Association binding recommendation method based on heterogeneous graph and auto-encoder

Abstract

The invention discloses an association binding recommendation method based on a heterogeneous graph and an auto-encoder. The association binding recommendation method comprises the following steps: collecting description data in an existing Mashup application, a service binding package Bundle and a related service Service; selecting text description from the description data to perform text embedding operation; a calling relation and a service cooperation relation are selected from the description data, and a Mashup-Bundle-Service heterogeneous graph is constructed; constructing a graph propagation module, and extracting the potential implicit correlation of sparse interaction in the heterogeneous graph; a prediction module is constructed, a space attention layer is constructed, and sparse and discrete distribution of different subgraphs is fused; a self-encoding layer is introduced into the prediction module, service binding features are inscribed from the perspective of services, and semantic features of different services are adaptively aggregated; and fusing multi-model training, and integrating different types of feature descriptions. According to the method, the recall rate and the accuracy of the recommended service binding package in the Mashup development process can be greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of service bundle recommendation, and in particular relates to an associated bundle recommendation method based on a heterogeneous graph and an autoencoder. Background Art

[0002] In recent years, with the emergence of new software technologies such as cloud computing, mobile computing, and blockchain, more and more companies and organizations have begun to adjust their development strategies, publish data, resources, or related services to the Internet in the form of Web services, and realize their value propositions through Web APIs while improving the utilization rate of information and their own competitiveness. However, most traditional services follow the Simple Object Access Protocol, usually providing single-functional services for the business needs of a specific field. In addition, there are also problems such as complex technical systems and poor scalability, and it is gradually difficult to adapt to the complex and changeable application scenarios in real life.

[0003] To overcome the problems brought by traditional services, composite services based on the Mashup technology have emerged as the times require and are favored by a large number of enterprises and developers. Mashup services can integrate single-functional services and various data resources, allowing users to integrate existing Web service resources and create composite Web applications to respond to complex business needs. To complete the development of Mashup, developers usually select appropriate and compatible Web services based on the requirements of Mashup. When selecting Web services, solving the compatibility and efficiency problems between various service combinations has become an important challenge in developing Mashup. And service bundles play an important role in supporting the integration of multiple services. Service bundles mainly focus on the constraints and dependencies between services and can capture the compatibility between each Web service. Such service bundles that can consider the compatibility between Web services have a significant impact on Mashup development, thus giving rise to the corresponding field of service bundle recommendation, which provides more effective bundle recommendations through the historical interactions of Mashup, Bundle, and Web services.

[0004] In recent years, researchers have increasingly tended to use graph neural networks (GNNs) to capture information between Mashup-Service-Bundles. By using the historical calls between Mashups and Services, Bundles, and the compositional relationships between Bundles and Services, an interaction graph regarding Mashups, Bundles, and Services is constructed. For example, Dawar modeled the recommendation problem as a graph search problem and provided different service combinations by finding the smallest Steiner tree group in the interaction graph. Csbr proposed a service bundle recommendation method based on compositional semantics to mine the composite semantics of bundles in a semantic service bundle repository and recommend service packages that can cover the functional requirements of Mashups as completely as possible. These GNN-based methods have achieved remarkable success and demonstrated state-of-the-art performance. However, the existing technologies still have the following problems:

[0005] 1) First, the semantics of a service bundle are not just the aggregation of the semantics of the individual services it contains. The semantics generated by a group of collaborative services may be significantly different from their original semantics. Previous methods only simply aggregated the semantics of individual services and could not well express the semantic information of service bundles, ultimately resulting in poor quality of the learned service bundle representations.

[0006] 2) Second, existing methods cannot ensure the compatibility between services in a service bundle. In most current Mashup developments, only individual component services are recommended separately, ignoring the constraint relationships between services.

[0007] 3) Compared with the recommendation of individual services, the recommendation of service bundles has sparser interactions in historical data, resulting in a lack of interaction information of service bundles obtained in collaborative filtering-based recommendations. These limitations may lead to poor representations of Mashups, Bundles, and Services. Summary of the Invention

[0008] The present invention provides a correlation bundle recommendation method based on a heterogeneous graph and an autoencoder, which can greatly improve the recall rate and accuracy of recommending service bundles during the Mashup development process.

