A socialized recommendation method and device based on user multi-intention and a medium

Abstract

The application discloses a kind of socialized recommendation method, device and medium based on user multi-intention, wherein method includes: obtaining data set, constructs graph structure data;Build the recommendation model containing two parts graph convolution network, for modeling user commodity interaction relationship and user social relationship in turn;Intention space mapping;Interaction intention decoupling learning;Self-adapting denoising;User social multi-influence learning;Unsupervised learning;Combination obtained final commodity feature vector, final user feature vector and unsupervised learning to train and learn recommendation model;The user and commodity feature vector that recommendation model finally learns are scored and predicted, and the recommended commodity order is obtained.The application fully mines the multi-intention influence in user commodity interaction, and the multi-influence of user social relationship, learns more fine-grained user representation and commodity representation, and obtains more accurate recommendation result.The application can be widely applied in the technical field of recommendation system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical fields of artificial intelligence, deep learning and recommendation system, and particularly relates to a socialized recommendation method and device based on user multi-intention and a medium. BACKGROUND

[0002] With the rapid development of the Internet, the information on the network presents explosive growth, and the recommendation system can well solve the problem of "information overload", so the development of the recommendation system is more and more concerned by people. The improvement of economic level and the development of network technology promote the development of online shopping, online viewing and online reading and other network applications. These applications contain a large amount of user and commodity data. The mainstream recommendation algorithm today mainly adopts the method of collaborative filtering, which can effectively filter a large amount of data to obtain truly useful information and realize personalized recommendation. As an important branch of the recommendation system, socialized recommendation has attracted more and more attention in recent years with the rapid development of social media. However, the current socialized recommendation algorithm still has three important challenges that have not been well solved.

[0003] (1) In the current socialized recommendation algorithm, the interaction relationship between users and commodities is often considered singlely, and the intention influence behind the interaction behavior of users and commodities is not mined, which usually leads to suboptimal user features and commodity features learned, so it is crucial to decouple the intention factors behind the user and commodity interaction.

[0004] (2) The current socialized recommendation algorithm also often considers the social relationship of users singlely, and does not mine the multiple influences between users. In real life, the social influence between users is complex, for example, when the social friends are students, they may buy the same commodity because of the price factor due to the social influence, but when the social friends have worked, they may buy the same commodity due to the social influence because of factors other than price. Therefore, it is also very important to decouple the multiple influences between users.

[0005] (3) The current social relationship often has noise and also has a data sparsity problem, so it is very important to effectively remove the noise edges in the social relationship and use appropriate self-supervised learning to enhance the data. SUMMARY

[0006] In order to at least partially solve one of the technical problems in the prior art, the purpose of the present application is to provide a socialized recommendation method, device and medium based on user multi-intention.

[0007] The technical solution adopted by the present application is:

[0008] A socialized recommendation method based on user multi-intention, comprising the following steps:

[0009] Collecting a data set with user commodity interaction relationship and user social relationship;

[0010] Storing data with sparse matrix, constructing graph structure data that can be processed by graph convolution network;

[0011] Building a recommendation model containing two parts of graph convolution network, for modeling user commodity interaction relationship and user social relationship in turn;

[0012] Intention space mapping: using multilayer perceptron to map initial user feature vector and initial commodity feature vector to different intention space, in order to mine multi-intention influence;

[0013] Interaction intention decoupling learning: building a graph neural network in each intention space for representation learning, to get user feature vector and commodity feature vector under each intention; for commodity node, summing up commodity feature vector under each intention to get final commodity feature vector;

[0014] Adaptive denoising: using adaptive Dropout module to remove irrelevant edges in social relationship graph, to retain important edges;

[0015] User social multi-influence learning: according to the obtained user feature vector under each intention, using graph neural network to model user social graph under each intention, summing up the obtained user feature vector under each intention to get final user feature vector;

[0016] Unsupervised learning: introducing two adaptive Dropout modules to get two social views for contrastive learning;

[0017] Combining the obtained final commodity feature vector, final user feature vector and unsupervised learning to train and learn the recommendation model;

[0018] Scoring and predicting the user and commodity feature vectors learned by the recommendation model finally, to get the recommended commodity order.

