A recommendation method for multi-modal service features based on meta-path fusion

Abstract

The application discloses a kind of based on the recommendation method of multi-modal service feature of meta-path fusion, belong to service recommendation and representation learning field.It includes: collection Mashup application and Service service, and constructs historical call matrix and the multi-modal feature of Service service;Meta-path and similarity calculation are used to Service service, respectively obtain meta-path isomorphism and modal isomorphism;Introduce multilayer perceptron and filter, further combined with all isomorphism, contrast loss and fusion modal feature are obtained by graph convolution network;Based on fusion modal feature and historical call matrix, respectively calculate to obtain second contrast loss and BRP loss;Regularization loss is introduced, to minimize all loss sum reverse training all network parameters, obtain the service retrieval table of training completion.The method of the application can accurately fuse the features of multiple modalities of traditional services, significantly enhance the generalization capability in cold start and sparse scenarios.

Description

Technical Field

[0001] This invention belongs to the field of service recommendation and representation learning, and specifically relates to a recommendation method based on multimodal service features of meta-path fusion. Technical Background In today's era of accelerated digital transformation, emerging software technologies such as cloud computing, mobile computing, and blockchain are experiencing explosive growth. Their distributed architecture, high availability, and trusted collaborative capabilities are profoundly reshaping the technological ecosystem of enterprise applications. This technological revolution has prompted many enterprises and organizations to proactively adjust their development strategies, encapsulating core business data, critical computing resources, and specialized business capabilities into standardized web services and publishing them on the internet. Through web application programming interfaces (Web APIs), they can achieve cross-system and cross-domain resource collaboration and value transfer, thereby maximizing the utilization of data assets and strengthening their core competitive advantages in the market.

[0002] To overcome the limitations of traditional web services, service composition solutions based on mashup technology have emerged. The core of this approach is to integrate dispersed atomic-level web services with multi-source data resources through a loosely coupled architecture. This allows developers to flexibly integrate existing web service resources and quickly build composite applications that meet complex business needs, thus gaining widespread popularity among enterprises and developers.

[0003] In Mashup service composition development scenarios, service developers need to select suitable service components from a massive number of atomic-level web services and build composite applications with complex business functions through multi-service collaborative integration to meet diverse actual business needs. With the large-scale development of cloud computing and web service technologies, service resources in the internet ecosystem are experiencing explosive growth. Various services differ significantly in terms of functionality, performance, interface protocols, and compatibility. The expansion and heterogeneity of the service resource pool mean that service developers need to accurately select from a massive number of candidate services—not only matching core business function requirements but also considering multiple dimensions such as service compatibility, response efficiency, and reliability. This process significantly increases the decision-making complexity, substantially increasing the time cost and difficulty of service selection, and seriously affecting the development efficiency and quality of Mashup applications.

[0004] Against this backdrop, service recommendation systems based on the needs of service developers have become a key technical solution to address the aforementioned pain points. Their core objective is to reduce the decision-making threshold for service selection and improve the accuracy and efficiency of service combination through intelligent recommendation mechanisms.

[0005] However, current mainstream service recommendation systems still have technical limitations: (1) Most existing solutions focus on modeling and utilizing the single-modal features of services (such as functional description text, performance index data, etc.), making it difficult to comprehensively and collaboratively model and integrate the multimodal features of services (including functional semantic features, interface structure features, running performance features, historical call feedback features, etc.). This results in biases in the representation learning of services by the recommendation system, making it impossible to fully capture the comprehensive characteristics of services and the potential needs of developers, thereby affecting the accuracy and applicability of the recommendation results and making it difficult to meet the actual needs of service composition development. (2) Existing methods usually use pre-trained models to extract different modal features, which results in a lot of noise in the modal feature extraction stage. Furthermore, it is not possible to extract features suitable for recommendation tasks through pre-trained models. Therefore, the current multimodal service recommendation has the problem of poor quality of initial modal features. (3) Current fusion strategies typically perform fusion only once, which can easily lead to misalignment during representation learning at different stages, lacking fine-grained alignment strategies. Therefore, there is an urgent need for a service recommendation model that can accurately fuse features from multiple modalities of traditional services for recommendation. Summary of the Invention

[0006] This invention addresses the problems of modal denoising, incomplete modal feature fusion, and poor initial modal representation learning in existing methods, and proposes a recommendation method for multimodal service features based on meta-path fusion.

