A service recommendation method based on representation decoupling and user modality preference capture

Abstract

This invention discloses a service recommendation method based on representation decoupling and user modality preference capture, belonging to the fields of service recommendation and representation learning. The method includes: first, constructing a historical interaction graph between users and services; then, obtaining corresponding semantic codes for the service's functional text description and category through two RQVAEs; second, calculating the number of semantic matches between services based on different semantic codes to obtain corresponding association strength matrices; further, combining the historical interaction graph to generate two modality enhancement graphs; based on the two enhancement graphs, obtaining user preferences in the corresponding modality and service features in the corresponding modality; further obtaining interaction probabilities and calculating the BRP loss function; and finally, obtaining a service retrieval table. This invention provides specific preference criteria for refined supply-side recommendation and advertising optimization, comprehensively improving the personalization and accuracy of recommendation models.

Description

Technical Field

[0001] This invention belongs to the field of service recommendation and representation learning, and specifically relates to a service recommendation method based on representation decoupling and user modality preference capture. Technical Background In recent years, service recommendation systems have gradually become a core support tool for various web applications. Their core value lies in breaking down the matching barriers between massive amounts of online information and users' personalized needs, helping users efficiently and accurately discover relevant content that meets their needs from the exponentially growing available information resources, significantly reducing information retrieval costs, and improving user efficiency and experience.

[0002] At the same time, with the rapid development of multimedia technology, multimodal information associated with various services has shown an explosive growth trend. This information, which covers multiple forms such as images, text, and videos, can more comprehensively and three-dimensionally outline the core characteristics and value connotations of items by virtue of its complementarity and richness. This helps users to understand their own interests and preferences more deeply, and significantly enriches and optimizes the user's interaction experience with the platform.

[0003] It is worth noting that this multi-dimensional and richly layered multimodal information not only profoundly influences the formation and evolution of user preferences, but also drives technological innovation in the recommendation field, giving rise to the emerging multimodal recommendation field (MMRec). The core focus of this field is to break through the limitations of traditional single-modal recommendation by effectively integrating and mining the correlational value inherent in multiple modalities of information to construct recommendation models that better meet users' real needs. This results in more personalized, accurate, and effective recommendation services, further improving the core performance and user satisfaction of recommendation systems.

[0004] Existing methods primarily employ modality fusion from multiple perspectives to obtain users' overall preferences for items. For example, attention concatenation, exemplified by DRAGON, weights and concatenates pre-trained vectors from various modalities before inputting them into a collaborative filtering network, achieving cross-modal information compression and preliminary semantic alignment. LATTICE first constructs an item graph for each modality and runs graph convolution, then pools or concatenates the output multimodal node representations to jointly encode structural and content information. MGCN, after obtaining information from multiple modalities, allows each modality to independently produce recommendation scores, which are then aggregated through gating or weighted voting to ensure the independent contribution of each signal to the final ranking. These fusion strategies all achieve good performance in graph neural network-based recommendation systems.

[0005] However, the existing methods still have the following drawbacks: (1) Current recommendation methods tend to integrate the features of services to obtain the overall preferences of users, making it difficult to distinguish user preferences on the supply side of goods.

[0006] (2) The current user and service interactions are too sparse, making it difficult to build relevant connections between various independent services, resulting in poor representation learning.

[0007] (3) Current multimodal recommendation methods directly use the multimodal features of services obtained from pre-trained networks without filtering or refining the noise information contained in the features. However, in actual recommendation scenarios, the morphological information of services often inevitably contains noise components.

