QoS prediction method and device

Abstract

The invention discloses a QoS prediction method and device, and is applied to the technical field of communication, and the method comprises the steps: obtaining user entity data and service entity data; performing embedding learning on the user entity data and the service entity data to obtain a user feature embedding matrix and a service feature embedding matrix; performing feature interaction on the user feature embedding matrix and the service feature embedding matrix to obtain interactive fusion features; and performing QoS prediction according to the interactive fusion features to obtain a service quality result of the service entity relative to the user entity. According to the method, the dynamic relationship between the user entity and the service entity can be captured without depending on a graph structure through embedded learning, and the attention of the user entity to different service entities and the attention of the service entity to different user entities can be accurately captured through feature interaction; and the nonlinear dynamic interaction relationship between the user entity and the service entity is effectively expressed, and prediction is realized on the basis, so that the QoS prediction precision can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a QoS prediction method and apparatus. Background Technology

[0002] Quality of Service (QoS), a crucial non-functional attribute in the Web services domain, encompasses core performance metrics such as response time and throughput. QoS prediction aims to estimate service call performance by uncovering the underlying interaction relationships between users and services. Related technologies employ methods such as Collaborative Filtering (CF), latent factor models, and Graph Neural Networks (GNNs) to capture feature representations and achieve QoS prediction. However, these technologies struggle to achieve accurate and stable QoS prediction in the absence of explicit graph structures and under conditions of sparse user-service interactions. Summary of the Invention

[0003] This application provides a QoS prediction method and apparatus to improve QoS prediction accuracy.

[0004] On the one hand, embodiments of this application provide a QoS prediction method, the method including: Obtain user entity data and service entity data; Embedding learning is performed on the user entity data and the service entity data to obtain the user feature embedding matrix and the service feature embedding matrix; The user feature embedding matrix and the service feature embedding matrix are subjected to feature interaction to obtain interactive fusion features; Based on the interaction fusion features, QoS prediction is performed to obtain the service quality result of the service entity relative to the user entity.

[0005] Further, in one embodiment, the embedding learning of the user entity data and the service entity data to obtain the user feature embedding matrix and the service feature embedding matrix includes: Assign corresponding indexes to each data item in the target entity data; The indexes of each data item in the target entity data are embedded into a continuous vector space to obtain the transformation matrix of each data item in the target entity data; The transformation matrix of each data item in the target entity data is optimized to obtain the embedding matrix of each data item in the target entity data; The embedding matrices of each data item in the target entity data are integrated and processed to obtain the target feature embedding matrix; Wherein, the target entity data is either the user entity data or the service entity data. When the target entity data is the user entity data, the target feature embedding matrix is ​​the user feature embedding matrix. When the target entity data is the service entity data, the target feature embedding matrix is ​​the service feature embedding matrix.

[0006] Further, in one embodiment, optimizing the transformation matrix of each data item in the target entity data to obtain the embedding matrix of each data item in the target entity data includes: Noise prediction is performed on the transformation matrix of each data item in the target entity data to obtain the noise prediction component of each data item in the target entity data; Using the noise prediction components of each data item in the target entity data, the transformation matrix of each data item in the target entity data is subjected to denoising diffusion reconstruction processing to obtain the embedding matrix of each data item in the target entity data.

[0007] Furthermore, in one embodiment, the process of integrating the embedding matrices of each data item in the target entity data to obtain the target feature embedding matrix includes: The embedding matrix of each data item in the target entity data is averaged to obtain the mean embedding matrix of the target entity data. The mean embedding matrix of the target entity data is normalized to obtain the target feature embedding matrix.

[0008] Further, in one embodiment, the step of performing feature interaction between the user feature embedding matrix and the service feature embedding matrix to obtain interactive fusion features includes: The user feature embedding matrix and the service feature embedding matrix are processed as inputs to obtain a first input feature and a second input feature; wherein the first input feature and the second input feature are features obtained through different input processes; Multi-head attention processing is performed on the first input feature and the second input feature to obtain a first attention feature and a second attention feature; wherein the first attention feature and the second attention feature are features obtained through different queries, keys and values; The first attention feature and the second attention feature are fused together to obtain the interaction fusion feature.