[0007] S1. Collect multiple Mashup applications and Service services, and construct a historical call matrix, which consists of historical call records of all Mashup applications and all Service services; encode the functional text description and category of each Service service to obtain multimodal features containing both functional text description and category modalities; S2, based on the historical call matrix, obtains the meta-path Service isomorphism graph by defining the first meta-path and the second meta-path; calculates the similarity between services and forms service pairs according to the multimodal characteristics of all services, and all service pairs form a modal service isomorphism graph containing two modalities; S3 uses a multilayer perceptron and filters to convert the multimodal features of each Service into a final embedding vector; the final embedding vector is then input into a graph convolutional network along with the metapath Service isomorphism graph and the modal Service isomorphism graph to obtain fused modal features, and the first contrastive loss is further calculated. S4, construct the Mashup-Service isomorphic graph from the historical call matrix, and calculate the second contrastive loss and BPR loss by combining the fused modal features; S5, introduce regularization loss, add the first contrastive loss, the second contrastive loss, the BPR loss and the regularization loss to obtain the total loss; minimize the total loss to train all parameters in the encoder, filter, multilayer perceptron and graph convolutional network in reverse to obtain the trained Service retrieval table; the Service retrieval table contains all services; S6. Calculate the similarity score between the Mashup application to be recommended and each Service in the Service retrieval table, and use the Service with a score higher than the threshold as the recommended service for the Mashup application to be recommended.

[0008] Furthermore, in S2, the first meta-path and the second meta-path are respectively: ; ; Where A is the historical call matrix, For the first-order path, This is the second-order path, and the T in the upper right corner is the transpose.

[0009] Furthermore, S2 includes a Service isomorphism graph with two modalities, specifically: For each Service, the similarity between it and each of the other Services is calculated based on multimodal features. The similarity results are sorted from high to low, and the top K Services are paired with the current Service to form Service pairs. Traverse all Services to obtain all Service pairs, and further construct a Service matrix. The element values ​​in the matrix are the corresponding similarity results. Normalizing the item matrix yields a Service isomorphism graph containing two modalities.

[0010] Furthermore, in step S3, the final embedding vector for each Service is determined through the following steps: First, using a multilayer perceptron, the multimodal features of the i-th Service are mapped to a latent space of a unified dimension to form intermediate projection features. Then, using a fast Fourier transform, the intermediate projection features are converted to a frequency spectrum to obtain the corresponding complex spectrum features. Secondly, a learnable weighted filter is constructed for each modality of the i-th Service, and a dot product operation is performed with the complex spectral features: ; in, Let m be the modal spectrum feature of the i-th Service after filtering, where m is the total modality, which includes two modalities, a and c. Here, a is the functional text description of the Service, and c is the category of the Service. Let m be the learnable weighted filter corresponding to mode m. For element-wise multiplication, The modal spectrum features of the i-th Service after Fourier transform are given. It is the space of real numbers; Finally, through inverse fast Fourier transform, By restoring to the real number domain, we obtain the final embedding vector corresponding to the i-th Service.

[0011] Furthermore, in S3, the first contrastive loss is calculated using contrastive learning, and the specific formula is as follows: ; ; ; in, Collaborative filtering features for Services Collaborative filtering features for negative samples of Service. For the characteristics of the Service under submodal a, For the characteristics of the Service in submodal c, For temperature coefficient, A collection of all Services. Let exp be the negative sample Service corresponding to Service s, and let exp be the exponential function. For the loss of all services under submodal a, For the loss of all Services under submodal c, As the first comparison loss, This is a collection of negative sample services.

[0012] Furthermore, the negative sample Service services of a Service are specifically defined as follows: for each Service service, the Service services that do not form a Service pair with it are the negative sample Service services corresponding to the current Service service, and a set of negative sample Service services is obtained by randomly sampling from all negative sample Service services.

[0013] Furthermore, in S4, the second contrastive loss is calculated using contrastive learning, specifically as follows: ; in, For a submode in mode m, Service Collaborative filtering characterization, Service Submodal characterization, Service Submodal characterization, For temperature coefficient, Service The corresponding negative samples, This is the set of negative sample Services corresponding to the Service. A collection of all Services.