[0008] Therefore, in order to more accurately capture users' interests and preferences and decouple users' interests in different modalities to achieve accurate recommendations, it is urgent to study a method based on decoupling user representations and capturing modal preferences. Summary of the Invention

[0009] To address existing technical challenges such as modal denoising, sparse user service interactions, and difficulty in mining user preferences for modal representations, this invention proposes a recommendation method based on representation decoupling and modal preference capture. The specific technical solution is as follows: S1. Collect multiple historical users and multiple services from the target service recommendation scenario, and construct a historical interaction graph based on the historical interaction relationships between all historical users and all services; input the functional text description and category of each service into the first RQVAE and the second RQVAE respectively to obtain the first semantic code and the second semantic code; S2, calculate the number of semantic matches of functional text descriptions among all services based on the first semantic encoding, and then obtain the first association strength matrix; calculate the number of semantic matches of categories among all services based on the second semantic encoding, and then obtain the second association strength matrix; S3, the historical interaction graph is enhanced using the first association strength matrix and the second association strength matrix respectively to obtain the first modality enhanced graph and the second modality enhanced graph; S4, input the first modality augmented graph, the second modality augmented graph, the service collaborative filtering representation, and the user collaborative filtering representation into the graph convolutional network to further obtain the interaction probability between each user and each service; S5. Calculate the BRP loss based on the interaction probability. Update the service collaborative filtering representation, user collaborative filtering representation and graph convolutional network parameters based on the BRP loss. Repeat S4-S5 until the BRP loss converges to obtain the trained service collaborative filtering representation and form a service retrieval table. S6. Calculate the similarity score between the user to be recommended and each service in the service retrieval table, and select the top K services with the highest similarity scores as the services to be recommended to the user.

[0010] Furthermore, in S1, a historical interaction graph is constructed based on the historical interaction relationships between all historical users and all services, specifically as follows: First, a historical interaction matrix is ​​constructed based on the historical interaction relationships between all historical users and all services. In the historical interaction matrix, if there is a historical interaction between a historical user and a service, the element at the corresponding position in the matrix is ​​1, and otherwise it is 0. Secondly, based on the historical interaction matrix, a historical interaction graph is further constructed.

[0011] Furthermore, in S1, both the first RQVAE and the second RQVAE are pre-trained RQVAEs. The pre-training process of the first RQVAE is as follows: S101, construct the first RQVAE containing an encoder, a decoder, and several quantizers containing codebooks; at the same time, randomly initialize the codebook of each quantizer; S102, the functional text descriptions of all services are encoded into discrete encoded sequences by an encoder; S103, the discrete encoded sequence is sequentially input into several quantizers, and each quantizer performs quantization processing through residuals; the outputs of each quantizer are then combined and restored to modal features by a decoder; S104, the reconstruction loss is calculated based on the modal features; the submission loss is calculated based on the discrete encoded sequence; the reconstruction loss and the submission loss are added together according to the weights to obtain the total loss; S105, update all parameters of the first RQVAE according to the total loss until the total loss converges, and obtain the codebook and the first RQVAE that has been trained. The pre-training method for the second RQVAE is the same as that for the first RQVAE, wherein the input of the second RQVAE encoder is the category of all services.

[0012] Furthermore, in S103, each quantizer performs quantization processing using a residual method, specifically as follows: The first quantizer matches the input discrete encoded sequence with the nearest codeword in the current quantizer codebook and outputs the matching result. The second and all subsequent quantizers obtain the corresponding matching results by matching the output of the previous quantizer with the nearest codeword in the current quantizer codebook.

[0013] Furthermore, in S104, the formula for calculating the reconstruction loss is as follows: ; in, Indicates the reconstruction loss. Represents the mathematical expectation. Represents the conditional probability distribution. Indicates the encoder input, This represents the latent vector output by the encoder. This represents the modal characteristics output by the decoder. Represents the L2 norm; The formula for calculating the submission loss is as follows: ; in, Indicates the submission of losses. This indicates that the gradient is not calculated. Indicates hyperparameters, For the first l The index of the codeword matched by each quantizer, where L represents the number of quantizers. Indicates the first l Service features after matching each quantizer Indicates the first l The vector closest to the quantizer, i.e., the quantizer l The codeword matched by each quantizer, where j represents the j-th service.

[0014] Further, in S2, the number of semantic matches in the functional text descriptions among all services is calculated, specifically as follows: ; in, Indicates an indicator function, This function calculates the matching degree between two services, where t represents the t-th service. This represents the first semantic encoding of the t-th service.

