A Method, Device and Storage Medium for Debiasing User Multi-Interest Recommendation

Abstract

The present invention discloses a method, device, and storage medium for debiased user multi-interest recommendation. The method comprises: collecting a dataset of user-item interaction records, obtaining a training set based on the dataset; storing the data in sparse matrix form to construct graph-structured data that can be processed by a graph convolutional neural network; learning node feature vectors on a user-item relationship graph; learning node embedding vectors; optimizing global relationships to obtain more potential semantic relationships; hyperedge representation and interest prototype optimization; obtaining multi-interest representations; debiased auxiliary learning; combining a supervised loss function with a contrastive loss function, using an optimizer for backpropagation to optimize network parameters; iteratively training the training set data until the model converges; and performing recommendation prediction. Through contrastive learning tasks and adaptive sampling, the present invention eliminates bias caused by public psychological interference when capturing user interests, improves recommendation accuracy, and is widely applicable in the field of machine learning technology.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method, device, and storage medium for debiased user multi-interest recommendation. Background Art

[0002] In the context of big data, efficiently extracting and mining useful information from massive amounts of data has become a highly sought-after research topic, attracting increasing attention from experts and scholars both domestically and internationally. Processing and mining this data is highly beneficial for both users and service providers. With the increasing development and adoption of artificial intelligence (AI) technology, more and more researchers are devoting significant time and effort to the study of personalized recommendation systems. Recommendation tasks have become both common and important, and the development of better and more accurate recommendations is a pressing goal.

[0003] In recent years, the application of graph neural networks in recommender systems has significantly improved recommendation performance, alleviating the data sparsity problem to some extent. However, acquiring global information through iterative multi-layer graph neural networks can easily lead to oversmoothing, causing node representations to become uniform, ultimately affecting the acquisition of global information. Furthermore, many existing recommender system algorithms simply and generally optimize user and item representations, failing to reflect users' diverse interests. However, people are often influenced by multiple interests when performing certain behaviors. For example, a person may like a movie because of their interest in the director, the lead actor, or the subject matter. Different users in different real-world scenarios have different interests. By studying users' multiple interests, we can more precisely pinpoint the specific reasons or primary reasons that motivate a user to take a particular action, thereby improving the accuracy and efficiency of recommendations. Numerous experts and scholars are currently engaged in mining multiple interests, but existing methods mostly use clustering algorithms that can only cluster by specifying a specific number of interests and cannot adaptively change the number of interests of different users. Summary of the Invention

[0004] In order to at least partially solve one of the technical problems existing in the prior art, the present invention aims to provide a method, device and storage medium for debiased user multi-interest recommendation.

[0005] The technical solution adopted in the present invention is:

[0006] A method for debiasing user multi-interest recommendation, comprising the following steps:

[0007] Collect a dataset of user-product interaction records and obtain a training set based on the dataset;

[0008] According to the interactive relationship of the data set, the data is stored in the form of a sparse matrix, which is constructed into graph structure data that can be processed by the graph convolutional neural network;

[0009] Use a graph convolution submodule to learn node feature vectors on the user-item relationship graph;

[0010] Node embedding vector learning: Use a multi-layer convolutional network layer to obtain high-level semantic information, sum the feature outputs of each layer to obtain the product feature vector, and fuse the user representation with the user's multiple interest representations to obtain the user's final embedding vector;

[0011] Use a hypergraph convolutional neural network to optimize global relationships and obtain more potential semantic relationships;

[0012] Hyperedge representation and interest prototype optimization: Using multi-layer hypergraph convolutional network iteration, the optimized hyperedge global representation is applied to the update of interest prototypes, ultimately making the existing interest representation based on the existing interaction information and its inherent latent semantics;

[0013] Acquiring multi-interest representations: Calculate the similarity between the user representation and the prototype representations of the interest in the large interest pool, sort the similarities in descending order, and then select the K interests in the interest pool that are most similar to the user as the user's multi-interest representation through personalized selection of the similarity threshold. That is, select the K most similar interests until the similarity gap with the K+1th interest is maximized.

[0014] Debiasing-assisted learning: Debiased data is obtained through adaptive sampling. To ensure adaptive debiasing, the data with the greatest bias is selected to form positive and negative pairs. A contrastive learning task is constructed, and the contrastive loss is calculated to assist learning.

