Graph-based hierarchical hybrid enhanced service combination recommendation method
By introducing MLP residual embedding and a multi-level hybrid enhancement strategy, combined with the Mixup method, the problem of inaccurate user preference modeling in existing technologies is solved, and the accuracy and diversity of service composition recommendations are improved.
Patent Information
- Application Number
- CN202511080144.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-04
- Publication Date
- 2025-11-18
AI Technical Summary
Existing technologies rely on randomly initialized embeddings in service combination recommendations, which leads to users focusing excessively on sparse interaction data and failing to fully reflect potential preferences. Furthermore, graph neural networks are prone to oversmoothing and information perturbation during propagation, resulting in inaccurate recommendations.
We employ an embedded enhancement module based on MLP residuals and a multi-level hybrid enhancement strategy. By combining the graph-after-hybrid enhancement module and the cross-view hybrid enhancement module with the Mixup method, we capture explicit structural information and unstructured interaction features, alleviate the oversmoothing problem, and improve the user preference modeling capability.
Effective integration of explicit and implicit features enhances user preference modeling capabilities and recommendation quality, thereby improving the accuracy and diversity of service combination recommendations.
Smart Images
Figure CN120974100A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of service composition recommendation, and more particularly to a hierarchical hybrid enhancement service composition recommendation method based on graph structures. Background Technology
[0002] With the continuous evolution of e-commerce, online service platforms, and enterprise-level solutions, service composition is gradually shifting from traditional process-based integration to a more intelligent, user-centric recommendation model. Automatically generating and recommending service compositions based on user needs has become an important research direction in service computing and intelligent service systems. Compared to recommending individual services, service composition recommendations can provide complete and collaborative service solutions, thereby improving user task completion efficiency and experience, enhancing the overall perceived value of services, and also helping platforms improve service utilization and business revenue.
[0003] Early research primarily focused on user-service combination interaction data for user preference modeling, supplemented by user-individual service interaction data. In recent years, while more methods have attempted to incorporate graph neural networks into service combination recommendation modeling—constructing heterogeneous graphs between users and services, user-service combinations, and atomic services and service combinations, and using graph neural networks (such as GNNs and GCNs) to aggregate neighboring node information—this approach relies on randomly initialized raw embeddings. This leads to the learned user preferences overemphasizing existing, sparse interaction data, failing to comprehensively reflect users' potential preferences. In deeper graph propagation, oversmoothing of representations can easily occur, potentially weakening the model's ability to accurately capture user preferences. Although some methods introduce contrastive learning on top of graph neural networks, they do not fully consider the structural conflicts of features between views, easily introducing information perturbations during multi-view fusion and failing to effectively supplement features not covered by the graph topology, ultimately resulting in inaccurate learned user preferences. Summary of the Invention
[0004] In view of the aforementioned shortcomings of existing technologies, this invention provides a graph-based hierarchical hybrid enhancement service composition recommendation method. By introducing an embedding enhancement module based on MLP residuals, it is possible to model unstructured latent features and capture user preference information for atomic services and their combinations beyond graph propagation. By introducing a graph-based multi-level hybrid enhancement strategy in multi-view embedding learning, the graph post-hybrid enhancement module effectively integrates unstructured features while preserving structural information, alleviating the oversmoothing problem of graph neural networks. The cross-view hybrid enhancement module performs linear interpolation on user features from different views, thereby improving user preference modeling capabilities and recommendation quality.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A graph-based hierarchical hybrid enhancement service composition recommendation method includes the following steps:
[0007] Step 1: Initialize the original embedding; perform MLP enhancement on the original embedding based on the MLP residual to generate the MLP-enhanced embedding;
[0008] Step 2: Construct a graph structure propagation model and a bipartite graph;
[0009] The graph structure propagation model includes a graph post-hybrid enhancement module and a cross-view hybrid enhancement module;
[0010] Step 3: Input the bipartite graph, the original embedding, and the MLP-enhanced embedding into the LightGCN network of the graph post-hybrid enhancement module to generate the graph-enhanced original embedding and the graph-enhanced MLP embedding, respectively.