[0009] A correlation bundle recommendation method based on a heterogeneous graph and an autoencoder includes the following steps:

[0010] (1) Collect the description data in existing Mashup applications, service bundles Bundle, and related services Service;

[0011] (2) Select the text descriptions of Mashup applications, service bundles (Bundles), and related services (Services) from the description data for text embedding operations to obtain text feature embeddings;

[0012] (3) Select the call relationships between Mashup applications and service bundles (Bundles), the call relationships between Mashup applications and services (Services), and the collaboration relationships between service bundles (Bundles) and services (Services) from the description data to construct a Mashup-Bundle-Service heterogeneous graph, including a Mashup-Bundle subgraph, a Mashup-Service subgraph, and a Bundle-Service subgraph;

[0013] (4) Based on the Mashup-Bundle-Service heterogeneous graph and the text feature embeddings, construct a graph propagation module to extract the potential implicit correlations with sparse interactions in the heterogeneous graph, and obtain the representations of Mashup, Bundle, and Service under each subgraph;

[0014] (5) Construct a spatial attention layer to aggregate the Service representations under multiple subgraphs to obtain the aggregated Service feature E s* , after average pooling and max pooling operations, input it into the spatial attention machine layer to obtain the aggregated Service feature U SA (E s* );

[0015] (6) Construct an autoencoder layer to encode the semantics of the service bundle from the perspective of the service. According to the aggregated Service feature U SA (E s* ) obtained after spatial attention calculation, obtain the reconstructed Bundle feature aggregated from the service features

[0016] (7) Aggregate multiple Bundle features to obtain the aggregated Bundle feature E b* , after average pooling and max pooling operations, input it into the spatial attention layer to obtain the aggregated Bundle feature U SA (E b* );

[0017] (8) Adopt a multi-model fusion method to aggregate the features obtained in steps (4) to (7), and then pass the fused features through an activation function to obtain the prediction score of Mashup for Bundle.

[0018] In step (1), the described description data includes the Mashup application name, the text description of the Mashup application, the Mashup application category, the service name, the text description of the service, the service category, the service provider, the text description of the service bundle, the call relationship between the Mashup application and the service, the call relationship between the Mashup application and the service bundle, and the collaboration relationship between the service bundle and the service.

[0019] In step (3), during the construction process, the Mashup application-related encoding, the service bundle-related encoding, and the service-related encoding are used as nodes. When there is a call relationship between any two different types of nodes, an edge is established on the heterogeneous graph, thereby obtaining the Mashup-Bundle-Service heterogeneous graph.

[0020] In step (4), the described graph propagation module performs graph convolution operations on the three input subgraphs respectively to extract structural representations, and obtains the feature representations of Mashup, Bundle, and Service under each subgraph; specifically including: the representation e m of Mashup and the representation e b of Bundle in the Mashup-Bundle subgraph, the representation of Mashup and the representation of Service in the Mashup-Service subgraph, the representation of Bundle and the representation of Service in the Bundle-Service subgraph. At the same time, the representation of Service in the Bundle-Service subgraph is further aggregated to obtain the representation

[0021] The described graph propagation module adopts an improved LightGCN, and removes the feature transformation and non-linear activation operations in GCN from LightGCN to form an improved LightGCN. <00>

[0022] The specific process of step (5) is as follows:

[0023] First, the representations of Service in the Mashup-Service subgraph and the representation of Service in the Bundle-Service subgraph are aggregated through average pooling and max pooling operations to generate two two-dimensional graphs

[0024]

[0025] In the spatial attention mechanism, first, two-dimensional maps obtained from average pooling and max pooling operations are concatenated to describe features; then, a convolutional layer is applied to generate a spatial attention map:

[0026]

[0027] where σ represents the sigmoid activation function, and conv 7×7 represents a 7×7 convolutional kernel; E s* represents the representation of Service in the Mashup-Service subgraph and the representation of Service in the Bundle-Service subgraph are concatenated to form the aggregated Service feature;

[0028] Finally, the spatial attention map M s (E) is combined with the adaptive pooling representation of the subgraph feature to obtain the aggregated Service feature U SA (E s* ) after spatial attention calculation.