[0019] Further, the socialized recommendation method further comprises the step of preprocessing the data set:

[0020] Filtering invalid users and commodities according to preset conditions, and retaining valid user and commodity nodes;

[0021] Dividing the data set, randomly selecting two interaction items of the user, and putting them into the validation set and the test set respectively, and the remaining interaction items are the training set.

[0022] Further, the intention space mapping comprises:

[0023] Using a multi-layer perceptron, the randomly initialized feature vector is mapped to different intent spaces:

[0024] E f = MLP(E)

[0025] H f = MLP(H)

[0026] where E represents the total user feature vector, H represents the total commodity feature vector, and f represents the fth intent.

[0027] Further, the decoupled intent interaction learning includes:

[0028] Under each intent, the initial adjacency matrix corresponding to the user-commodity interaction is assigned a value of 1, representing that under the initial state, each intent has the same influence on user-commodity interaction.

[0029] The user feature vector and the commodity feature vector are input into the graph convolution network for K iterations, i.e., message passing and message aggregation, to continuously update the adjacency matrix of each intent, so that important intents can be decoupled.

[0030] Using a multi-layer graph convolution network, high-order neighbor information is obtained; using an average operation, the feature vectors learned by each layer are aggregated, and the formula is as follows:

[0031]

[0032]

[0033] wherein, represents the user feature vector of the fth intent in the lth network, represents the commodity feature vector of the fth intent in the lth network; E f represents the user feature vector under the fth intent, H f represents the user feature vector under the fth intent.

[0034] Finally, a summation operation is used to sum the commodity feature vectors under each intent to obtain the final commodity feature vector.

[0035] Further, the adaptive denoising includes:

[0036] A multi-layer perceptron is used to calculate the dropout coefficient m ij of each social edge, which determines whether to discard the social edge; the specific calculation formula is as follows:

[0037] D1(i,i') = MLP(e i ||e i′})

[0038] In the formula, || represents splicing operation;

[0039] In order to make the whole model can be trained end to end, the reparameterization technique is adopted, a mask matrix M1 of the noise edge in the social graph is constructed; the calculation formula is as follows:

[0040] M1(i,i')=σ((logσ-log(1-σ)+M(i,i') / τ)

[0041] Wherein, sigma represents sigmoid activation function, tau is temperature coefficient, i and i' represent two users on the social edge respectively.

[0042] Further, the user social multi-intention learning comprises:

[0043] Under each intention, the weight coefficient of the social edge is calculated by using the attention mechanism, the weight values of the same social edge under different intentions are normalized by using the softmax, then the dot product operation is performed with the mask matrix, the weight of the noise edge is removed, the final social adjacency matrix is obtained, then the message passing and message aggregation are performed in the graph neural network; the multi-layer graph neural network is adopted to capture high-order neighbor information; the formula is as follows:

[0044]

[0045] Wherein, E1 represents the user feature vector learned by the first social denoising view, and the user feature vector E2 is also learned from the second social denoising view;

[0046] The average operation is adopted to aggregate the feature vectors of the two views to obtain the final user feature vector.

[0047] Further, the obtained final commodity feature vector, final user feature vector and unsupervised learning are combined to train and learn the recommendation model, comprising:

[0048] The sum of the supervised loss and the unsupervised loss is used as the target loss function of the recommendation model, wherein the supervised loss is the loss function in the node feature vector learning, and the unsupervised loss is the loss function in the unsupervised learning;

[0049] The gradient descent method is used to update the parameters of the recommendation model until the target loss function reaches the preset threshold.

[0050] Another technical solution adopted by the application is:

[0051] A social recommendation device based on user multi-intention, comprising:

[0052] At least one processor;

[0053] at least one memory for storing at least one program;

[0054] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0055] Another technical solution adopted by the present application is:

[0056] A computer-readable storage medium, wherein a processor-executable program is stored, the processor-executable program is used to execute the method as described above when executed by a processor.