[0015] Furthermore, S3 specifically refers to: S301, transpose the historical interaction graph to obtain the transpose matrix of the historical interaction graph; S302, combine the historical interaction graph, the transpose of the historical interaction graph, and the two all-zero matrices in a block matrix form to obtain the initial enhanced graph; S303, the initial enhancement map is further enhanced using the first correlation strength matrix to obtain the first modal enhancement map; the initial enhancement map is further enhanced using the second correlation strength matrix to obtain the second modal enhancement map; specifically: ; ; in, This is a hyperparameter, where m is the number of users. and These represent the first correlation strength matrix and the second correlation strength matrix, respectively. Represents a historical interaction diagram. and These represent the first and second mode augmentation maps, respectively, in the upper right corner. This indicates transpose.

[0016] Further, in step S4, the user's preferences in the text description modality and the service features in the text description modality are obtained based on the first modality enhancement graph, specifically as follows: S401, initialize the corresponding initial user collaborative filtering features and initial service collaborative filtering features for each user and service respectively; S402, the initial user collaborative filtering features and the initial service collaborative filtering features are passed through each layer of LightGCN, and combined with the first modality augmentation map to obtain the user embedding vector and service embedding vector output by each layer of LightGCN. S403, combine the user embedding vectors of all layers of LightGCN to obtain the user's preferences in the text description modality; combine the service embedding vectors of all layers of LightGCN to obtain the service features in the text description modality.

[0017] Furthermore, in step S4, the formula for calculating the interaction probability between the user and the service is specifically as follows: ; in, This represents the sigmoid activation function. This represents a multilayer perceptron. This represents the vector concatenation operation. This represents the interest vector obtained based on user i's preferences in the functional text description modality and service j's features in the functional text description modality. This represents the interest vector obtained based on user i's preferences in the functional text description modality and service j's features in the category modality. This represents the interest vector obtained based on user i's preferences in the category modality and service j's features in the functional text description modality. This represents the interest vector obtained based on user i's preferences in the category modality and service j's features in the category modality.

[0018] Furthermore, the interest vector obtained based on user i's preferences in the functional text description modality and service j's features in the functional text description modality... The calculation formula is as follows: ; in, This indicates multiplication by element position. This represents the characteristics of service j in the functional text description modality. This indicates user i's preferences in the functional text description modality.

[0019] The beneficial effects of this invention are: Compared with existing multimodal recommendation technologies, this invention specifically addresses the core problems of noisy modal features, sparse user-service interactions, and difficulty in mining user modal preferences.

[0020] (1) By constructing RQVAE models for different modalities, continuous modal features are transformed into discrete semantic IDs. Combined with double loss function pre-training, modal feature denoising is achieved, filtering redundant noise while retaining core semantics, thus improving the accuracy of feature representation. (2) Based on semantic ID matching, potential semantic associations between services are mined, and an association strength matrix is ​​constructed to improve the service feature representation, enhance the original user-service interaction graph, make up for the defect of poor representation learning under sparse interaction, and enable the model to learn more comprehensive service features and interaction rules. (3) By constructing enhanced interactive views specific to text and category modalities and inputting them into LightGCN for graph convolution calculation, the decoupled representation and accurate mining of user preferences of different modalities are realized, breaking through the limitation of existing technologies that can only obtain the overall user preferences. This provides specific preference basis for refined recommendation and advertising optimization on the supply side, and improves the personalization and accuracy of the recommendation model as a whole. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the pre-training process of RQVAE in the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the process for capturing user modal preferences in this invention.

[0023] Figure 3 This is an overall framework diagram of the present invention. Detailed Implementation

[0024] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0025] In service recommendation scenarios, we use and Let I and J represent the sets of users and services, respectively, where I and J are the total number of users and services. The modal characteristics of a service are represented as follows: ,in It is the space of real numbers. It is modal The dimension of the feature is B, which is the set of modalities. In this invention, we mainly consider two dimensions: B = {des, cat}, where des and cat represent the text description modal feature and the category modal feature of the service, respectively.