[0009] Further, in one embodiment, the input processing of the user feature embedding matrix and the service feature embedding matrix to obtain the first input feature and the second input feature includes: The user feature embedding matrix and the service feature embedding matrix are fused to obtain joint features; The joint features are processed based on the first bias and the first weight to obtain the first input features; The joint features are processed based on the second bias and the second weight to obtain the second input features; Wherein, the first bias and the second bias are biases with different values, and the first weight and the second weight are weights with different values.

[0010] Further, in one embodiment, the step of performing multi-head attention processing on the first input feature and the second input feature to obtain the first attention feature and the second attention feature includes: Using the first input feature as the query and the second input feature as the key and value, a multi-head attention operation is performed on the first input feature and the second input feature to obtain the first attention feature.

[0011] Further, in one embodiment, the step of performing multi-head attention processing on the first input feature and the second input feature to obtain the first attention feature and the second attention feature includes: Using the second input feature as the query and the first input feature as the key and value, a multi-head attention operation is performed on the first input feature and the second input feature to obtain the second attention feature.

[0012] Further, in one embodiment, the step of performing QoS prediction based on the interaction fusion features to obtain the service quality result of the service entity relative to the user entity includes: The interaction fusion features are nonlinearly mapped and predicted using a multilayer perceptron to obtain the service quality result of the service entity relative to the user entity.

[0013] On the other hand, embodiments of this application provide a QoS prediction apparatus, including: The acquisition module is used to acquire user entity data and service entity data. The first processing module is used to perform embedding learning on the user entity data and the service entity data to obtain a user feature embedding matrix and a service feature embedding matrix. The second processing module is used to perform feature interaction between the user feature embedding matrix and the service feature embedding matrix to obtain interactive fusion features. The third processing module is used to perform QoS prediction based on the interaction fusion features to obtain the service quality result of the service entity relative to the user entity.

[0014] According to an embodiment of this application, a QoS prediction method and apparatus are provided, comprising: acquiring user entity data and service entity data; performing embedding learning on the user entity data and service entity data to obtain a user feature embedding matrix and a service feature embedding matrix; performing feature interaction between the user feature embedding matrix and the service feature embedding matrix to obtain interactive fusion features; and performing QoS prediction based on the interactive fusion features to obtain the service quality result of the service entity relative to the user entity. This application, through embedding learning, can capture the dynamic relationship between user entities and service entities without relying on graph structures, while through feature interaction, it can accurately capture the attention of user entities to different service entities and the attention of service entities to different user entities, and effectively express the nonlinear dynamic interaction relationship between user entities and service entities. Based on this, prediction can be performed, effectively improving the accuracy of QoS prediction.

[0015] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0016] Figure 1 This is a flowchart of a QoS prediction method provided in this application; Figure 2 This is a schematic diagram of the embedded learning principle provided in this application; Figure 3 This is a schematic diagram of the feature interaction provided in this application; Figure 4 This is a performance comparison chart of existing embedded learning methods and the embedded learning method provided in this application in terms of response time metrics; Figure 5 This is a performance comparison chart of existing embedded learning methods and the embedded learning method provided in this application in terms of throughput metrics; Figure 6 This is a performance comparison chart of existing feature interaction methods and the feature interaction method of this application in terms of response time. Figure 7 This is a performance comparison chart of existing feature interaction methods and the feature interaction method of this application in terms of throughput. Figure 8 This is an experimental comparison graph of hyperparameter embedding dimensions; Figure 9 This is an experimental comparison chart of the number of hyperparameter attention heads; Figure 10 This is a structural diagram of a QoS prediction device provided in this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0019] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] QoS, a crucial non-functional attribute in the Web services domain, encompasses core performance metrics such as response time and throughput. QoS prediction aims to estimate service call performance by uncovering the underlying interaction relationships between users and services. With the rapid development of cloud computing and IoT technologies, Web services offering similar functions but varying QoS attributes are experiencing explosive growth. Accurate QoS prediction methods are essential for supporting a broader service ecosystem; their prediction results have decisive significance for key downstream applications and can directly serve core aspects such as service selection, service composition, and service recommendation. The lack of accurate QoS prediction methods can easily lead to the selection of suboptimal services, significantly degrading user experience and overall system performance. Therefore, establishing efficient and reliable QoS prediction methods has become a core research hotspot.

