session-based recommendation
CN114676315A (en) Method and system for constructing attribute fusion interaction recommendation model based on enhanced graph convolution
Cao et al. Implicit user relationships across sessions enhanced graph for session-based recommendation
CN114861072B (en) Graph convolution network recommendation method and device based on interlayer combination mechanism
Wang et al. Jointly modeling intra-and inter-transaction dependencies with hierarchical attentive transaction embeddings for next-item recommendation
Zhang et al. Graph spring network and informative anchor selection for session-based recommendation
CN113704439B (en) Conversation recommendation method based on multi-source information heteromorphic graph

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