[0026] User-service interaction matrix Used to represent the observed historical interaction relationships between users and items, where if the user With items If there has been interaction, then the elements in the interaction matrix... ,on the contrary Then it is 0.

[0027] When a given user When considering the interacted Services and their modal information, the goal of multimodal recommendation is to predict user behavior. The Service that may be accessed in the future can be formalized as follows: in It has parameters The mapping function, User The next service that users are most likely to access. It is a matrix of historical interactions between users and all services, i.e., a historical interaction graph.

[0028] The above service recommendation scenarios can be corresponding to: (1) In the e-commerce application development scenario, developer D plans to develop an intelligent shopping guide application to automatically recommend products and jump to the e-commerce platform to place an order based on the user's shopping needs. The user is developer D, and the services include the product search service of an e-commerce platform, the online payment service of a payment platform, and the logistics tracking service of a logistics company.

[0029] Therefore, the technical problem of this invention can be concretized as follows: a user has already invoked product search, but has not yet invoked payment and logistics services. How can the system quantify whether the user is most likely to invoke online payment service or logistics tracking service next, and thus generate a recommendation service?

[0030] (2) In the content creation tool development scenario, developer F plans to develop a short video intelligent editing tool to automatically add background music, subtitles and special effects after uploading a video. The user is developer F, and the services include video processing services of a cloud platform, music recommendation of a music platform and subtitle generation services of an AI platform.

[0031] Therefore, the technical problem of this invention can be concretized as follows: the user has already called the video processing service, but has not yet called the music recommendation service and the subtitle generation service. How can the system quantify whether the user is most likely to call the music recommendation service or the subtitle generation service next, and thus generate a recommendation service?

[0032] It should be noted that the method of the present invention is not limited to the two scenarios mentioned above, and can be applied to any scenario involving user and service recommendations.

[0033] Based on the above application scenario analysis, this example proposes a multimodal service recommendation method. The overall framework of the recommendation method is as follows: Figure 2 and 3 As shown, the specific steps are as follows: (1) Collect information from different modalities such as language descriptions and visual images in existing services, and collect historical interaction relationships between users and services.

[0034] This invention first utilizes the official registration interface provided by the Programmable Web platform to collect full description metadata of Mashups and Services in batches, as well as the actual call association information between Mashups and Services.

[0035] The full description metadata of a Mashup includes the Mashup name, functional description text for developers, and the category to which the Mashup belongs. The functional description text is mainly used to describe the business functions and actual application scenarios of the Mashup. The full description metadata of a Service includes the Service name, functional description text for developers, the category to which it belongs, and the service provider.

[0036] Under this premise, this method treats the functional description text and category information in the Service as two heterogeneous modalities, performs independent preprocessing operations on each, and then sequentially completes feature encoding through a pre-trained language model and a category embedding model (in this invention, the pre-trained language model and the category embedding model uniformly adopt the corresponding structure in BGE-M3) to obtain the primary modality feature vector. ,in It is a textual description modal feature of the Service. These are the category modal features of the Service, thus providing a semantically unified and information-rich basic representation for subsequent multimodal feature fusion and recommendation-related calculations.

[0037] It should be noted that this invention does not limit the number of modalities of a Service. In specific operations, different parts of the full description metadata of a Service can be treated as different modalities and processed separately as needed.

[0038] Since Mashup is a hybrid service developed by the user, the calling relationship between Mashup and Service can be used as the calling relationship between the user and Service, thus obtaining the historical interaction matrix A, which is the historical interaction graph.

[0039] (2) Construct multiple residual quantization variational autoencoders (RQVAEs) and pre-train them to obtain semantic identifiers of the Service in different modalities. RQVAE is constructed based on a discrete mapping mechanism of feature encoding and residual vector quantization of variational autoencoders. It can encode the continuous feature vectors of each modality of a Service into structured and discrete semantic identifiers. Through hierarchical residual quantization, it achieves a tight mapping from continuous features to fixed semantic IDs. This not only preserves the deep semantic associations of features from different modalities, but also transforms high-dimensional continuous features into low-redundancy, indexable, and alignable discrete semantic codes. This effectively improves the representation stability and retrieval matching efficiency of multimodal features, and provides standardized and highly discriminative semantic identifiers for subsequent multimodal fusion and recommendation computation.