[0014] Furthermore, in S4, the BPR loss is calculated using Bayesian personalized ranking, and the specific calculation formula is as follows: ; ; ; Where r and h are the positive sample service from the Mashup application p and the negative sample service from the Mashup application p, respectively. A triple represents a set of Mashup applications, positive samples that have interacted with them, and negative samples that have not interacted with them; For the sigmoid function, and These are the final representations of Service r and Service h, respectively. The final representation of p applied to Mashup, and These are intermediate parameters.

[0015] Furthermore, the negative sample service of a Mashup application is specifically defined as follows: for each Mashup application, the service that has not interacted with it is the negative sample service corresponding to the Mashup application.

[0016] The beneficial effects of this invention are: (1) In view of the problem that the existing technology directly reuses pre-trained features, the recommendation task cannot intervene, and irrelevant information is amplified with the gradient, the present invention embeds learnable frequency domain-channel dual filtering before the features enter the recommendation process, and removes redundant signals that are irrelevant to the calling behavior in real time, so that the subsequent layers always operate on clean representations, fundamentally improving robustness.

[0017] (2) In view of the lack of alignment constraints in existing single-stage fusion, and the semantic drift between modalities due to dimensional differences and conflicting optimization objectives, this invention uses a three-level progressive alignment of “encoding-fusion-decoding” to connect mutual information maximization, dynamic weight calibration and distillation loop constraints to ensure that the representation of each modality is updated synchronously in space, scale and semantics throughout the process, and eliminates the recommendation bias caused by alignment error.

[0018] (3) In view of the problem that existing solutions rely only on the “Mashup-Service” bipartite graph, the node neighborhood is single and the training signal is sparse, the present invention constructs Service isomorphic graph and Mashup compatible graph in parallel to form a multi-topological view, so that the same modality can be fully learned on the most suitable structure first, and then complemented by a learnable cross-view fusion gate, which significantly enhances the generalization ability in cold start and sparse scenarios, while maintaining logarithmic linear complexity and not increasing the additional computing power burden. Attached Figure Description

[0019] Figure 1 This is an overall framework diagram of the method of the present invention.

[0020] Figure 2 This is a flowchart illustrating the modal feature extraction module.

[0021] Figure 3 This is a flowchart illustrating the process of the initial integration module.

[0022] Figure 4 This is a flowchart illustrating the collaborative filtering feature extraction module.

[0023] Figure 5 This is a flowchart illustrating the process of the later-stage fusion module. 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 the recommended scenario, let the Mashup set be... The Service collection is Where I and J represent the total number of Mashups and Services, respectively. In a learnable recommendation framework, each Mashup... With Service Each corresponds to a trainable embedding vector: .in d For the embedding dimension, the index is either Mashup or Service with the id.

[0026] Furthermore, the multimodal characteristics of a Service are denoted as , For modality Regarding the feature dimensions, this method considers both descriptive and categorical modalities.

[0027] The historical call matrix is ​​composed of the historical call relationships between all Mashups and all Services in the collection. This is used to record observed call relationships, where, for element values ​​in the matrix, when the element value... When the value is 0, it means that the Mashup corresponding to the row has called the Service corresponding to the column. When the element value is 0, it means that the Mashup corresponding to the row has not called the Service corresponding to the column.

[0028] Given a set of Mashups, consider the call history of a Mashup for all Services in the Service set. and modal information for each Service. The goal of the multimodal recommendation task is to train a parametric mapping function. This leads to the obtained embedding vectors after Service training. .

[0029] in, For Mashup The predicted score vector of the Service most likely to be called next.

[0030] Based on the above principles, this invention proposes a method for multimodal service recommendation, such as... Figure 1 As shown, the specific steps include: (1) Collect information on different modalities, such as language descriptions and categories, from existing Mashup applications and Service services. Also collect historical call relationships between Mashups and Services. And use pre-trained models to initially encode the modal features of each service into modal features.