[0011] The graph-enhanced original embedding and the graph-enhanced MLP embedding are interpolated in the Mixup unit of the graph-enhanced post-graph mixing enhancement module to generate an interpolated embedding;
[0012] The interpolation embedding includes the final global user embedding and the final local user embedding;
[0013] Step 4: Input the final global user embedding and the final local user embedding into the cross-view blending enhancement module for cross-view interpolation to obtain the final user embedding.
[0014] Preferably, in step 1, the original embedding includes an initialized service composition embedding, an atomic service embedding, a global user embedding, and a local user embedding; the original embedding is then enhanced by MLP to generate MLP service composition embedding, MLP atomic service embedding, MLP global user embedding, and MLP local user embedding, respectively.
[0015] Preferably, in step 2, the bipartite graph includes a user-service composite bipartite graph and a user-atomic service bipartite graph.
[0016] Preferably, step 3 includes step 3-1:
[0017] The global user embedding, the service composition embedding, and the user-service composition bipartite graph are input into the LightGCN network to obtain graph-enhanced global user embedding and graph-enhanced service composition embedding.
[0018] The MLP global user embedding, the MLP service composition embedding, and the user-service composition bipartite graph are input into the LightGCN network to obtain the graph-augmented MLP global user embedding and the graph-augmented MLP service composition embedding.
[0019] The local user embedding, the atomic service embedding, and the user-atomic service bipartite graph are input into the LightGCN network to obtain graph-enhanced local user embedding and graph-enhanced atomic service embedding.
[0020] The MLP local user embedding, the MLP atomic service embedding, and the user-atomic service bipartite graph are input into the LightGCN network to obtain the graph-enhanced MLP local user embedding and the graph-enhanced MLP atomic service embedding.
[0021] Preferably, step 3 includes step 3-2:
[0022] The graph-enhanced MLP global user embedding and the graph-enhanced global user embedding are input into the Mixup unit and interpolated to generate the final global user embedding.
[0023] The graph-enhanced MLP local user embedding and the graph-enhanced local user embedding are input into the Mixup unit and interpolated to generate the final local user embedding.
[0024] Preferably, step 3-2 further includes:
[0025] The interpolation embedding also includes final atomic service embedding and final global service composition embedding;
[0026] The graph-enhanced MLP service composition embedding and the graph-enhanced service composition embedding are input into the Mixup unit for interpolation to generate the final global service composition embedding.
[0027] The graph-enhanced MLP atomic service embedding and the graph-enhanced atomic service embedding are input into the Mixup unit for interpolation to generate the final atomic service embedding; the final atomic service embedding is aggregated by average pooling to obtain the final local service composite embedding.
[0028] Preferably, the method further includes step 5, which involves embedding the end-user with both the global service composition embedding and the local service composition embedding.
[0029] Similarity is calculated to obtain the predicted score for each candidate service combination. The calculation formula is as follows:
[0030]
[0031] in, The end-user embedding, Embedded for the final global service composition, Embedded for the final service combination.
[0032] Compared with the prior art, the beneficial effects of the present invention are reflected in:
[0033] By combining the Mixup method with graph neural networks, this approach effectively enhances the embedding representation capability by simultaneously capturing explicit structural information and unstructured interaction features. Specifically, the method first initializes users with two independent embeddings in both the global and local views. A multilayer perceptron combined with a residual network is used to extract unstructured latent features from the initialized embeddings, constructing enhanced feature representations for subsequent user preference feature modeling. Secondly, a graph-based hierarchical mixing mechanism is employed. The Mixup method is used to linearly interpolate the explicit and latent features learned by the graph neural network within each view, mitigating the over-smoothing phenomenon that easily occurs during graph structure propagation. Building upon this, a second Mixup method is applied across views to enhance the diversity of multi-view representations, modeling diverse user preference expressions. Finally, a recommendation score is calculated to predict user preferences for service combinations, while a loss function is calculated to optimize the learned features, thereby improving the model's recommendation performance. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of the method framework of Embodiment 1 of the present invention;
[0035] Figure 2 This is a schematic diagram of the post-figure hybrid enhancement module of Embodiment 1 of the present invention (taking the global view as an example). Detailed Implementation
[0036] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.