[0040] The RQVAE (Residual Quantized Variational Auto Encoder) constructed in this invention mainly consists of an encoder, a decoder, and multiple quantizers containing codebooks, wherein the decoder is only used in the pre-training stage of RQVAE.

[0041] like Figure 1 The following is the pre-training process for RQVAE, with the specific steps as follows: First, random initialization. L Different codebooks , where L represents the number of quantizers.

[0042] Secondly, the primary modality feature vector is encoded using the RQVAE encoder. Mapped to source respectively L Different codebooks Discrete coding sequences, each layer of the codebook has ,in For the Lth level codebook, For the first k The first layer of the codebook y Each code character.

[0043] After the hidden state is mapped to the vector by the encoder, the discrete encoded sequence is sequentially input into multiple quantizers. The hidden state is quantized using residuals: the first quantizer matches the input discrete encoded sequence with the nearest neighbor in the codebook; the second quantizer matches the residual result of the first quantizer with the nearest neighbor in the codebook; and subsequent quantizers iteratively execute this process. (In a specific embodiment of this invention, a three-layer residual quantization is designed, namely…) L =3, corresponding to three codebooks each containing 128 codewords), the specific calculation logic is as follows: in, The input to the encoder, denoted as RQVAE, is typically the initial characteristics of the service. Indicates the first l Service characteristics after layer matching Indicates the first l The closest vector in the layer, i.e., the first... l The codewords matched by the layer; For the first l The index of the codeword matched by the layer. The encoder representing RQVAE consists of multiple layers of MLP.

[0044] After residual quantization through L layers, the modal features of Service can be mapped to the corresponding semantic ID information. At this point, the semantic ID representation vectors output by each quantizer are merged and restored to the original modal features through the decoder, which are then used for reconstruction loss. in, This represents the characteristics of the decoded item. The decoder for RQVAE also consists of a multi-layered MLP.

[0045] After obtaining the decoded modal features, in order to train the RQVAE structure, this invention mainly constructs two corresponding losses: reconstruction loss and reconstruction loss. and submitting losses .

[0046] The reconstruction loss is mainly used to constrain the difference between the reconstructed vector and the original input vector, ensuring that the semantic ID can accurately restore the core semantics of the modality features. Its expression is as follows: in, Represents the mathematical expectation. Represents the conditional probability distribution. This represents the latent vector output by the encoder; This represents the L2 norm.

[0047] The submission loss is primarily used to constrain the "binding relationship" between the quantized discrete codes and continuous latent vectors, preventing codebook collapse (i.e., some codebook vectors being idle), thereby ensuring that each semantic ID can be effectively mapped to its corresponding modal feature. Its expression is as follows: in, This indicates that the gradient is not calculated. This represents the hyperparameter used to balance the intensity of losses on both sides.

[0048] Finally, during the pre-training phase, the total loss of RQVAE is: in, and These are the weight parameters.

[0049] The RQVAE parameters are updated in reverse according to the loss function until convergence, resulting in a trained codebook and an RQVAE that can map all modal information of the Service to semantic IDs. The specific representation of the RQVAE is as follows: in, This represents a function used to extract modal information from an item. Indicates the extracted first j Semantic encoding of an item.

[0050] Based on the above pre-training method, this invention constructs a corresponding RQVAE for each modality of the Service and performs pre-training operations, ultimately training two specific modality-aware RQVAEs to obtain the Service modality semantic ID under the corresponding modality, as specifically represented as follows: in, and They represent the first j Semantic encoding of the text description modality and category modality of a Service. and They represent the first j The feature vectors of the text description modality and category modality of a Service.

[0051] (3) Based on the semantic identifiers between items, the correlation of items under different modal features is obtained, and the different weights of items under the current modality are represented according to the proportion of similar identifiers. Based on the semantic ID number obtained in step (2), this step mainly analyzes the similarity between current services under a specific modality. In order to represent the strength of the correlation between services, this invention mainly uses the method of comparing the number of semantic ID matches between each service for analysis. The specific formula is as follows: in, Indicates an indicator function, This function represents the degree of matching between two services. t Indicates the first t A Service.