[0022] Early QoS prediction methods primarily relied on correlation-based strategies, especially CF (Consciousness-Based Flow), whose core idea was to estimate the similarity of users or services using methods such as the Pearson correlation coefficient, thereby achieving neighborhood-based prediction. Although such methods are intuitive and highly interpretable, their scalability and predictive performance exhibit significant limitations when faced with sparse and incomplete data.

[0023] To overcome the limitations of collaborative filtering methods, subsequent research has gradually shifted towards latent factor models to capture the implicit interactions between users and services. For example, neighborhood latent factor models combine latent relationships between service providers with geographic information to improve QoS prediction performance; some studies further introduce contextualized latent features in cloud computing environments, significantly improving prediction accuracy through personalized interaction modeling; other studies simultaneously model user-service and environment-environment interactions, highlighting the crucial role of contextual information in the prediction process and thus significantly improving prediction accuracy; still others introduce methods such as Adaptive Matrix Factorization (AMF) to further enhance prediction capabilities by mining latent relationships. Although latent factor models effectively improve prediction accuracy, they still face challenges in terms of computational complexity and adaptability to dynamic environments.

[0024] To enhance the model's ability to characterize the interactions between users and services, some studies have proposed Neural Collaborative Filtering (NCF), which replaces the inner product calculation method in traditional collaborative filtering with a multilayer perceptron structure, thereby uncovering higher-order association features between users and services more deeply. Other studies have further improved predictive performance by introducing multi-stage, multi-index fusion strategies, emphasizing the importance of comprehensive feature integration in improving predictive reliability. However, these methods generally fail to fully capture the unique features of Web service scenarios. The fundamental reason lies in insufficient mining of the underlying graph topology of service interactions, leading to poor performance in cold-start scenarios, especially when the density of the user-service matrix is ​​extremely low. Furthermore, these methods often fail to effectively handle contextual features, typically requiring manual analysis to process them. Moreover, contextual features must be hard-coded and embedded into the optimization objective, limiting the model's adaptability and generalization ability.

[0025] In recent years, graph neural networks (Graph Neural Networks), as a deep learning (DL) model based on graph structures, have been widely applied in fields such as social network analysis and sequence recommendation. These models can effectively capture high-order dependencies between nodes and edges, thus exhibiting superior performance in complex relationship modeling and pattern recognition. Existing research has shown that applying Graph Neural Networks to QoS prediction tasks can significantly improve prediction accuracy. For example, related research has used Graph Neural Networks to construct user and service interaction graphs, achieving more accurate QoS value predictions. However, this approach generally relies on predefined static graph structures and is difficult to adapt to dynamically changing service environments. When new users or service nodes join the system, the entire graph structure often needs to be reconstructed and retrained, which limits the model's scalability and real-time performance.

[0026] Based on the above, the relevant technologies have the following problems: (1) Lack of explicit graph structure: Existing graph neural network models generally rely on predefined graph structures (i.e., user-service bipartite graphs) to achieve feature propagation and relationship modeling. However, on the one hand, with the expansion of the service ecosystem, the construction and inference of graph structures face scalability challenges such as high computational complexity and large storage overhead. On the other hand, the interaction records between users and services are often sparse, and there is no stable topological relationship between users and services, which makes it difficult to construct an effective graph structure. Under the condition of no graph structure (i.e., lack of explicit topological connections), graph neural network models cannot directly model the potential relationships between users and services, and their modeling ability is limited, making it difficult to achieve high-precision QoS prediction.

[0027] (2) Insufficient modeling ability under sparse interaction: Under the condition of high data sparsity, the number of historical interactions between users and services is limited. Collaborative filtering and latent factorization methods are highly sensitive to data sparsity and are difficult to characterize the complex nonlinear high-order interaction relationship between users and services. This makes it difficult for the model to capture interaction patterns and preference features. Especially in cold start or low-density matrix scenarios, information loss and overfitting problems are likely to occur, which seriously affect the stability and generalization performance of QoS prediction.