[0037] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.
[0038] Example 1:
[0039] like Figure 1 The illustrated hierarchical hybrid augmented service composition recommendation system based on graph structure includes the following steps:
[0040] Step 1: Perform MLP enhancement on the original embedding based on the MLP residual to generate the MLP-enhanced embedding;
[0041] Initialize the original embedding and the multi-view user embedding; perform MLP enhancement on the original embedding and the multi-view user embedding based on the MLP residual:
[0042] To express users' multifaceted preferences, a global view and local views are first constructed, representing user preferences in combined service interactions and user preferences in interactions with atomic services, respectively. Then, independent user embeddings are initialized in both views to capture user preference features across different views. A residual augmentation network is constructed using a multilayer perceptron (MLP) to augment the original embeddings, thereby extracting unstructured implicit preference information from users across different views.
[0043] Step 1-1: Initialize Service Composition Embedding and atomic service embedding
[0044] Where C represents composite-level services, A represents atomic-level services, s represents service category embedding, and r represents random initialization.
[0045] Step 1-2: Initialize the multi-view user embedding
[0046] To represent a user's global and local preferences for service compositions, this invention initializes two independent embeddings, denoted as the global user embedding. and local user embedding
[0047] Steps 1-3: Embedding service composition based on MLP residual pairs Atomic service embedding Enhance MLP with multi-view user embedding;
[0048] To model implicit interaction features, service composition is embedded. and global user embedding The input is fed into an MLP network with residual connections to obtain an MLP service composition embedding. and MLP global user embedding
[0049] Embed atomic services and local user embedding It is also input into another residual network with the same structure to obtain the MLP atomic service embedding. and MLP local user embedding
[0050] The MLP network with residual connections in the global view is as follows:
[0051]
[0052] The MLP network with residual connections in the local view is as follows:
[0053]
[0054] Here, concat is a method to concatenate two embeddings, W is the trainable weight, k is the number of network propagation layers, and split is a method to separate two embeddings.
[0055] Step 2: Construct a graph structure propagation model and a bipartite graph;
[0056] The graph-based hierarchical hybrid enhancement employs a hierarchical mixup strategy. The graph structure propagation model comprises two modules: a post-graph hybrid enhancement module and a cross-view hybrid enhancement module. To model structured information, a bipartite graph is constructed in each view using existing users, service compositions, service sets, and known interaction relationships. This bipartite graph is then input into the post-graph hybrid enhancement module. The module first uses a LightGCN network to propagate the MLP-enhanced global user embeddings, local user embeddings, atomic service embeddings, and service composition embeddings into the graph structure, aggregating the interaction information of neighboring nodes to model explicit structural features. The original embeddings (service composition embeddings, atomic service embeddings, global user embeddings, and local user embeddings) without MLP enhancement undergo the same modeling process to model multi-layered user preference information.
[0057] In the global view, construct a user-service composition bipartite graph for the user and service composition. To model the explicit interactions between them, where It is a collection of users. It is a set of service compositions, ε uc It is the interaction edge of the user and service combination.
[0058] In the local view, construct a user-atomic service bipartite graph. in Denotes the set of atomic services, ε ua Interaction edges that serve users and atoms.
[0059] The graph structure propagation model includes a graph post-blending enhancement module and a cross-view blending enhancement module.
[0060] The post-graph hybrid enhancement module includes a LightGCN network and a Mixup unit, which is used to execute the Mixup method.
[0061] Step 3: Combine the bipartite graph with the original embedding and the MLP-enhanced embedding input graphs in the LightGCN network of the post-mixing enhancement module to generate graph-enhanced original embeddings and graph-enhanced MLP embeddings, respectively.