[0031] like Figure 2 As shown, in the modal feature extraction module, this invention first calls the open registration interface provided by the Programmable Web official website to batch retrieve the complete description metadata of Mashups and Services. The complete description metadata of a Mashup includes the Mashup name, a textual description of the Mashup's functions for developers, and the Mashup category. The textual description of the Mashup's functions focuses on explaining the business capabilities and use cases that the target Mashup can provide. The complete description metadata of a Service includes the Service name, a textual description of the Service's functions, the Service category, information about the Service provider, and records of the actual call relationships between Mashups and Services.

[0032] Based on this, this method treats the "functional text description" field and the "category" field of the Service as two heterogeneous modalities. After preprocessing the two modalities independently, they are sequentially encoded by a pre-trained language model and a category embedding model (in this specific embodiment, both the pre-trained language model and the category embedding model are BGE-M3), thereby generating the corresponding primary modality feature vectors. , This provides a unified and semantically rich input representation for subsequent multimodal fusion and recommendation computation. The actual call relationships between all Mashups and all Services are recorded to form a historical call matrix A.

[0033] (2) Construct a denoising scheme based on the frequency domain model and obtain the denoised modal features.

[0034] In this step, the present invention abandons the traditional spatial attention mechanism and introduces a frequency domain processing framework to achieve intermodal information complementarity and noise suppression in one step.

[0035] The specific process is as follows: First, for the i-th Service, the original features of each modality are mapped to a latent space of a unified dimension using a multilayer perceptron, forming intermediate projected features: in, , These are the projected features and the initial Service modality features, respectively. , These are the learnable linear layer parameters.

[0036] Then, a Fast Fourier Transform is used to transform the above projection features to the frequency domain to obtain the corresponding complex spectral features: in, The features are those obtained after the Fourier transform, where n is the total number of Services. The Fourier transform function is given by the following formula: in, Let be a discrete signal sequence in the time domain, k be the index of the signal in the frequency domain, exp() be the exponential function, q be the variable for time-domain summation, and j be the imaginary unit.

[0037] Based on this, a learnable complex weighted filter is constructed for each mode, and point-by-point multiplication is performed on the spectral features to dynamically suppress high-frequency noise and mode-specific redundant information. in, The filtered single-mode spectral features, These are the trainable parameters for the corresponding linear layer. This is for element-wise multiplication.

[0038] Finally, the filtered single-mode spectral features are obtained through inverse fast Fourier transform. Synchronously restore to the real feature space, and output the final embedding vector that retains modality specificity and is rich in cross-modal semantics for use in subsequent steps (4).

[0039] (3) Construct Service isomorphism graphs, which mainly include isomorphism graphs based on Service modality similarity and Service isomorphism graphs based on metapath design.

[0040] To extract effective information about services across different modalities, this invention constructs different service isomorphic graphs to provide a technical foundation for subsequent GNN-based learning of modal representations. Specifically, this invention designs the following two types of service isomorphic graphs: 1) Service isomorphism graph built based on metapath.

[0041] Existing methods overemphasize the similarity between modalities, resulting in a lack of collaborative information within the modal data. This leads to the easy neglect of collaborative information in modal features during initial feature learning. Meta-path design can predefine paths in a graph and uncover the connections between corresponding nodes. Therefore, this invention, based on meta-path analysis, uses the original Mashup-Service call graph to retrieve meta-paths and constructs an isomorphic graph based on the start and end points of the meta-paths. In this invention, meta-paths are mainly constructed by retrieving "SMS" paths and "SMSMS" paths. The representation based on these two paths is as follows: After defining several meta-paths, this invention uses certain weights to fuse the subgraphs generated by the meta-paths, constructing a Service isomorphic graph based on meta-paths. : in, and These are different weighting coefficients.

[0042] 2) Service isomorphism graph constructed based on modal similarity.

[0043] Because the original platform only provides "Mashup–Service" call records and lacks explicit "Service–Service" relationships, the message source of the graph neural network is singular, making it prone to semantic drift and gradient explosion. Furthermore, visual screenshots or text descriptions of different Services often contain comparable semantic signals, but these are not fully utilized. To address this, this invention introduces an additional Service isomorphism graph, using modal similarity as edge weights to construct an isomorphism graph based on modal features. This not only supplements structural information but also provides stable semantic anchors for subsequent cross-modal fusion.