[0028] In summary, existing technologies struggle to achieve accurate and stable QoS prediction in the absence of explicit graph structures and under conditions of sparse user-service interactions. Therefore, there is an urgent need for a prediction method that can still exhibit strong generalization and robustness even without a graph structure, in order to improve adaptability and reliability in real-world, complex service environments.

[0029] In view of this, embodiments of this application provide a QoS prediction method and apparatus, which aim to effectively improve the accuracy of QoS prediction.

[0030] The QoS prediction method provided in this application can be applied to terminals, servers, or software running on either terminal or server. Terminals can be tablets, laptops, desktop computers, etc., but are not limited to these. Servers can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Furthermore, a server can be a node server in a blockchain network, but is not limited to this. Blockchain is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0031] Reference Figure 1 , Figure 1 This is a flowchart of a QoS prediction method provided in this application, which may include the following steps S101-S104.

[0032] S101, Obtain user entity data and service entity data.

[0033] It should be noted that user entity data refers to the contextual data of a user entity, which may include, but is not limited to, the user entity's Autonomous System (AS) information and its region information. Furthermore, service entity data refers to the contextual data of a service entity, which may include, but is not limited to, the service entity's AS information, deployment location information, and provider information. For ease of understanding, user entity data can be represented as... , This indicates the amount of context data for the user entity. The first character representing the user entity Context data; service entity data can be represented as , This indicates the amount of context data for the service entity. The first one represents the service entity Context data, .

[0034] It is understandable that a service entity refers to a web service; a user entity refers to an object associated with a user, which may be an account, but is not limited to this.

[0035] In this step, firstly, the service entity and the user entity are identified; then, the context data of the service entity, i.e., the service entity data, is obtained; at the same time, the context data of the user entity, i.e., the user entity data, is obtained.

[0036] S102, perform embedding learning on user entity data and service entity data to obtain user feature embedding matrix and service feature embedding matrix.

[0037] In this step, embedding learning is performed on user entity data to obtain a user feature embedding matrix; simultaneously, embedding learning is performed on service entity data to obtain a service feature embedding matrix. Embedding learning includes embedding processing and denoising diffusion processing. Embedding processing aims to capture the latent features of user entities and service entities, generating representations that can be used for prediction. Denoising diffusion processing aims to capture the dynamic relationships between user entities and service entities without relying on graph structure, which helps provide accurate feature representations for subsequent prediction processing.

[0038] S103, the user feature embedding matrix and the service feature embedding matrix are interacted to obtain interactive fusion features.

[0039] In this step, the user feature embedding matrix and the service feature embedding matrix are interacted to obtain interactive fusion features. This allows for a more flexible and accurate capture of the attention of user entities to different service entities and the attention of service entities to different user entities. It also effectively expresses the nonlinear dynamic interaction relationship between user entities and service entities, which helps to provide accurate feature representations for subsequent prediction processing.

[0040] S104. Based on the interaction fusion characteristics, perform QoS prediction to obtain the service quality result of the service entity relative to the user entity.

[0041] In this step, prediction is performed based on the interaction fusion features, thereby obtaining the service quality result of the service entity relative to the user entity, thus achieving QoS prediction. For example, the interaction fusion features are input into a pre-trained machine learning model, which outputs the service quality result of the service entity relative to the user entity. This machine learning model can be trained from multiple sample data and the labels of each sample data. The sample data includes user entity sample data and service entity sample data. The labels of the sample data are the service quality result of the service entity in the service entity sample data relative to the user entity in the user entity sample data. It should be understood that the same applies to user entity sample data and service entity sample data.

[0042] The steps described above will be explained in further detail below.

[0043] In some implementations, refer to Figure 2In step S102 above, embedding learning is performed on user entity data and service entity data to obtain user feature embedding matrix and service feature embedding matrix, including the following steps S201-S204, wherein the target entity data is either user entity data or service entity data. When the target entity data is user entity data, the target feature embedding matrix is ​​the user feature embedding matrix; when the target entity data is service entity data, the target feature embedding matrix is ​​the service feature embedding matrix. S201, assign corresponding indexes to each data item in the target entity data.