[0062] Step 3-1: Input the original embedding and the MLP-enhanced embedding together with the bipartite graph into the graph post-hybrid enhancement module to generate the graph-enhanced original embedding and the graph-enhanced MLP embedding, respectively.
[0063] This invention employs dual information propagation for the embedding of input.
[0064] In the global view:
[0065] like Figure 2 As shown, embed the global user Service composition embedding The user-service composite bipartite graph is input into LightGCN to obtain the original graph augmentation embedding; the MLP global user embedding is then applied. MLP service composition embedding The user-service composite bipartite graph is input into the LightGCN network to obtain the graph-augmented MLP embedding. The entire process is represented as follows:
[0066]
[0067] Where L is the number of layers in the network propagation. These represent the original user embedding and service composition embedding of graph enhancement under the global view, namely, the global user embedding and the service composition embedding of graph enhancement.
[0068] These represent the graph-enhanced MLP user embedding and service composition embedding under the global view, namely the graph-enhanced MLP global user embedding and the graph-enhanced MLP service composition embedding.
[0069] In partial view:
[0070] Embed local users Atomic service embedding The user-atomic service bipartite graph is input into LightGCN to obtain the original graph augmentation embedding; the local user embedding of MLP is then applied. MLP atomic service embedding The user-atomic service bipartite graph is input into the LightGCN network to obtain the graph-augmented MLP embedding. The entire process is represented as follows:
[0071]
[0072] Where L is the number of layers in the network propagation. These represent the original user embedding and atomic service embedding of graph augmentation under the local view, namely, the local user embedding and the atomic service embedding of graph augmentation.
[0073] These represent the graph-enhanced MLP user embedding and atomic service embedding under the local view, respectively, namely the graph-enhanced MLP local user embedding and the graph-enhanced MLP atomic service embedding.
[0074] Step 3-2: Input the graph augmentation original embedding and the graph augmentation MLP embedding into the Mixup unit of the graph post-hybridization module for interpolation to generate the interpolated embedding;
[0075] To integrate the diverse representations of dominant and latent features, and to alleviate the oversmoothing of single-graph propagation, this invention interpolates the two types of representations separately, as shown below:
[0076]
[0077] Where λ1 is an adjustable hyperparameter for Mixup. and These represent the user embedding and service composition embedding representations under the global view, namely the final global user embedding and the final global service composition embedding.
[0078] Using λ2 as the interpolation coefficient for the graph post-hybridization enhancement module, the user embedding in the local view (final local user embedding) is obtained. And atomic service embedding (final atomic service embedding) The specific formula is as follows:
[0079]
[0080] To obtain the service composition embedding representation under the local view, a construction was made. Used to obtain Where ε ca The dependency relationships between service compositions and atomic services are defined. An average pooling method is used to aggregate the service composition embedding representation under the local view.
[0081]
[0082] in This indicates the number of atomic services in a service composition. This indicates the embedding of the service composition under the final partial view.
[0083] Step 4: Perform cross-view blending enhancements:
[0084] In practical applications, user preferences for service compositions depend not only on the overall value of the composition but also on the individual atomic services within it. To comprehensively capture user preference information from different perspectives, this invention performs cross-view interpolation fusion of user embeddings between the global and local views, integrating overall composition preferences with atomic service preferences. This strategy effectively alleviates the sparsity of service composition interaction data, enhances the expressive power of user preferences, and thus obtains more accurate and richer representations of user interests, contributing to improved service composition recommendation performance.
[0085] Between the global view and the local view, the view blending enhancement module allows for user embedding. and Cross-view interpolation is performed to obtain the final user embedding representation, as shown in the following formula:
[0086]
[0087] Step 5: Calculate the recommendation score:
[0088] The similarity between the end-user embedding and the service composition embeddings under the global view and local view is calculated to obtain the predicted score for each candidate service composition. The calculation formula is as follows:
[0089]
[0090] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without improvement or equivalent substitution, are all within the protection scope of the present invention.