[0044] like Figure 2 As shown, the specific construction steps are as follows: First, using the modal features of the Service side as input, the semantic association strength between any two Services is calculated through similarity, taking cosine similarity as an example: in, This represents the similarity score between Service 𝑖 and Service 𝑗 in modality 𝑚. It is a norm.

[0045] To extract the most relevant features from the neighbors, this invention only retains the top K similar edges for each Service: in, Let be the edge weight between Services and in modality . After pairing each service into a service pair, the final result is a service matrix composed of all service pairs. .

[0046] To mitigate the explosive gradient problem, the Service-Service matrix is ​​further normalized as follows: in, for The angle matrix, This is the final Service isomorphic graph obtained from the construction.

[0047] (4) Construct a preliminary fusion module, obtain Service representations under different views based on the multiple Service isomorphic graphs constructed in step (3), and fuse preliminary modal representations based on different views.

[0048] like Figure 3 As shown, in step (3), the Service isomorphic graph generated based on the metapath is obtained. Modality-specific Serive isomorphism graphs generated based on modality similarity Then, the present invention will use the final embedding vector obtained in step (2) Each isomorphic graph is input into a graph convolutional network, and a GNN is used to perform graph convolution operations in each graph to obtain the Service representation under each view.

[0049] In this invention, LightGCN is mainly used for graph convolution. A single-layer LightGCN is as follows: in, For the Laplace diagram norm, and Service s First-order neighbors and in isomorphic graphs with s Connected Service s' First-order neighbors, For Service After the first Feature embedding after convolutional layers Neighboring nodes After the first Feature embedding after convolutional layers.

[0050] The initial input is the initial modal characteristics of the Service.

[0051] After obtaining the corresponding Service modal features in each Service isomorphic graph, in order to integrate collaboration information and modal information, this invention combines each layer to perform preliminary fusion of Service features, so that the currently obtained Service modal features can contain a portion of collaboration information.

[0052] This invention uses contrastive learning for the first step of preliminary fusion. For each Service, its positive samples are the other Service nodes that form a Service pair with it, and its negative samples are Service nodes that do not form a Service pair with it, and are randomly sampled from them.

[0053] Specific loss function as follows: in, For Service The collaborative filtering features are obtained by convolution of the Service isomorphic graph generated based on meta-paths; The collaborative filtering features of the Service as negative samples are obtained by convolution of the Service isomorphic graph generated based on the metapath. For Service exist Modal features For Service exist Modal features For temperature coefficient, A collection of Services For Service The corresponding negative samples, A collection of negative sample services.

[0054] The first step of fusion enables the two modal information to acquire some collaborative information, thus completing the early fusion of modal information and collaborative information.

[0055] After comparative learning, it can be assumed that some collaborative information has been incorporated into the original modal features. At this point, the features of all services under modality m are... It is considered to be a modal feature after collaborative information fusion, and is denoted as ,and It contains two modes.

[0056] (5) Construct a collaborative filtering feature extraction module to obtain collaborative information based on historical interaction data.

[0057] like Figure 4 As shown, in the collaborative filtering module, this invention uses graph convolution based on the historical call matrix A to initialize a collaborative filtering feature for each Service and each Mashup, thus constructing Service and Mashup representations with collaborative information.

[0058] The formula for graph convolution is: in, For the Laplace diagram norm, and They are Mashup p and Service respectively First-order neighbors, and They are Mashup p and Service respectively In the Embedding after layer propagation.

[0059] After obtaining the embeddings of each layer, the layers are combined to obtain collaborative filtering representations, and then the final representations of the Service and Mashup are obtained through a local attention mechanism. , , recorded as , : Where l is the number of layers in the GNN. This represents the features of Service after graph convolution at layer l. This represents the features of the Mashup after convolution of the l-th graph, where and Initialize to , After traversing all levels, and Updated to their respective final representations.

[0060] (6) Modal features and collaborative filtering features are fused in the later stage to generate unified Mashup and Service representations, and different types of feature descriptions are integrated to generate predictions.

[0061] like Figure 5 As shown, based on the results calculated in steps (4) and (5), in the later fusion module, the modal features are fused by calculating the InfoNCE loss. and final expression The feature space distance between them is reduced, so that they can be fused in the later stage and a prediction can be constructed.