[0015] Combine the supervised loss function with the contrastive loss function, use the Adam optimizer for backpropagation, and optimize the network parameters; iteratively train the training set data until the model converges;

[0016] Recommendation prediction: The user and item embedding vectors finally learned by the model are scored and predicted to obtain the recommended item order.

[0017] Furthermore, obtaining a training set according to the data set includes:

[0018] Invalid users are filtered out based on the condition that the number of user-item interactions is less than 2, and valid users and corresponding item nodes are retained. The dataset is divided, and one interaction is randomly selected from the validation set and the test set for each user. The remaining interaction items are used as the training set. To evaluate the prediction results, negative sampling is required on the validation set and the test set.

[0019] The processing of debiased data is to count the popularity of each product and then set a threshold. The threshold is different for different users and is calculated based on the ratio of the number of products interacted with by each user. That is, if it is greater than the margin value, it is a recognized popular product, and if it is less than the pool value, it is defined as a niche product.

[0020] Furthermore, the learning of node feature vectors on the user-item relationship graph includes:

[0021] Input users and products into the graph convolutional neural network encoder and output the node features in their respective relationship graphs;

[0022] From the updated user and product representations, we obtain the interaction representation and hyperedges, which are then fed into the hypergraph convolutional neural network. While updating the interaction representation, we also update the hyperedges. Then, we feed the interest representation into another hypergraph neural network, ultimately optimizing and updating the interest representation.

[0023] The final user and product feature vectors are obtained by adding and averaging the outputs of all layers and the initialized feature vector.

[0024] Furthermore, the acquisition of multiple interest representations includes:

[0025] The similarity between user representation and interests is obtained through cosine similarity, and the similarity is sorted in descending order according to the size of the similarity. The K most similar interests are automatically selected until the similarity interval with the K+1th interest is the largest, thus obtaining the representation of the user's multiple interests.

[0026] Furthermore, the feature vector finally obtained by the model is used to obtain the supervised loss of the model recommendation prediction using the Margin-Loss loss function.

[0027] Furthermore, the debiasing assisted learning includes:

[0028] By adaptively sampling the biased and debiased data, we construct a contrastive learning task to eliminate the bias caused by public psychological interference when capturing user interests. We introduce supervisory signals to further optimize the representation vectors of users and products, and obtain the unsupervised loss of the model contrastive learning.

[0029] Furthermore, the model's target loss function is the sum of supervised loss and unsupervised loss. Using the gradient descent method, the model parameters are continuously updated until the target loss is minimized, and high-quality expressions of users and products are learned to achieve accurate recommendation predictions.

[0030] Another technical solution adopted in the present invention is:

[0031] A debiased user multi-interest recommendation device, comprising:

[0032] at least one processor;

[0033] at least one memory for storing at least one program;

[0034] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0035] Another technical solution adopted in the present invention is:

[0036] A computer-readable storage medium stores a program executable by a processor, wherein the program executable by the processor is used to perform the method described above when executed by the processor.

[0037] The beneficial effects of the present invention are as follows: the present invention eliminates the deviation caused by public psychological interference when capturing user interests through comparative learning tasks and adaptive sampling methods, further optimizes the representation vectors of users and products, and uses the user's final representation embedding vector and the product's final representation embedding vector to make preference prediction more accurate, thereby enhancing the interpretability and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following introduction is made to the drawings of the embodiments of the present invention or the related technical solutions in the prior art. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative work.

[0039] Figure 1 This is a workflow diagram of a multi-interest mining recommendation method based on a hypergraph neural network in an embodiment of the present invention;

[0040] Figure 2 It is a framework diagram of the recommendation model in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention. The step numbers in the following embodiments are provided for ease of explanation only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0042] In the description of the present invention, it should be understood that descriptions involving orientations, such as up, down, front, back, left, right, etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limitations on the present invention.

[0043] In the description of the present invention, "several" means one or more, "many" means more than two, "greater than," "less than," and "exceed" are understood to exclude the number itself, while "above," "below," and "within" are understood to include the number itself. The use of "first" and "second" in the description is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, implicitly specifying the number of the indicated technical features, or implicitly specifying the order of the indicated technical features.