[0057] The beneficial effects of the present application are: the present application fully excavates the multi-intention influence in the user commodity interaction and the multi-influence of the user social relationship, learns the user representation and the commodity representation of finer granularity, and obtains more accurate recommendation results; meanwhile, the adaptive Dropout mode is introduced, the social edges in the social graph are adaptively discarded, the social edges irrelevant to the task are deleted, two social views are constructed, then the graph convolution network is used to learn two user feature vectors for self-supervised contrast learning, the data sparsity problem is alleviated, and the robustness and generalization performance of the model are further improved. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following introduces the drawings of the related technical solutions in the embodiments of the present application or the prior art. It should be understood that the drawings in the following introduction are only for the convenience of clearly describing part of the embodiments of the technical solutions of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the premise.

[0059] Figure 1 is a workflow diagram of a socialized recommendation method based on user multi-intention and self-supervised learning in an embodiment of the present application;

[0060] Figure 2 is a framework diagram of a socialized recommendation model in an embodiment of the present application. DETAILED DESCRIPTION

[0061] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application. For the step numbers in the following embodiments, they are only set for the convenience of description and explanation, and the order between the steps is not limited in any way, and the execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0062] In the description of the present application, it should be understood that the orientation description, such as the orientation or position relationship indicated by up, down, front, back, left, right and the like, is based on the orientation or position relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0063] In the description of the present application, the meaning of several is one or more, the meaning of multiple is more than two, greater than, less than, more than, etc. are understood as not including the number, above, below, within, etc. are understood as including the number. If the first, second is described, it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features or the order of the indicated technical features.

[0064] In the description of the present application, unless otherwise explicitly limited, the words such as setting, installing, connecting, etc. should be broadly understood, and the person skilled in the art can reasonably determine the specific meaning of the above words in the present application in combination with the specific content of the technical solution.

[0065] As shown in Figure 1 and Figure 2 The embodiment provides a social recommendation method based on user multi-intention and self-supervised learning. A perception machine is used to map a single user feature vector and a commodity feature vector to multiple intention spaces. Then, a two-part graph convolution network is used to model a user commodity interaction graph and a social graph in sequence, and to mine multi-intention influence. Meanwhile, when modeling the social graph, an adaptive Dropout module is used to discard irrelevant edges of the social graph, two social views are constructed for comparative learning, and the data sparsity problem is alleviated. This method can fully mine multi-intention influence in user commodity interaction behavior, mine different influences in social relationships, obtain user representations and commodity representations closer to reality, and thus achieve better recommendation effect. Meanwhile, because self-supervised learning is adopted, the model can obtain better robustness and generalization. The method specifically comprises the following steps:

[0066] S1, model data preparation. User-item interaction data and user social data are collected from public datasets CiaoDVD, Yelp, Brightkite, and the data is preprocessed. The preprocessing includes: removing user nodes and item nodes with less than 5 user-item interactions, and excluding users with social relationships of 0, to avoid the cold start problem of user and item data, and to affect the model training effect. Then the data set is divided, two interaction items of the user are randomly selected and put into the validation set and test set respectively, and the remaining interaction items are the training set. In order to verify the effect of the model on the validation set and the test set, according to the 1:100 positive and negative sample ratio, 100 negative samples are randomly selected for each positive sample in the validation set and the test set. Then the user-item interaction relationship and the user social relationship are stored in the form of a sparse matrix for graph neural network modeling. Finally, a feature vector is randomly initialized for each user node and item node for subsequent model learning. The initial feature vector is represented as:

[0067] E = {e1, …, ei, …, ei} i ,…,e I}

[0068] H = h1, …, hj, …, hj} j ,…,h J}

[0069] Where E represents the total user feature vector, H represents the total item feature vector; e i represents a user feature vector, a total of I users; h i represents a product feature vector, a total of J products;

[0070] S2, user-item interaction intention decoupling learning. Mainly includes intention space mapping of feature vectors, intention decoupling graph neural network for iterative updating of feature vectors, and multi-layer graph neural network output aggregation.