[0029] In step (6), the aggregated Service feature U SA (E s* ) obtained after spatial attention calculation is input into the autoencoder to obtain the reconstructed Bundle feature aggregated from the service features The specific formula is as follows:

[0030]

[0031] x = U SA (E s* )

[0032] where W AE and b AE are the weight matrix and bias vector in the autoencoder respectively, and f θ (x) represents the encoder in the autoencoder, and represents the decoder in the autoencoder.

[0033] The specific process of step (7) is as follows:

[0034] First, multiple Bundle features are aggregated through average pooling and max pooling operations, and the formula is as follows:

[0035]

[0036] In the formula, and respectively represent the two-dimensional graphs obtained by performing average pooling and max pooling operations on the aggregated Bundle feature E b* ; E b* represents e b connected to form the aggregated Bundle feature;

[0037] In the spatial attention mechanism, first, the two-dimensional graphs and obtained from the average pooling and max pooling operations are connected to describe the features; then, a convolutional layer is applied to generate the spatial attention map:

[0038]

[0039] where σ represents the sigmoid activation function, and conv 7×7 represents a 7×7 convolutional kernel;

[0040] Finally, the spatial attention map M b (E b* ) is combined with the adaptive pooling representation of the subgraph features to obtain the aggregated Bundle feature U SA (E b* ).

[0041] In step (8), a multi-model fusion method is used to aggregate the features obtained in steps (4) to (7), as shown below:

[0042]

[0043] In the formula, e m represents the representation of Mashup in the Mashup-Bundle subgraph; represents the representation of Bundle in the Bundle-Service subgraph; represents the representation of Mashup in the Mashup-Service subgraph; U SA (E b* ) represents the aggregated Bundle feature after spatial attention calculation.

[0044] The fused features are passed through the activation function, and the formula is as follows:

[0045]

[0046] In the formula, W TL and b TL are the weight matrix and bias vector respectively. Compared with the prior art, the present invention has the following beneficial effects:

[0047] Compared to existing methods, this invention effectively addresses three issues in the service bundle recommendation task. It uses an autoencoder to map service features to service bundles, addressing the misalignment of textual semantics and constructing a more complete service bundle profile. It also employs a spatial attention mechanism to address inter-service compatibility issues. It also refines service feature representation and employs multiple views to obtain node representations to address the sparsity of historical interaction data. This method characterizes mashups, service bundles, and web services from multiple perspectives. Ultimately, it significantly improves the recall and accuracy of service bundle recommendations during mashup development. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of the association bundling recommendation method based on heterogeneous graphs and autoencoders of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0050] In the service ecosystem, the embodiment of the present invention uses B={b1,b2,...,b |B|}、S={s1,s2,...,s |S|} and M={m1,m2,...,m |M|} respectively represent the set of service bundles, web services and mashups, where |B|, |S|, |M| respectively represent the number of service bundles, web services and mashups. For each service bundle b, there are b i ={s1,s2,...,s n}, represents the services included in the service bundle, where n represents the number of web services included in the service bundle. Each new Mashup is associated with a set of sparse attributes, denoted as m i ={D mi ,T mi ,C mi ,B mi}. Among them, D mi It is a text description used to introduce functional information. mi is a set of labels indicating the category C to which it belongs mi , and B mi is a set of service packages that indicate their participation in hybrid applications. Similarly, the sparse attribute set of each service in a service package is defined as s i ={D si ,T si ,C si}, where Dsi is a text description for introducing functional information, T si is a set of tags used to indicate the category C to which it belongs si . Given three sets B, S, M, and a sequence of keyword DW = {q1, q2,..., q N}, where q i represents a keyword and N is the total number of keywords. The goal of service bundle recommendation is to build a model that can generate different service packages and recommend the top N service packages that have not been interacted with in history to developers as accurately as possible.