[0044] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0045] like Figure 1 and Figure 2 As shown, this embodiment provides a multi-interest mining recommendation method based on a hypergraph neural network. First, the acquired data set is divided into a training set, a validation set, and a test set, preprocessed, and a relationship diagram of product interactions is constructed, and then fed into a graph convolutional network for encoding. Subsequently, the debiased interaction data and the biased interaction data are fed into the hypergraph convolutional neural network respectively and a comparative learning task is constructed. The supervision signal is introduced and the unsupervised loss is calculated. Finally, the supervised loss (Margin-Loss) and the unsupervised loss are jointly trained to perform the final recommendation prediction. The method specifically includes the following steps:

[0046] S1. Build a data set to obtain a data set containing user personal information, interaction records, and product category information on the Internet platform, and further filter, divide, and encode it.

[0047] Datasets are collected. E-commerce platforms, video sites, and review websites all have user-item interaction records. Considering that user interests are dynamic, we selected datasets with time information when collecting data. Currently available Gowalla, Taobao, and Yelp review datasets can all serve as candidate datasets for filtering, preprocessing, and dataset partitioning. The user-item unique identifiers are converted into numerical values, and the mapping process eliminates incomplete data.

[0048] After data collection, the step of preprocessing the data is also included: select user sets and item sets with a number of user-item interactions greater than 2, construct them into an adjacency matrix form, and store them in a file in CSR format. Adaptive sampling to obtain debiased data and biased data runs through each step of training. Specifically, after data collection, users with an interaction number of less than 2 are eliminated to ensure high data quality. Subsequently, the interaction items of each user are divided, and one interaction record is randomly selected as the validation and test sets, and the rest are used as training sets. Finally, in order to verify and test the prediction effect of the model recommendation, the validation set and test set are evaluated in the form of positive and negative samples m:n, that is, the all-ranking method, with all existing interactions of each user as positive samples and all samples with missing interactions as negative samples.

[0049] S2. Node feature vector learning: The relationship data stored in a graph structure is input into the user-product interaction graph convolutional network for feature learning, and the aggregated and updated node features are output. The graph convolutional neural network mainly includes two steps: message passing and message aggregation. Taking product features as an example, the user feature update is similar. The specific process is shown in the following formula:

[0050]

[0051]

[0052] Where m represents the node feature, N represents the total number of products or users, and u i represents the i-th user, v j represents the jth product, H is the updated node feature, and l refers to the lth layer, involving a multi-layer convolutional neural network.

[0053] Initialize user and item representations and feed them into the GCN encoder along with interaction information to optimize their representations. Since the public bias between users and popular items arises during interaction, it's necessary to initialize the interaction representation with the help of user and item initialization. Initialize interest prototypes. Optimizing the representation of the interest prototypes themselves can also better optimize the user-selected interest representations, ultimately enhancing the user's final representation. To better capture global relationships, the potential relationships between interests in the original interest pool are more explicitly represented, and hyperedge nodes are initialized. While optimizing hyperedge nodes, the interest prototype representations are also optimized as a supplement.

[0054] S3. Learning the feature vector of interest prototype. Initialize the biased interaction features and debiased interaction features respectively, and send them into the interactive hypergraph convolutional neural network to update the hyperedge representing the global relationship. Then send the updated hyperedge into the interest hypergraph convolutional neural network, and finally obtain the updated interest prototype vector. Formula 3 represents the initialization of biased interaction features. The same is true for debiased interaction features. Formula 4 represents the detailed process of updating hyperedges and interest prototypes in the hypergraph convolution calculation. The different number of convolution layers is determined by superimposing multiple layers of nonlinear mapping. represents the biased interactive feature representation of the lth layer, W (u) represents the representation of the hyperedge node, δ represents the nonlinear transformation (try using different activation functions such as ReLU, LeakyReLU, etc.), Represents the prototype representation of interest in a biased situation.

[0055]

[0056]

[0057] The updated user and product representations are converted into interaction representations and input into the hypergraph convolutional neural network together with the hyperedges. When the interaction representation is updated, the hyperedges are also updated. Then, the interest representation is input into another hypergraph neural network, and finally the interest representation is optimized and updated:

[0058] 1) The interactive hypergraph convolutional neural network iterates two layers, updating the biased interaction data and the debiased interaction data respectively to prepare for the subsequent contrastive learning task. At the same time, the hyperedge nodes can also be updated to optimize the interest prototype.

[0059] 2) The interest prototype hypergraph neural network uses three layers of hypergraph convolution to iteratively update the interest prototype, which can, to a certain extent, eliminate the bias caused by public psychology when capturing user interests.