[0071] Wherein, step S2 specifically includes steps S21-S23:

[0072] S21, preset the number of intentions, construct the corresponding multi-layer perceptron according to the number of intentions, and then use the multi-layer perceptron to perform multi-intention space mapping on a single user feature vector and item feature vector, as shown in the following formula:

[0073] E f = MLP f (E)

[0074] e = {e1, …, ei, …, ei} f ,…e F}

[0075] H f = MLPf (H)

[0076] h = {h1, …, h f ,…,h F}

[0077] where F represents the total number of intentions, f represents the fth intention; E f represents the user feature vector of the fth intention, H f represents the commodity feature vector of the fth intention.

[0078] S22, under each intention, a graph neural network is designed to perform message passing and message aggregation, and the influence of the intention on the user-commodity interaction is adaptively obtained. Adopting the iterative update principle, there will be K iterations in one GCN network, k ∈ {1, …, K}. At the beginning, the initial value of the user-commodity adjacency matrix under each intention is set to 1, denoted as G f (e, h) = 1, representing that the influence of each intention on the user-commodity interaction behavior is the same in the initial state.

[0079] Then before each iteration, it is necessary to keep the distribution stable under all intentions, fully excavate important intentions, and normalize these weight values using the softmax function. Then message passing and aggregation are performed on the graph neural network. Then the iteration update operation is performed. The formula of the above operation is as follows:

[0080]

[0081]

[0082]

[0083] wherein, represents the value of the adjacency matrix corresponding to the fth intention at the kth iteration, tanh is a nonlinear activation function.

[0084] After K iterations in one interaction intention decoupling layer, the total user feature vector and commodity feature vector need to be updated to fuse the information in multiple intention spaces, so that the total feature vector can be mapped to different intention spaces in the next operation. The formula is as follows:

[0085]

[0086]

[0087] wherein, l represents the total number of graph neural networks, l represents the total number of layers.

[0088] S23, using multi-layer graph convolution network, obtaining high-order neighbor information. The learned feature vector of each layer is aggregated by using the average operation. Finally, the product feature vector under each intent is summed to obtain the final product feature vector, as follows:

[0089]

[0090]

[0091]

[0092] wherein, represents the user feature vector of the fth intent in the lth layer network, represents the product feature vector of the fth intent in the lth layer network; E f represents the user feature vector under the fth intent, H f represents the user feature vector under the fth intent.

[0093] S3, adaptive denoising module. The social relationship of the user usually has noise problems, and reasonable removal of these noise relationships is crucial for effective use of social relationships. We use a multi-layer perceptron to calculate the dropout coefficient m ij of each social edge to determine whether to discard the social edge. The calculation formula is as follows:

[0094] D1(i,i') = MLP(e i ||e i′})

[0095] wherein, || represents splicing operation.

[0096] In order to make the whole model can be trained end to end, using the reparameterization trick, a mask matrix M1 is constructed to discard the noise edges in the social graph. We use the same method to construct another mask matrix M2 for contrastive learning. The calculation formula is as follows:

[0097] M1(i,i') = σ((logσ-log(1-σ)+M(i,i') / τ)

[0098] wherein, σ represents sigmoid activation function, and τ is temperature coefficient.

[0099] S4, user social multi-influence learning. The social relationship between users has multiple influences, and accurate mining of the main intent influence in the social relationship is crucial for maximizing the use of social relationships. First, the attention mechanism is used to calculate the weight coefficient of the social relationship under each intent. Then, the weight coefficient W fNormalization is performed so that the model can give higher attention weight to the primary intent. Then, the noise edges on W f are removed using M1 to avoid irrelevant social edges introducing disruptive information. The formula of the above steps is as follows:

[0100]

[0101]

[0102] W f = W f ⊙ M1

[0103] where ⊙ represents the dot product operation, E1 represents the user feature vector learned from the first social denoising view, and the user feature vector E2 is also learned from the second social denoising view.