Claims
1. A graph-based hierarchical hybrid enhancement service composition recommendation method, characterized in that, Includes the following steps: Step 1: Initialize the original embedding; perform MLP enhancement on the original embedding based on the MLP residual to generate the MLP-enhanced embedding; Step 2: Construct a graph structure propagation model and a bipartite graph; The graph structure propagation model includes a graph post-hybrid enhancement module and a cross-view hybrid enhancement module; Step 3: Input the bipartite graph, the original embedding, and the MLP-enhanced embedding into the LightGCN network of the graph post-hybrid enhancement module to generate the graph-enhanced original embedding and the graph-enhanced MLP embedding, respectively. The graph-enhanced original embedding and the graph-enhanced MLP embedding are interpolated in the Mixup unit of the graph-enhanced post-graph mixing enhancement module to generate an interpolated embedding; The interpolation embedding includes the final global user embedding and the final local user embedding; Step 4: Input the final global user embedding and the final local user embedding into the cross-view blending enhancement module for cross-view interpolation to obtain the final user embedding.
2. The graph-based hierarchical hybrid enhancement service composition recommendation method according to claim 1, characterized in that, In step 1, the original embedding includes the initialized service composition embedding, atomic service embedding, global user embedding, and local user embedding; MLP enhancement is performed on the original embedding to generate MLP service composition embedding, MLP atomic service embedding, MLP global user embedding, and MLP local user embedding, respectively.
3. The graph-based hierarchical hybrid enhancement service composition recommendation method according to claim 2, characterized in that, In step 2, the bipartite graph includes a user-service composite bipartite graph and a user-atomic service bipartite graph.
4. The graph-based hierarchical hybrid enhancement service composition recommendation method according to claim 3, characterized in that, Step 3 includes step 3-1: The global user embedding, the service composition embedding, and the user-service composition bipartite graph are input into the LightGCN network to obtain graph-enhanced global user embedding and graph-enhanced service composition embedding. The MLP global user embedding, the MLP service composition embedding, and the user-service composition bipartite graph are input into the LightGCN network to obtain the graph-augmented MLP global user embedding and the graph-augmented MLP service composition embedding. The local user embedding, the atomic service embedding, and the user-atomic service bipartite graph are input into the LightGCN network to obtain graph-enhanced local user embedding and graph-enhanced atomic service embedding. The MLP local user embedding, the MLP atomic service embedding, and the user-atomic service bipartite graph are input into the LightGCN network to obtain the graph-enhanced MLP local user embedding and the graph-enhanced MLP atomic service embedding.
5. The graph-based hierarchical hybrid enhancement service composition recommendation method according to claim 4, characterized in that, Step 3 includes step 3-2: The graph-enhanced MLP global user embedding and the graph-enhanced global user embedding are input into the Mixup unit and interpolated to generate the final global user embedding. The graph-enhanced MLP local user embedding and the graph-enhanced local user embedding are input into the Mixup unit and interpolated to generate the final local user embedding.
6. The graph-based hierarchical hybrid enhancement service composition recommendation method according to claim 4, characterized in that, Step 3-2 further includes: The interpolation embedding also includes final atomic service embedding and final global service composition embedding; The graph-enhanced MLP service composition embedding and the graph-enhanced service composition embedding are input into the Mixup unit for interpolation to generate the final global service composition embedding. The graph-enhanced MLP atomic service embedding and the graph-enhanced atomic service embedding are input into the Mixup unit for interpolation to generate the final atomic service embedding; the final atomic service embedding is aggregated by average pooling to obtain the final local service composite embedding.
7. The graph-based hierarchical hybrid enhancement service composition recommendation method according to claim 6, characterized in that, The method also includes step 5, which calculates the similarity between the end-user embedding and the global service composition embedding and the local service composition embedding, respectively, to obtain the prediction score for each candidate service composition. The calculation formula is as follows: in, The end-user embedding, Embedded for the final global service composition, Embedded for the final service combination.