[0060] S4. Optimize the user's multi-interest representation. Calculate the cosine similarity between the original user's representation and the prototype representation of the interest in the large interest pool, as shown in Formula 5. Through personalized selection of the similarity threshold, select the K interests in the interest pool that are closest to the user as the user's multi-interest representation. That is, select the most similar K interests until the similarity interval with the K+1th interest is maximized.

[0061]

[0062] The representation of the first K interests is added to the inner product of the initial representation of the user's multiple interests that combines the user and the original interest prototype representation to obtain the final representation of the user's multiple interests. The detailed process can be referred to Formula 6, e u and e int They represent the user initialization representation and interest prototype initialization standard respectively, and Z represents the user's multi-interest representation that integrates K interest prototypes.

[0063]

[0064] The above process goes through multiple layers of iteration, and the representation of each feature vector at the lth layer is stored for subsequent fusion, and the residual result is considered in the fusion process, where Λ (u,l) represents the final representation of the lth layer that integrates multiple user interests, Λ (v,l) is the product representation. The user and product representations used for rating calculation are the sum of the results of each layer, ψ (u) represents the user's final embedding vector, ψ (v) Represented as the final embedding vector of the product.

[0065] Λ (u,l) =Z (u,l) +H (u,l) +Λ (u,l-1) (7)

[0066] Λ (v,l) =H (v,l) +Λ (v,l-1) (8)

[0067]

[0068] S5. Since the user-product feature vectors will be affected by the herd mentality of users caused by popular products during the learning optimization process, in order to eliminate the deviation caused by the herd mentality to the experimental results, the present invention uses adaptive sampling and comparative learning tasks to jointly reduce the effect of the deviation. In the adaptive sampling process, the most popular data, i.e. biased data, and the least popular data, i.e. debiased data, are sampled respectively, and these two batches of data are sent to the interactive hypergraph neural network respectively, so that the hyperedge representation representing the global relationship is optimized. In order to ensure adaptive debiasing, the data with the largest bias and the data with the smallest bias are selected to form a positive and negative pair, and a comparative learning task is constructed. The comparative loss is calculated to assist learning, and InfoNCE is used to maximize the consistency of the positive examples, and a greater weight is placed on the learning of the debiased features, minimizing the consistency of the negative columns. Its mathematical expression is shown in Formula 10:

[0069]

[0070] Among them, s is the similarity function (cosine similarity or other similarity functions are acceptable), which calculates the similarity between vectors, and τ is the temperature coefficient, which is used to automatically control the difficulty of negative samples. con The hypergraph representation obtained by optimizing the biased interactive hypergraph convolutional neural network is represented as a negative sample. Similarly, E int Represented as positive samples, by using contrastive learning tasks, the positive and negative samples can be separated to a certain extent, and ultimately more weight can be placed on the positive samples for optimization.

[0071] Among them, the positive and negative sample pairs: the positive sample pair is a debiased interaction representation, and the negative sample pair is a biased interaction representation. The distance between them is maximized. The above positive and negative sample pair features are normalized and substituted into the InfoNCE expression to calculate the contrast loss of each sample. The total contrast loss is obtained by adding up all the sample contrast losses and dividing by the number of samples.

[0072] S6. Recommendation prediction. The model finally outputs comprehensive user and product embedding vectors, and uses the Margin-Loss loss function to obtain the supervised loss L of the recommendation task. margin , added to the above unsupervised loss to obtain the final model loss. β is an adjustable hyperparameter used to control the weight of the unsupervised loss generated by contrastive learning:

[0073] L=L margin +β*L cl (11)

[0074] The goal of this neural network is to minimize loss. Using gradient descent, the gradient of the loss with respect to each parameter is calculated and back-propagated into the network to continuously update the model parameters. The updated network parameters for each training round, as well as the parameters of the user, product, and interest prototype feature vectors, are input into the network. After calculation, embedding vectors for users (integrating multiple user interest representations) and products are calculated for the test phase. Using the prepared m:n positive and negative columns, a predicted score is determined by the inner product of the two embedding vectors. A higher value indicates a higher user preference for the product, and vice versa.