[0104] Then, under each intent, message passing and message aggregation are performed on the graph neural network, and the formula is as follows:

[0105]

[0106] wherein and The left upper corner l represents the number of layers of GCN, and there are L layers in total. D is the degree matrix of , and δ represents the LeakyRelu activation function. After L layers of GCN, the final representation of the user feature vector can be obtained:

[0107]

[0108] wherein E1 represents the user feature vector learned from the first social denoising view, and the user feature vector E2 is also learned from the second social denoising view.

[0109] S5, unsupervised loss. The user feature vectors E1 and E2 of the two social views obtained from S4 are subjected to self-supervised contrastive learning. For user i, the negative sample is the user in the same batch who does not have an edge in the social graph, denoted as I - , then the self-supervised loss function is:

[0110]

[0111] wherein sim is the cosine similarity function, and τ is the temperature coefficient.

[0112] S6, supervised loss. The user feature vectors obtained from the two views are subjected to average operation to obtain the final user feature vector. Then, the BPRLoss loss function is used to obtain the supervised loss Loss bpr :

[0113] score ij = δ ((e i ) T h j )

[0114]

[0115] where δ is sigmoid function, j - represents negative sample, and λ is a regularization coefficient.

[0116] S7, loss fusion. The sum of the supervised loss and the unsupervised loss is taken as the target loss function of the recommendation model, the gradient of the loss to each parameter is calculated by using the gradient descent method, and is propagated back to the network to constantly update the model parameters. The network parameters updated in each round of training and the user and commodity feature vectors are input into the network, and the embedding vectors of the user and the commodity in the test stage are obtained through calculation.

[0117] S8, recommendation prediction. The prepared 1:100 positive and negative examples of the commodity are used to perform inner product operation on the feature vectors and the user feature vectors to perform recommendation prediction scoring. Wherein, the higher the inner product value, the greater the possibility of the user selecting the commodity, and vice versa. Then, the scores are sorted in descending order. Whether the top 10 commodities have positive examples can be used to represent the hit rate, and if a hit is found, the position of the commodity in the top 10 is further used to represent the prediction accuracy, and the earlier the position, the more accurate the algorithm.

[0118] In summary, the method of the embodiment has at least the following advantages and beneficial effects compared with the prior art:

[0119] (1) The present application is based on the existence of multiple intention influences in the user-commodity interaction process. Compared with other methods that use a single feature vector to represent a node, multiple feature vectors are used to represent the same node in different intention spaces. In different intention spaces, the same graph convolution network is used to model the user-commodity interaction graph, and the iterative update principle is used to calculate the weight of the interaction edge, and the same edge is normalized to decouple the main intention of the interaction, so that the model learns more fine-grained user representation and commodity representation, and obtains better recommendation performance.

[0120] (2) The present application also considers multiple influences on the social graph. Similarly, in different intention spaces, a graph neural network is used to model the social graph. Before the message passing of the graph neural network, an attention mechanism is used to calculate the weight of the same social edge in different intention spaces, and then a normalization operation is performed to make the same edge have different weights in different intention spaces, so as to mine the dominant influence of the social relationship and learn more fine-grained user representation.

[0121] (3) Compared with the self-supervised manner of constructing multiple views by using random Dropout, the application adopts adaptive Dropout operation, avoids destroying the overall graph structure by deleting useless edges irrelevant to the task, improves the robustness and generalization of the model, and alleviates the data sparsity problem.

[0122] The embodiment also provides a socialized recommendation device based on user multi-intention and self-supervised learning, comprising:

[0123] at least one processor;

[0124] at least one memory for storing at least one program;

[0125] When the at least one program is executed by the at least one processor, the at least one processor implements Figure 1 the method.

[0126] The socialized recommendation device based on user multi-intention and self-supervised learning of the embodiment can execute the socialized recommendation method based on user multi-intention and self-supervised learning provided by the method embodiment, can execute the implementation steps of any combination of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0127] The application also discloses a computer program product or a computer program, which comprises computer instructions stored in a computer readable storage medium. A processor of a computer device can read the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes Figure 1 the method.