[0062] in, For Service Collaborative filtering characterization, For Service Modal characterization, For the corresponding negative sample Service Modal characterization, For temperature coefficient, For Service The corresponding negative samples, This is the set of negative sample Services corresponding to the Service.

[0063] Bayesian Personalized Ranking (BPR) loss has been widely used as the primary optimization objective in recent service package recommendation models. It predicts the interaction probabilities between the Mashup and positive and negative service samples and optimizes the model to ensure that the prediction score of positive samples is significantly higher than that of negative samples.

[0064] The formal definition of the BPR loss function is as follows: Where r and h are the positive sample Service of Mashup p and the negative sample randomly selected from the negative sample Service of Mashup p, respectively. A triple represents a set of Mashups, positive samples that have interacted with them, and negative samples that have not interacted with them; For the sigmoid function, and These are the final representations of Service r and Service h, respectively. This is the final representation of Mashup p.

[0065] 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.

[0066] Finally, the multiple losses are combined to obtain the final loss function. : Among them, losses , and These correspond to the definitions provided in the first three sections. For L2 regularization, it is represented as , , and These are different weight coefficients, W and b are the weights and biases of the linear layer, respectively, and z is the z-th linear layer.

[0067] Based on the final loss function The network parameters are updated in reverse order until convergence, thus obtaining the completed embedding vector for each Service. And a Service lookup table containing all Services.

[0068] During the inference phase, for the input Mashup, the requirements provided by the Mashup are converted into corresponding feature vectors. Then, based on the embedding vector of each Service in the Service retrieval table, the predicted scores of all Services are calculated using similarity. The Services are then sorted from high to low according to their scores, and the top-ranked Services are selected as recommended services.

[0069] 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 recommendation method based on multimodal service features of meta-path fusion, characterized in that, include: S1. Collect multiple Mashup applications and Service services, and construct a historical call matrix, which consists of historical call records of all Mashup applications and all Service services; encode the functional text description and category of each Service service to obtain multimodal features containing both functional text description and category modalities; S2, based on the historical call matrix, obtains the meta-path Service isomorphism graph by defining the first meta-path and the second meta-path; calculates the similarity between services and forms service pairs according to the multimodal characteristics of all services, and all service pairs form a modal service isomorphism graph containing two modalities; S3 uses a multilayer perceptron and filters to convert the multimodal features of each Service into a final embedding vector; the final embedding vector is then input into a graph convolutional network along with the metapath Service isomorphism graph and the modal Service isomorphism graph to obtain fused modal features, and the first contrastive loss is further calculated. S4, construct the Mashup-Service isomorphic graph from the historical call matrix, and calculate the second contrastive loss and BPR loss by combining the fused modal features; S5, introduce regularization loss, add the first contrastive loss, the second contrastive loss, the BPR loss and the regularization loss to obtain the total loss; minimize the total loss to train all parameters in the encoder, filter, multilayer perceptron and graph convolutional network in reverse to obtain the trained Service retrieval table; the Service retrieval table contains all services; S6. Calculate the similarity score between the Mashup application to be recommended and each Service in the Service retrieval table, and use the Service with a score higher than the threshold as the recommended service for the Mashup application to be recommended.

2. The recommendation method based on multimodal service features according to claim 1, characterized in that, In S2, the first meta-path and the second meta-path are respectively: ； ； Where A is the historical call matrix, For the first-order path, This is the second-order path, and the T in the upper right corner is the transpose.

3. The recommendation method based on multimodal service features according to claim 1, characterized in that, S2 contains a Service isomorphism graph with two modalities, specifically: For each Service, the similarity between it and each of the other Services is calculated based on multimodal features. The similarity results are sorted from high to low, and the top K Services are paired with the current Service to form Service pairs. Traverse all Services to obtain all Service pairs, and further construct a Service matrix. The element values ​​in the matrix are the corresponding similarity results. Normalizing the item matrix yields a Service isomorphism graph containing two modalities.