[0075] The calculation of the predicted score demonstrates a significant advantage of the present invention over other multi-interest mining methods. Other methods simply calculate and rank the user's fixed K multi-interest representations separately, separating the multiple representations of a single user and calculating the inner product with the product representations. While this can reflect the distinctness between interests, it loses the overall integrity of each user and also the user's original characteristic representation. The present invention, on the other hand, sums the user's original representation with the multi-interest representations and adaptively integrates the K multi-interest representations before calculating the score. This reflects both the distinctness between user interests and the overall integrity of the user. The score sequence table is sorted in descending order, and the product with the highest predicted score is the one with the strongest user preference. The hit rate, one of the evaluation metrics of the present invention, can be represented by the number of positive examples among the top 20 scoring products. Furthermore, considering that the ranking information of the recommendation list involves another evaluation metric—normalized discounted cumulative gain—the model training is completed until the recommendation prediction effect is optimal, making accurate recommendations and obtaining a list of products that the user may like.

[0076] In summary, compared with the prior art, this embodiment has at least the following advantages and beneficial effects:

[0077] (1) The present invention starts from the fact that different users have multiple interests. A unified user representation cannot reflect the user's multiple interests, and the number of multiple interests of each user is also different. Most existing multi-interest mining methods use clustering algorithms, which can only perform clustering by specifying the number of interests, and cannot achieve adaptive changes in the number of multiple interests of different users. Compared with the existing technology, the present invention selects the K most similar interests of each user from a large-scale interest prototype pool until the interval with K+1 similarities is the largest, adaptively learns the number of multiple interests of each user, and finally fuses the multiple interest representations of each user to generate the user's multi-interest representation.

[0078] (2) The hypergraph technology used in the present invention can more accurately describe multivariate relationships. It has strong capabilities in depicting and mining nonlinear high-order relationships and can fully utilize ordinary binary relationship data to mine more information. It can learn the relationship between any two nodes. In addition, the construction of traditional hypergraphs is based on the prior relationship between nodes, while the hypergraph used in the present invention is adaptive, that is, the hyperedge nodes are learnable, and their global relationships can be learned efficiently. Introducing noise increases the robustness of the model, helping the model learn more powerful feature representations, can effectively handle the noise of real data, enhance its generalization performance, and prevent overfitting.

[0079] (3) The present invention mentions the use of a hypergraph neural network to optimize interest prototypes. It first updates the biased interaction representation and the debiased interaction representation through a two-layer hypergraph convolutional neural network. While updating the interaction representation, it also updates the hyperedge, extracting the global potential relationship of the existing interaction information. It then inputs it into another hypergraph neural network with the interest prototype to update the interest representation. The acquisition of the interaction representation requires the use of the user-item GCN encoder to optimize the original representation of users and items based on the existing interaction information. It iterates multiple layers, which can alleviate the problem of sparse interaction information data to a certain extent.

[0080] (4) A clever design of adaptive debiasing sampling divides the data into three parts: biased, debiased, and blurred. Using contrastive learning tasks, debiasing is enhanced and bias is weakened, automatically and implicitly placing greater weight on learning debiased representations. Furthermore, performing InfoNCE calculations on the data is equivalent to introducing supervisory signals, enabling deeper mutual supervision and better learning of feature representations. Contrastive learning focuses on learning the differences between instances of different classes, distancing biased data and bringing debiased data closer.

[0081] This embodiment further provides a device for debiasing user multi-interest recommendation, comprising:

[0082] at least one processor;

[0083] at least one memory for storing at least one program;

[0084] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0085] A debiased user multi-interest recommendation device of this embodiment can execute a debiased user multi-interest recommendation method provided by the method embodiment of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0086] The present application also discloses a computer program product or computer program, which includes 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 performs Figure 1 The method shown.

[0087] This embodiment also provides a storage medium storing instructions or programs that can execute a de-biased user multi-interest recommendation method provided by an embodiment of the method of the present invention. When the instructions or program are run, any combination of implementation steps of the method embodiment can be executed, and the corresponding functions and beneficial effects of the method can be obtained.