[0051] As Figure 1 shown, an association bundling recommendation method based on a heterogeneous graph and an autoencoder includes the following steps:

[0052] (1) Collect description data in existing Mashup applications, service bundles, and related services.

[0053] First, obtain the description information of Mashup applications, service bundles, and services from the Programmable Web service online registration platform. The description information includes Mashup application name, Mashup application text description information, Mashup application category, service name, service text description, service category, service provider, text description of service bundles, call relationships between Mashup applications and services, call relationships between Mashup applications and service bundles, collaboration relationships between service bundles and services, etc. The text description information mainly focuses on introducing the target Mashup and the provided functional information, and the service bundle is a set composed of several functionally compatible services.

[0054] (2) Build an embedding module to perform text embedding operations on the text description information and attributes of Mashup applications, service bundles, and services.

[0055] Based on the text description information obtained in step (1), in this embodiment, a data preprocessing method is used to parse the texts of Mashup applications, service bundles, and candidate services, establish vocabulary tables respectively, vectorize the description text sets by word indexing, and obtain a text set vector matrix represented by word indexing. The preprocessing method mainly includes: text filtering, abbreviation replacement, part-of-speech restoration, and word vectorization.

[0056] In the text filtering method, regular expressions are used to filter invalid words such as tags, punctuation marks, non-characters, stop words (such as am, is, etc.).

[0057] In the abbreviation replacement method, the complete word spelling is used to replace the abbreviated word to normalize the text description (e.g., cannot / can’t).

[0058] In the part-of-speech reduction method, the StandardAnalyzer and PosterStemFilter tools in the Lucene package are used to convert all terms into a lexical format without prefixes and suffixes.

[0059] In the word vectorization method, a vocabulary of existing words is constructed, and the word vectors of the Glove.6B.300d pre-trained model are used to initialize the embeddings of the existing words, and the vectors of the cold-start words are randomly initialized. All words in the vocabulary are independent and have a one-to-one index. Cold-start words refer to words not covered in the existing vocabulary.

[0060] For the text description inputs of Mashup applications, service bundles, and candidate services, their text feature embeddings are obtained respectively through the mapping of the above text set vector matrix.

[0061] (3) According to the call relationships between Mashup applications and service bundles, the call relationships between Mashup applications and services, and the collaboration relationships between service bundles and services, construct a Mashup-Bundle-Service heterogeneous graph.

[0062] Based on the call relationships between Mashup applications and service bundles, the call relationships between Mashup applications and services, and the collaboration relationships between service bundles and services obtained in step (1), construct a Mashup-Bundle-Service heterogeneous graph, and select category and provider as supplementary information for the heterogeneous graph.

[0063] The Mashup-Bundle-Service heterogeneous graph consists of a Mashup-Bundle subgraph, a Mashup-Service subgraph, and a Bundle-Service subgraph. In the embodiment, when constructing, the Mashup application-related codes, service bundle-related codes, and service-related codes are used as nodes. When there is a call relationship between any two different types of nodes, an edge is established on the heterogeneous graph to construct the Mashup-Bundle-Service heterogeneous graph.

[0064] (4) Construct a graph propagation module to extract the potential implicit correlations of sparse interactions in the heterogeneous graph.

[0065] Based on the Mashup-Bundle-Service heterogeneous graph obtained in step (3) and the text feature embedding obtained in step (2), construct a graph propagation module to extract the potential implicit correlations of sparse interactions in the heterogeneous graph, and obtain the interaction feature vectors of different subgraphs respectively.

[0066] The graph propagation module will perform graph convolution operations on the input subgraphs respectively to extract structural representations. Specifically, first, the Mashup-Bundle-Service heterogeneous graph obtained in step (3) is split into Mashups-Bundles subgraph, Mashups-Services subgraph, and Bundles-Services subgraph according to the types of nodes and edges. When performing message passing on the three views, mainly the feature transformation and non-linear activation operations in GCN are removed to form an improved LightGCN, and LightGCN is used to perform convolution operations on the input sparse subgraphs to extract structural representations. By removing the feature transformation and non-linear activation operations, the computational complexity can be reduced, the computational efficiency can be greatly improved, and the training time can be reduced. Specifically, the representations obtained through double-layer LightGCN convolution are used as structural representations. After information passing on the three subgraphs, the features of Mashup, Bundle, and Service in each view are obtained, which are formalized as:

[0067]

[0068] Among them, respectively represent the representations of Bundle and Mashup at the k-th layer, p mb is the Laplacian norm, N m and N b are the one-hop neighbors of Mashup m and Bundle b respectively. When calculating the representations of each view in the graph propagation layer, the representations of each layer need to be added according to a certain weight. α k represents the weight coefficient of each layer when aggregating the representations of each layer. Similarly, in the Mashup-Service view and the Bundle-Service view, there are:

[0069]

[0070] Among them, are the representations of Mashup and Service in the Mashup-Service subgraph respectively, are the representations of Bundle and Service in the Bundle-Service subgraph respectively.

[0071] To obtain the representation of the Bundle from multiple perspectives, in the Bundle-Service subgraph, the representations of the Services are further aggregated to obtain the representation of the Bundle, which is formalized as:

[0072]

[0073] where is the aggregated Bundle representation, is the one-hop neighbor of Bundle b, is the representation of the Service in the Bundle-Service subgraph.

[0074] (5) Construct a prediction module, construct a spatial attention layer, and fuse the sparse and discrete distributions of different subgraphs.

[0075] Based on the subgraph features obtained in step (4), first apply average pooling and max pooling operations along the subgraph features and concatenate them to generate an effective concatenated feature descriptor. On the concatenated feature descriptor, generate a spatial attention map by applying a convolutional layer.

[0076] The average pooling and max pooling operations are to aggregate the representations of the Services in the Mashup-Service subgraph and the representations of the Services in the Bundle-Service subgraph to generate two two-dimensional maps

[0077]

[0078] In the spatial attention mechanism, first concatenate the two-dimensional maps obtained from the average pooling and max pooling operations to describe the features. Generate a spatial attention map by applying a convolutional layer:

[0079]

[0080] where σ represents the sigmoid activation function, conv 7×7 represents a 7*7 convolutional kernel, and E S* represents the representation of the Service in the Mashup-Service subgraph and the representation of the Service in the Bundle-Service subgraph connected to form the aggregated Service feature.

[0081] Finally, use the inner product to combine the spatial attention with the adaptive pooling representation of the subgraph features Combine them to enrich the understanding and expression ability of the correlation between different sub - graphs, and obtain the aggregated Service feature U after spatial attention calculation SA (E s* ):

[0082]

[0083] (6) Introduce an auto - encoder layer in the prediction module to engrave service bundling features from the perspective of services and adaptively aggregate the semantic features of different services

[0084] Generally, each service has its unique features, attributes, and goals. However, when these services are combined, their semantic information is assembled and shared. Through this shared feature space and interaction, a more comprehensive and accurate semantic representation of the service package can be provided. To achieve a unified representation of service package features, we propose to introduce an auto - encoder to encode the semantics of the service bundle from the perspective of services. In the auto - encoder, the encoder network maps the features of the service to a vector space through a mapping function f θ (x):

[0085] h = f θ (x)=σ(wx + b)

[0086] where w is the weight matrix, b is the bias vector, and σ is the activation function

[0087] Then, the obtained hidden representation h will be mapped back to a reconstructed d - dimensional vector in the input space through the decoder network, that is Generally, the mapping function is also a compressed non - linear function

[0088]

[0089] where w' is the weight matrix of the decoder and b′ is the bias vector of the decoder

[0090] Input the aggregated Service feature U SA (E S* ) into the auto - encoder to obtain the reconstructed Bundle feature aggregated from service features The specific formula is as follows

[0091]

[0092] x = U SA (E S* )

[0093]

[0094] Among them, W AE and b AE are the weight matrix and bias vector in the autoencoder, respectively, f θ (x) represents the encoder in the autoencoder, Represents the decoder in the autoencoder.

[0095] (7) Aggregate multiple Bundle features, perform average pooling and maximum pooling operations, and input them into the spatial attention layer to obtain the aggregated Bundle features after spatial attention calculation.