[0052] After calculating the number of identical semantic IDs between each pair of nodes, the Service association strength matrix based on semantic IDs can be obtained. Where n is the total number of Services, and the expression for each element in the matrix is: .

[0053] By calculating the association strength matrix of each Service under different modalities, we can obtain .

[0054] (4) Based on the weights between Services, establish Service-Service related edges with different weights in the original user-Service interaction graph to construct user-Service interaction graphs under multiple views. After obtaining the association strength matrix under different modalities in step (3), in order to capture the specific preference strength of users when selecting Services, this invention further enhances the initial historical interaction graph and constructs the enhanced view under the corresponding modal view based on the Service association strength matrix under different modalities.

[0055] The specific steps are as follows: In the initial interaction diagram First, the interaction diagram is expanded into To facilitate subsequent graph convolution: In this context, T in the upper right corner represents the transpose, and m represents the number of users.

[0056] Based on this, in order to capture user preferences in different modalities, this invention further enhances the Service association strength matrix obtained in step (3) for different modalities, thereby obtaining two enhanced user-Service interaction views. , respectively represented as: in, This refers to the hyperparameters used to control the strength of the correlation matrix.

[0057] (5) Construct a collaborative filtering module to perform graph convolution under the interactive view enhanced with different modal information to obtain user preferences for different modalities. In the collaborative filtering module, this invention first randomly initializes an initial collaborative filtering feature for each Serive and Mashup. , Based on the enhanced historical interaction graph U, LightGCN is used to construct Service and Mashup representations with collaborative information.

[0058] The LightGCN formula is: in, It is the Laplace diagram norm. and These are Mashup a and services. First-order neighbors in the enhanced historical interaction graph U and These are services and Mashup a in the Embedding after layer propagation, where and Initialize to learnable parameters , .

[0059] After obtaining the embeddings of each layer, the layers are combined to obtain the collaborative filtering representation. , : in, Indicates the first k Service graph convolutional features of the layer Indicates the firstk Mashup graph convolutional features of the layers.

[0060] Finally, LightGCN can be represented as: To capture user and service features under different modal augmented views, based on the aforementioned graph convolution formula, this invention performs convolution on the two augmented interaction graphs respectively. By performing graph convolution on the graph, user and service features were obtained from two different views: in, , This represents user preferences under textual descriptions and category modalities, while , This represents the characteristics of a Service in terms of text description and category modality.

[0061] (6) Predict user and item preferences and representations under different modalities, and select items that users are interested in for recommendation.

[0062] Obtaining user preferences in different modalities , Next, the interest score of the Service needs to be predicted based on the user's preferences. To fully explore the potential user interactions across different modalities, this invention fully interacts with the learned features, for example, for the user... and Service The predicted interaction probability is: in, This represents the sigmoid activation function. This represents a multilayer perceptron. This represents the vector concatenation operation. This indicates multiplication by element position (dot product). and They represent Services j Features in categorical modality and textual description modality, and They represent Mashup iFeatures in categorical modality and textual description modality, This represents the interest vector obtained from the text description modal based on Mashup and the text description modal based on Service. This represents the interest vector obtained based on the text description modality of Mashup and the category modality of Service. This represents the interest vector obtained based on the category modality of Mashup and the textual description modality of Service. This represents the interest vector obtained based on the category modalities of the Mashup and the category modalities of the Service.

[0063] Bayesian Personalized Ranking (BPR) loss has been widely used as the primary optimization objective in recent service package recommendation models. It works by predicting the interaction probabilities between the Mashup and both positive and negative service samples, and optimizing the model to ensure that the predicted scores for positive samples are significantly higher than those for negative samples. The formal definition of the BPR loss function is as follows: Where r and h are the positive sample Service of Mashup a and the negative sample randomly selected from the negative sample Service of Mashup a, respectively. A triple represents a set of Mashups, positive sample services, and negative sample services. For each Mashup, the positive sample service is the service that has interacted with it, and the negative sample service is the service that has not interacted with it. These services are randomly sampled from service nodes that are not linked to the prediction Mashup.