[0088] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operation and logic flow presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0089] Furthermore, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise indicated, one or more of the functions and / or features described may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It will also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Therefore, a person skilled in the art using ordinary skill will be able to implement the present invention set forth in the claims without undue experimentation. It will also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0090] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0091] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0092] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0093] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0094] In the above description of this specification, reference to the terms "one embodiment / example," "another embodiment / example," or "certain embodiments / examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0095] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0096] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for debiasing user multi-interest recommendation, characterized in that: The following steps are involved: Collect a dataset of user-product interaction records and obtain a training set based on the dataset; According to the interactive relationship of the data set, the data is stored in the form of a sparse matrix, which is constructed into graph structure data that can be processed by the graph convolutional neural network; Use a graph convolution submodule to learn node feature vectors on the user-item relationship graph; Node embedding vector learning: Use a multi-layer convolutional network layer to obtain high-level semantic information, sum the feature outputs of each layer to obtain the product feature vector, and fuse the user representation with the user's multiple interest representations to obtain the user's final embedding vector; Use a hypergraph convolutional neural network to optimize global relationships and obtain more potential semantic relationships; Hyperedge representation and interest prototype optimization: Using multi-layer hypergraph convolutional network iteration, the optimized hyperedge global representation is applied to the update of interest prototypes, ultimately making the existing interest representation based on the existing interaction information and its inherent latent semantics; Acquiring multi-interest representations: Calculate the similarity between the user representation and the interest prototype representations in the large interest pool, sort the similarities in descending order, and then select the K interests in the interest pool that are most similar to the user as the user's multi-interest representation through personalized selection of the similarity threshold; Debiasing-assisted learning: Debiased data is obtained through adaptive sampling. To ensure adaptive debiasing, the data with the greatest bias is selected to form positive and negative pairs. A contrastive learning task is constructed, and the contrastive loss is calculated to assist learning. Combine the supervised loss function with the contrastive loss function, use the Adam optimizer for backpropagation, and optimize the network parameters; iteratively train the training set data until the model converges; Recommendation prediction: The user and item embedding vectors finally learned by the model are scored and predicted to obtain the recommended item order.

2. The method for debiasing user multi-interest recommendation according to claim 1, characterized in that: The step of obtaining a training set according to the data set includes: Invalid users are filtered out based on the condition that the number of user-item interactions is less than 2, and valid users and corresponding item nodes are retained. The dataset is divided, and one interaction is randomly selected from the validation set and the test set for each user. The remaining interaction items are used as the training set. To evaluate the prediction results, negative sampling is required on the validation set and the test set. The processing of debiased data lies in counting the popularity of each product and then setting a threshold, which is calculated based on the ratio of the number of products interacted with by each user. That is, if it is greater than the margin value, it is a recognized popular product, and if it is less than the pool value, it is defined as a niche product.

3. The method for debiasing user multi-interest recommendation according to claim 1, characterized in that: The learning of node feature vectors on the user-item relationship graph includes: Input users and products into the graph convolutional neural network encoder and output the node features in their respective relationship graphs; From the updated user and product representations, we obtain the interaction representation and hyperedges, which are then fed into the hypergraph convolutional neural network. While updating the interaction representation, we also update the hyperedges. Then, we feed the interest representation into another hypergraph neural network, ultimately optimizing and updating the interest representation. The final user and product feature vectors are obtained by adding and averaging the outputs of all layers and the initialized feature vector.

4. The method for debiasing user multi-interest recommendation according to claim 1, characterized in that: The multi-interest representation acquisition includes: The similarity between user representation and interests is obtained through cosine similarity, and the similarity is sorted in descending order according to the size of the similarity. The K most similar interests are automatically selected until the similarity interval with the K+1th interest is the largest, thus obtaining the representation of the user's multiple interests.

5. The method for debiasing user multi-interest recommendation according to claim 1, characterized in that: The feature vector finally obtained by the model is used to obtain the supervised loss of the model recommendation prediction using the Margin-Loss loss function.

6. The method for debiasing user multi-interest recommendation according to claim 1, characterized in that: The debiasing-assisted learning includes: By adaptively sampling the biased and debiased data, we construct a contrastive learning task to eliminate the bias caused by public psychological interference when capturing user interests. We introduce supervisory signals to further optimize the representation vectors of users and products, and obtain the unsupervised loss of the model contrastive learning.

7. The method for debiasing user multi-interest recommendation according to claim 1, characterized in that: The model's target loss function is the sum of supervised loss and unsupervised loss. Using the gradient descent method, the model parameters are continuously updated until the target loss is minimized. This allows us to learn high-quality expressions of users and products and achieve accurate recommendation predictions.

8. A debiased user multi-interest recommendation device, characterized in that: include: 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 according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a program executable by a processor, characterized in that: The processor-executable program is used to perform the method according to any one of claims 1 to 7 when executed by the processor.

Images