[0096] Reconstructed Bundle features transformed and aggregated from service features After that, the multiple Bundle features are aggregated again and used as the input of the spatial attention layer to obtain the aggregated Bundle features U from multiple sources after spatial attention calculation. SA (E b* ):

[0097]

[0098] in, Represents the reconstructed Bundle feature that is transformed and aggregated from service features. and Respectively represent the aggregation Bundle feature E b* The two-dimensional image obtained by performing average pooling and maximum pooling operations, M b (·) indicates that convolution is used to calculate the spatial attention coefficient, U SA (E b* ) represents the aggregated Bundle feature U after spatial attention calculation SA (E b* ).

[0099] (8) Fusion of multi-model training and integration of different types of feature descriptions.

[0100] Based on the several feature vectors obtained in steps (4) to (7), in order to integrate different types of features and compensate for the heterogeneity of different models, multi-model fusion is used to aggregate the obtained features. The multi-model fusion formula is as follows:

[0101]

[0102]

[0103] Finally, we get the Mashup's predicted score for the Bundle:

[0104]

[0105] Among them, WTL and b TL are respectively the weight matrix and the bias vector in the linear layer of the prediction scores. Under the action of the activation function, all scores are in the interval (0, 1).

[0106] The embodiment of the present invention adopts soft margin loss to train the model parameters, which is specifically expressed as follows:

[0107]

[0108] where y m [b] represents whether the b-th Bundle meets the requirements of the Mashup, represents the predicted value of the model.

[0109] The above-described embodiments have detailed the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modification, supplement, and equivalent replacement made within the scope of the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An associated bundling recommendation method based on heterogeneous graphs and autoencoders, characterized in that, It includes the following steps: (1) Collect the description data in existing Mashup applications, service bundles (Bundles), and related services (Services); (2) Select the text descriptions of Mashup applications, service bundles (Bundles), and related services (Services) from the description data for text embedding operations to obtain text feature embeddings; (3) Select the call relationships between Mashup applications and service bundles (Bundles), the call relationships between Mashup applications and services (Services), and the collaboration relationships between service bundles (Bundles) and services (Services) from the description data to construct a Mashup-Bundle-Service heterogeneous graph, including Mashup-Bundle subgraphs, Mashup-Service subgraphs, and Bundle-Service subgraphs; (4) Based on the Mashup-Bundle-Service heterogeneous graph and the text feature embeddings, construct a graph propagation module to extract the potential implicit correlations of sparse interactions in the heterogeneous graph to obtain the representations of Mashup, Bundle, and Service under each subgraph; (5) Construct a spatial attention layer to obtain the Service representations under multiple subgraphs for aggregation, and obtain the aggregated Service feature E s* , after average pooling and max pooling operations, input it into the spatial attention machine layer to obtain the aggregated Service feature U after spatial attention calculation SA (E s* ); (6) Build an auto-encoding layer to encode the semantics of the service bundle from the perspective of the service, and based on the aggregated Service feature U after spatial attention calculation SA (E s* ), obtain the reconstructed Bundle feature transformed and aggregated from the service features (7) Aggregate multiple Bundle features to obtain the aggregated Bundle feature E b* , after average pooling and max pooling operations, input it into the spatial attention layer to obtain the aggregated Bundle feature U after spatial attention calculation SA (E b* ); (8) Adopt a multi-model fusion method to aggregate the features obtained in steps (4) to (7), and then pass the fused features through an activation function to obtain the prediction score of Mashup for Bundle.

2. The correlation bundling recommendation method based on heterogeneous graph and autoencoder according to claim 1, wherein In step (1), the description data includes Mashup application names, text descriptions of Mashup applications, Mashup application categories, service names, text descriptions of services, service categories, service providers, text descriptions of service bundles, call relationships between Mashup applications and services, call relationships between Mashup applications and service bundles, and collaboration relationships between service bundles and services.

3. The correlation bundling recommendation method based on heterogeneous graph and autoencoder according to claim 1, wherein In step (3), during the construction process, use the encodings related to Mashup applications, the encodings related to service bundles, and the encodings related to services as nodes. When there is a call relationship between any two different types of nodes, establish an edge on the heterogeneous graph to obtain the Mashup-Bundle-Service heterogeneous graph.