[0064] Based on the final loss function The process involves updating all learnable parameters in reverse order until convergence, thereby obtaining the collaborative filtering representation of each Service after training. And a Service lookup table containing all Services. It's important to note that the types of Services in the Service lookup table are limited by the data source selected during the training phase.

[0065] During the inference phase, for the user-input Mashup, the requirements provided by the Mashup are encoded into corresponding feature vectors. Based on the collaborative filtering representation of each Service in the Service retrieval table, similarity calculation methods such as cosine similarity are used to calculate the similarity between all Services and the input Mashup, which is the predicted score of the Service. The Services are then sorted from high to low according to their scores, and the top-ranked Services are selected as recommended services.

[0066] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A service recommendation method based on representation decoupling and user modality preference capture, characterized in that, include: S1. Collect multiple historical users and multiple services from the target service recommendation scenario, and construct a historical interaction graph based on the historical interaction relationships between all historical users and all services; input the functional text description and category of each service into the first RQVAE and the second RQVAE respectively to obtain the first semantic code and the second semantic code; S2, calculate the number of semantic matches of functional text descriptions among all services based on the first semantic encoding, and then obtain the first association strength matrix; calculate the number of semantic matches of categories among all services based on the second semantic encoding, and then obtain the second association strength matrix; S3, the historical interaction graph is enhanced using the first association strength matrix and the second association strength matrix respectively to obtain the first modality enhanced graph and the second modality enhanced graph; S4, input the first modality augmented graph, the second modality augmented graph, the service collaborative filtering representation, and the user collaborative filtering representation into the graph convolutional network to further obtain the interaction probability between each user and each service; S5. Calculate the BRP loss based on the interaction probability. Update the service collaborative filtering representation, user collaborative filtering representation and graph convolutional network parameters based on the BRP loss. Repeat S4-S5 until the BRP loss converges to obtain the trained service collaborative filtering representation and form a service retrieval table. S6. Calculate the similarity score between the user to be recommended and each service in the service retrieval table, and select the top K services with the highest similarity scores as the services to be recommended to the user.

2. The service recommendation method based on representation decoupling and user modality preference capture according to claim 1, characterized in that, In step S1, a historical interaction graph is constructed based on the historical interaction relationships between all historical users and all services, specifically as follows: First, a historical interaction matrix is ​​constructed based on the historical interaction relationships between all historical users and all services. In the historical interaction matrix, if there is a historical interaction between a historical user and a service, the element at the corresponding position in the matrix is ​​1, and otherwise it is 0. Secondly, based on the historical interaction matrix, a historical interaction graph is further constructed.

3. The service recommendation method based on representation decoupling and user modality preference capture according to claim 1, characterized in that, In S1, both the first RQVAE and the second RQVAE are pre-trained RQVAEs. The pre-training process of the first RQVAE is as follows: S101, construct the first RQVAE containing an encoder, a decoder, and several quantizers containing codebooks; at the same time, randomly initialize the codebook of each quantizer; S102, the functional text descriptions of all services are encoded into discrete encoded sequences by an encoder; S103, the discrete encoded sequence is sequentially input into several quantizers, and each quantizer performs quantization processing through residuals; the outputs of each quantizer are then combined and restored to modal features by a decoder; S104, the reconstruction loss is calculated based on the modal features; the submission loss is calculated based on the discrete encoded sequence; the reconstruction loss and the submission loss are added together according to the weights to obtain the total loss; S105, update all parameters of the first RQVAE according to the total loss until the total loss converges, and obtain the codebook and the first RQVAE that has been trained. The pre-training method for the second RQVAE is the same as that for the first RQVAE, wherein the input of the second RQVAE encoder is the category of all services.