[0128] The embodiment also provides a storage medium storing instructions or programs that can execute the socialized recommendation method based on user multi-intention and self-supervised learning provided by the method embodiment, and when the instructions or programs are executed, the implementation steps of any combination of the method embodiment can be executed, and the corresponding functions and beneficial effects of the method are possessed. In some alternative embodiments, the functions / operations mentioned in the block diagram can not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two blocks shown in succession can actually be executed substantially simultaneously or the blocks can sometimes be executed in reverse order. In addition, the embodiments presented and described in the flowcharts of the application are provided by way of example, and the purpose is to provide a more comprehensive understanding of the technology. The disclosed method is not limited to the operations and logical flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and in which sub-operations described as part of larger operations are independently executed.

[0129] Furthermore, although the present application is described in the context of functional modules, it is to be understood that one or more of the described functions and / or features can be integrated in a single physical device and / or software module, or one or more functions and / or features can be implemented in separate physical devices or software modules. It will also be appreciated that detailed discussion of the actual implementation of each module is not necessary to an understanding of the application. Rather, the actual implementation of the modules, in combination with their attributes, functions, and internal relationships, are to be understood within the context of the devices disclosed herein. Thus, those skilled in the art with access to patents, scientific journals, and other public sources known by those skilled in the art will be able to practice the application as set forth in the claims without undue experimentation, using ordinary skill in the art along with the present disclosure. It is also to be understood that the specific concepts disclosed are merely illustrative and that the scope of the present application is to be determined by the entire scope of the claims, along with all equivalents of the claims and their equivalents.

[0130] If the functions are implemented in software, the functions can be stored in or implemented as one or more computer program products, which can be incorporated into a computer-readable medium for use by or in connection with an apparatus, method or system as described herein. The computer-readable medium can be a machine-readable storage device, a machine-readable storage substrate, a memory device, or a computer-readable signal. The computer-readable medium can be, for example, but is not limited to, volatile or non-volatile memory, a floppy diskette, a compact disc read-only memory (CD-ROM), a magnetic tape, a flash drive, a hard disk drive, a programmable read-only memory (PROM), a random access memory (RAM), a reprogrammable read-only memory (REPROM) or a similar medium.

[0131] The logic and / or steps represented in the flowcharts and / or otherwise described herein, for example, can be embodied in non-transitory computer-readable media, which can be executed by an instruction execution system, apparatus, or device such as a computer-based system, a processor- based system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions. In this context, a "computer-readable medium" can be any means that can contain, store, communicate, propagate or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

[0132] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory.

[0133] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above described embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following technologies, or combinations thereof, can be used with the necessary logic gates and circuitry for implementing logic functions on data signals: discrete logic circuitry having logic gates for implementing logic functions on data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and so forth.

[0134] In the above description of the present specification, reference to the description of the terms "one embodiment", "another embodiment", or "certain embodiments" or the like 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 application. The illustrative appearances of the above described terms in the specification do not necessarily refer to the same embodiment or example. Moreover, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0135] Although embodiments of the present application have been shown and described, it would be recognized by those of ordinary skill in the art that various changes, modifications, substitutions, and alterations can be made hereto without departing from the spirit and scope of the application, which is defined by the following claims and their equivalents.

[0136] The above is a specific description of the preferred embodiments of the present application, but the present application is not limited to the above-described embodiments, and those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present application, and these equivalent modifications or substitutions are included in the scope defined by the claims of the present application.