4. The method for associated bundling recommendation based on heterogeneous graph and auto - encoder according to claim 1, characterized in that, In step (4), the graph propagation module performs graph convolution operations on the three input subgraphs respectively to extract structural representations, and obtains the feature representations of Mashup, Bundle, and Service under each subgraph; specifically including: the representation e of Mashup in the Mashup-Bundle subgraph m and the representation e of Bundle b , the representation of Mashup in the Mashup-Service subgraph and the representation of Service the representation of Bundle in the Bundle-Service subgraph and the representation of Service At the same time, the representation of Service in the Bundle-Service subgraph is further aggregated to obtain the representation of Bundle 5. The correlation bundling recommendation method based on heterogeneous graph and autoencoder according to claim 4, wherein The graph propagation module adopts an improved LightGCN, removing the feature transformation and non-linear activation operations in GCN from LightGCN to form an improved LightGCN.

6. The correlation bundling recommendation method based on heterogeneous graph and autoencoder according to claim 4, wherein The specific process of step (5) is as follows: First, aggregate the representations of Services in the Mashup-Service subgraph through average pooling and max pooling operations and the representations of Services in the Bundle-Service subgraph to generate two two-dimensional graphs In the spatial attention mechanism, first connect the two-dimensional graphs obtained by average pooling and max pooling operations to describe the features; then generate a spatial attention map by applying a convolutional layer: Among them, σ represents the sigmoid activation function, and conv 7×7 represents a 7×7 convolution kernel; E s* represents the representation of Service in the Mashup-Service subgraph and the representation of Service in the Bundle-Service subgraph are connected to form an aggregated Service feature; Finally, use the inner product to combine the spatial attention map M s (E) with the adaptive pooling representation of the subgraph features to obtain the aggregated Service features after spatial attention calculation U SA (E s* )。 7. The correlation bundling recommendation method based on heterogeneous graph and autoencoder according to claim 6, wherein In step (6), the aggregated Service feature U after spatial attention calculation SA (E s* ) is input into the autoencoder to obtain the reconstructed Bundle feature transformed and aggregated from the service feature The specific formula is as follows: x = U SA (E s* ) Among them, W AE and b AE are the weight matrix and bias vector in the autoencoder respectively, and f θ (x) represents the encoder in the autoencoder, represents the decoder in the autoencoder.

8. The correlation bundling recommendation method based on heterogeneous graph and auto - encoder according to claim 7, wherein, The specific process of step (7) is as follows: First aggregate multiple Bundle features through average pooling and max pooling operations. The formula is as follows: In the formula, and respectively represent two-dimensional graphs obtained by performing average pooling and max pooling operations on the aggregated Bundle feature E b* ; E b* represents e b connected to form the aggregated Bundle feature; In the spatial attention mechanism, first, the two-dimensional maps obtained from the average pooling and max pooling operations are and concatenated to describe the features; Then generate a spatial attention map by applying a convolutional layer: Among them, σ represents the sigmoid activation function, and conv 7×7 represents a 7×7 convolutional kernel; Finally, use the inner product to combine the spatial attention map M b (E b* ) with the adaptive pooling representation of the subgraph features to obtain the aggregated Bundle features after spatial attention calculation U SA (E b* )。 9. The method for associated bundling recommendation based on heterogeneous graph and autoencoder according to claim 8, wherein In step (8), adopt a multi-model fusion method to aggregate the features obtained in steps (4) to (7). The formula is as follows: where, e m represents the representation of the Mashup in the Mashup-Bundle subgraph; represents the representation of the Bundle in the Bundle-Service subgraph; represents the representation of the Mashup in the Mashup-Service subgraph; U SA (E b* ) represents the aggregated Bundle feature after spatial attention calculation.

10. The correlation bundling recommendation method based on heterogeneous graph and autoencoder according to claim 9, characterized in that, In step (8), pass the fused features through an activation function. The formula is as follows: where, W TL and b TL are the weight matrix and the bias vector respectively.

Images