4. The recommendation method based on multimodal service features according to claim 1, characterized in that, In S3, the final embedding vector of each Service is determined by the following steps: First, using a multilayer perceptron, the multimodal features of the i-th Service are mapped to a latent space of a unified dimension to form intermediate projection features. Then, using a fast Fourier transform, the intermediate projection features are converted to a frequency spectrum to obtain the corresponding complex spectrum features. Secondly, a learnable weighted filter is constructed for each modality of the i-th Service, and a dot product operation is performed with the complex spectral features: ； in, Let m be the modal spectrum feature of the i-th Service after filtering, where m is the total modality, which includes two modalities, a and c. Here, a is the functional text description of the Service, and c is the category of the Service. Let m be the learnable weighted filter corresponding to mode m. For element-wise multiplication, The modal spectrum features of the i-th Service after Fourier transform are given. It is the space of real numbers; Finally, through inverse fast Fourier transform, By restoring to the real number domain, we obtain the final embedding vector corresponding to the i-th Service.

5. The recommendation method based on multimodal service features according to claim 1, characterized in that, In S3, the first contrastive loss is calculated using contrastive learning, and the specific formula is as follows: ； ； ； in, Collaborative filtering features for Services Collaborative filtering features for negative samples of Service. For the characteristics of the Service under submodal a, For the characteristics of the Service in submodal c, For temperature coefficient, A collection of all Services. Let exp be the negative sample Service corresponding to Service s, and let exp be the exponential function. For the loss of all services under submodal a, For the loss of all Services under submodal c, As the first comparison loss, This is a collection of negative sample services.

6. The recommendation method based on multimodal service features according to claim 5, characterized in that, The negative sample service of a Service is defined as follows: for each Service, the Service that does not form a Service pair with it is the negative sample Service corresponding to the current Service. The negative sample Service is obtained by randomly sampling from all negative sample Service.

7. The recommendation method based on multimodal service features according to claim 1, characterized in that, In step S4, the second contrastive loss is calculated using contrastive learning, specifically as follows: ； in, For a submode in mode m, Service Collaborative filtering characterization, Service Submodal characterization, Service Submodal characterization, For temperature coefficient, Service The corresponding negative samples, This is the set of negative sample Services corresponding to the Service. A collection of all Services.

8. The recommendation method based on multimodal service features according to claim 1, characterized in that, In S4, the BPR loss is calculated using Bayesian personalized ranking, and the specific calculation formula is as follows: ； ； ； Where r and h are the positive sample service from the Mashup application p and the negative sample service from the Mashup application p, respectively. A triple represents a set of Mashup applications, positive samples that have interacted with them, and negative samples that have not interacted with them; For the sigmoid function, and These are the final representations of Service r and Service h, respectively. The final representation of p applied to Mashup, and These are intermediate parameters.

9. A recommendation method based on multimodal service features according to claim 8, characterized in that, The negative sample service of a Mashup application is defined as follows: for each Mashup application, the service that has not interacted with it is the negative sample service corresponding to the Mashup application.

10. A recommendation system based on meta-path fusion of multimodal service features, used to implement the recommendation method based on meta-path fusion of multimodal service features as described in claim 1, characterized in that, The system includes: The data collection module is used to collect data from multiple Mashup applications and Service services, and to build a historical call matrix, which consists of historical call records of all Mashup applications and all Service services. The modal feature extraction module is used to encode the functional text description and category of each Service to obtain multimodal features containing both functional text description and category modalities. Based on the historical call matrix, a meta-path Service isomorphism graph is obtained by defining the first meta-path and the second meta-path. Based on the multimodal features of all Services, the similarity between Services is calculated and Service pairs are formed. All Service pairs form a modal Service isomorphism graph containing both modalities. A multilayer perceptron and filters are used to convert the multimodal features of each Service into the final embedding vector. The initial fusion module is used to input the final embedding vector, along with the meta-path Service isomorphism graph and the modality Service isomorphism graph, into the graph convolutional network to obtain fused modality features, and further calculate the first contrastive loss. The collaborative filtering feature extraction module is used to construct a Mashup-Service isomorphic graph from the historical call matrix; The post-fusion module is used to calculate the second loss and BPR loss, introduce regularization loss, and add the first contrast loss, second contrast loss, BPR loss and regularization loss to obtain the total loss. The total loss is minimized to train all parameters in the encoder, filter, multilayer perceptron and graph convolutional network in reverse to obtain the trained Service retrieval table. The Mashup recommendation module calculates the similarity score between the Mashup application to be recommended and each Service in the Service lookup table, and recommends the Service with a score higher than a threshold as the recommended service for the Mashup application.

Images