Claims

1. A method for socialized recommendation based on user multi-intents, characterized in that, The method comprises the following steps: Collecting a dataset with user commodity interaction relationship and user social relationship; Storing the data in a sparse matrix to form graph structure data for graph convolution network processing; Building a recommendation model comprising two parts of graph convolution network to model the user commodity interaction relationship and the user social relationship in sequence; Intention space mapping: using a multi-layer perception to map an initial user feature vector and an initial commodity feature vector to different intention spaces; Interaction intention decoupling learning: building a graph neural network in each intention space to perform representation learning, to obtain a user feature vector and a commodity feature vector in each intention space; For a commodity node, summing the commodity feature vectors in each intention space to obtain a final commodity feature vector; Adaptive denoising: using an adaptive Dropout module to remove irrelevant edges in a social relationship graph to retain important edges; User social multi-influence learning: according to the obtained user feature vector in each intention space, using a graph neural network to model a user social graph in each intention space, and summing the obtained user feature vectors in each intention space to obtain a final user feature vector; Unsupervised learning: introducing two adaptive Dropout modules to obtain two social views for comparative learning; Training and learning the recommendation model in combination with the obtained final commodity feature vector, final user feature vector and unsupervised learning; Scoring and predicting the user and commodity feature vectors learned by the recommendation model to obtain a recommended commodity sequence; The adaptive denoising comprises: A multi-layer perceptron is used to compute a dropout coefficient for each social edge , which determines whether the social edge is dropped or not. The dropout coefficient is computed as follows: In the formula, || represents a splicing operation; In order to make the whole model can be trained end to end, the reparameterization trick is adopted, and a mask matrix of the noise edges in the discarded social graph is constructed ; The calculation formula is as follows: wherein, represents a sigmoid activation function, is a temperature coefficient, and denote two users on a social edge, respectively. 2.The method of claim 1, wherein, The intention space mapping comprises: Using a multi-layer perception to map a randomly initialized feature vector to different intention spaces; wherein, E represents a total user feature vector, H represents a total item feature vector, f represents the i-th intent. 3.The method of claim 1, wherein, The decoupled intention interaction learning comprises: In each intention space, assigning a value of 1 to an initial adjacency matrix corresponding to a user commodity interaction, representing that the influence of each intention on the user commodity interaction is the same in an initial state; The user feature vector and the commodity feature vector are input into a graph convolution network, and Sub-iterations are performed, i.e., message passing and message aggregation, to constantly update the incidence matrix of each intent, so that important intents can be decoupled. Using a multi-layer graph convolution network to obtain high-order neighbor information; and using an average operation to aggregate the learned feature vectors of each layer, as shown in the following formula: wherein, represents a user feature vector of the first layer network under the first intention, represents a product feature vector of the first layer network under the first intention; represents a user feature vector under the first intention, represents a user feature vector under the first intention; Finally, using a summation operation to sum the commodity feature vectors in each intention space to obtain a final commodity feature vector.

4. The method of claim 1, wherein, The user social multi-influence learning comprises: In each intention space, using an attention mechanism to calculate a weight coefficient of a social edge, using a softmax to normalize the weight values of the same social edge in different intention spaces, and then performing a dot product operation with a mask matrix to remove the weights of noise edges to obtain a final social adjacency matrix, and then performing message passing and message aggregation in a graph neural network; using a multi-layer graph neural network to capture high-order neighbor information; as shown in the following formula: wherein, representing a user feature vector learned from a first social denoising view, and additionally a user feature vector learned from a second social denoising view ; Using an average operation to aggregate the feature vectors of the two views to obtain a final user feature vector.

5. The method of claim 1, wherein, The training and learning of the recommendation model in combination with the obtained final commodity feature vector, final user feature vector and unsupervised learning comprises: Taking a sum of a supervised loss and an unsupervised loss as a target loss function of the recommendation model, wherein the supervised loss is a loss function in node feature vector learning, and the unsupervised loss is a loss function in unsupervised learning; The parameters of the recommendation model are updated using a gradient descent method until a preset threshold of a target loss function is reached.

6. The method of claim 1, wherein, The socialized recommendation method further comprises a step of preprocessing the data set: According to the preset condition, filter invalid users and goods, and keep valid user and goods nodes; The data set is divided, two interaction items of the user are randomly selected, and the remaining interaction items are put into the training set.

7. A user multi-intention based socialized recommendation apparatus, characterized by, Comprise: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method of any one of claims 1-6.

8. A computer readable storage medium having stored therein a program which is executable by a processor, characterized in that, The program executable by the processor is used to execute the method of any one of claims 1-6 when executed by the processor.

Images