4. The service recommendation method based on representation decoupling and user modality preference capture according to claim 3, characterized in that, In step S103, each quantizer performs quantization using a residual method, specifically as follows: The first quantizer matches the input discrete encoded sequence with the nearest codeword in the current quantizer codebook and outputs the matching result. The second and all subsequent quantizers obtain the corresponding matching results by matching the output of the previous quantizer with the nearest codeword in the current quantizer codebook.

5. The service recommendation method based on representation decoupling and user modality preference capture according to claim 3, characterized in that, In S104, the formula for calculating the reconstruction loss is as follows: ； in, Indicates the reconstruction loss. Represents the mathematical expectation. Represents the conditional probability distribution. Indicates the encoder input, This represents the latent vector output by the encoder. This represents the modal characteristics output by the decoder. Represents the L2 norm; The formula for calculating the submission loss is as follows: ； in, Indicates the submission of losses. This indicates that the gradient is not calculated. Indicates hyperparameters, For the first l The index of the codeword matched by each quantizer, where L represents the number of quantizers. Indicates the first l Service features after matching each quantizer Indicates the first l The vector closest to the quantizer, i.e., the quantizer l The codewords matched by the quantizer; j represents the j-th service.

6. The service recommendation method based on representation decoupling and user modality preference capture according to claim 1, characterized in that, In step S2, the number of semantic matches in the functional text descriptions among all services is calculated, specifically as follows: ； in, Indicates an indicator function, This function calculates the matching degree between two services, where t represents the t-th service. This represents the first semantic encoding of the t-th service.

7. The service recommendation method based on representation decoupling and user modality preference capture according to claim 1, characterized in that, Specifically, S3 is: S301, transpose the historical interaction graph to obtain the transpose matrix of the historical interaction graph; S302, combine the historical interaction graph, the transpose of the historical interaction graph, and the two all-zero matrices in a block matrix form to obtain the initial enhanced graph; S303, the initial enhancement map is further enhanced using the first correlation strength matrix to obtain the first modal enhancement map; the initial enhancement map is further enhanced using the second correlation strength matrix to obtain the second modal enhancement map, specifically as follows: ； ； in, This is a hyperparameter, where m is the number of users. and These represent the first correlation strength matrix and the second correlation strength matrix, respectively. Represents a historical interaction diagram. and These represent the first and second mode augmentation maps, respectively, in the upper right corner. This indicates transpose.

8. The service recommendation method based on representation decoupling and user modality preference capture according to claim 1, characterized in that, In step S4, the user's preferences in the text description modality and the service features in the text description modality are obtained based on the first modality enhancement graph, specifically as follows: S401, initialize the corresponding initial user collaborative filtering features and initial service collaborative filtering features for each user and service respectively; S402, the initial user collaborative filtering features and the initial service collaborative filtering features are passed through each layer of LightGCN, and combined with the first modality augmentation map to obtain the user embedding vector and service embedding vector output by each layer of LightGCN. S403, combine the user embedding vectors of all layers of LightGCN to obtain the user's preferences in the text description modality; combine the service embedding vectors of all layers of LightGCN to obtain the service features in the text description modality.

9. The service recommendation method based on representation decoupling and user modality preference capture according to claim 1, characterized in that, In S4, the formula for calculating the probability of interaction between the user and the service is as follows: ； in, This represents the sigmoid activation function. This represents a multilayer perceptron. This represents the vector concatenation operation. This represents the interest vector obtained based on user i's preferences in the functional text description modality and service j's features in the functional text description modality. This represents the interest vector obtained based on user i's preferences in the functional text description modality and service j's features in the category modality. This represents the interest vector obtained based on user i's preferences in the category modality and service j's features in the functional text description modality. This represents the interest vector obtained based on user i's preferences in the category modality and service j's features in the category modality.

10. A service recommendation method based on representation decoupling and user modality preference capture according to claim 9, characterized in that, The interest vector is obtained based on user i's preferences in the functional text description modality and service j's features in the functional text description modality. The calculation formula is as follows: ； in, This indicates multiplication by element position. This represents the characteristics of service j in the functional text description modality. This indicates user i's preferences in the functional text description modality.

Images