[0044] In this step, to eliminate the impact of entity heterogeneity on embedding learning, a function is used... Define a globally unique indexing scheme. Specifically, it assigns a unified index to the various contextual data of user entities and service entities. The various contextual data of user entities are denoted as... , The various contextual data of the service entity are recorded as follows: , .

[0045] S202, embed the indices of each data item in the target entity data into a continuous vector space to obtain the transformation matrix of each data item in the target entity data.

[0046] In this step, the indices of each data item in the target entity data are embedded to capture the latent features of the user entity and the service entity, thereby generating a representation that can be used for prediction, namely the embedding matrix. Specifically, for the indices of each data item in the target entity data, the current data index is embedded into a continuous vector space to obtain the transformation matrix of the current data. This method promotes similar indices to be adjacent to each other in the embedding space, thereby more effectively capturing the correlation and similarity between user entities or service entities, which helps to learn the implicit relationship between indices. All indices are embedded using the following formula (1): (1); In equation (1), Indicates allocation to the index Learnable embedding matrix; Indicates by embedding The obtained transformation matrix, i.e., the index The transformation matrix of the corresponding data; Indicates the dimension of index embedding; Indicates index The total number of unique indexes.

[0047] S203, optimize the transformation matrix of each data item in the target entity data to obtain the embedding matrix of each data item in the target entity data.

[0048] In this step, to improve the embedding operation in the previous step, an optimization and reconstruction mechanism for the embedding matrix is ​​constructed by using the DDPM-based denoising process framework. This mechanism is then used to optimize the transformation matrix of each data item in the target entity data, thereby obtaining the embedding matrix of each data item in the target entity data and thus obtaining a more accurate feature representation.

[0049] Furthermore, in step S203 above, the transformation matrix of each data item in the target entity data is optimized to obtain the embedding matrix of each data item in the target entity data, including: Noise prediction is performed on the transformation matrix of each data item in the target entity data to obtain the noise prediction components of each data item in the target entity data; By utilizing the noise prediction components of each data point in the target entity data, the transformation matrix of each data point in the target entity data is subjected to denoising diffusion reconstruction processing to obtain the embedding matrix of each data point in the target entity data.

[0050] Here, for the transformation matrix of each data item in the target entity data, we have: The transformation matrix of the current data is input into the conditional diffusion decoder, which outputs the noise prediction component of the current data. The conditional diffusion decoder consists of a multi-head attention layer, a regularization layer, and a multilayer perceptron connected in sequence. Unlike the DDPM mechanism, which uses U-Net to process continuous domain data, this implementation innovatively constructs a conditional diffusion decoder based on the Transformer architecture to address the discontinuity and global dependency characteristics of the discretized embedding matrix on the manifold structure. This design explicitly models the nonlinear interaction between discrete embedding units through a self-attention mechanism and uses a multi-head attention weight matrix to dynamically adjust the cross-dimensional denoising intensity distribution, ultimately achieving the probabilistic reconstruction of the ideal embedding state, as shown in the following formula (2): (2); In equation (2), Indicates the index predicted by the conditional diffusion decoder. The corresponding noise prediction component of the data; This represents the conditional diffusion decoder. During the denoising diffusion reconstruction process, the noise characteristics predicted by the network are explicitly taken into consideration, making the reconstructed embedding matrix closer to the ideal low-noise state. This allows the embedding matrix to better represent the inherent structure of the data and enhance the feature representation capability of the embedding.

[0051] After obtaining the noise prediction component of the current data, the transformation matrix of the current data is denoised and reconstructed using DDPM and the noise prediction component to obtain the embedding matrix of the current data. Specifically, to reconstruct the embedding matrix under ideal conditions, the noise prediction component of the current data needs to be subtracted from the embedding matrix. Since the embedding matrix uses the Kaiming initialization method during the initialization phase, for any embedding matrix... , , The identity matrix, whose dimensions are the same as those of the embedding vectors, is used to indicate that each dimension is independent and has the same variance. That is, the transformation matrix for each dimension satisfies the following formula (3): , (3).

[0052] In the forward diffusion process of DDPM, if we assume the diffusion step size is 1, then the following formula (4) is satisfied: (4).

[0053] This can be expressed as the following formula (5): , (5).

[0054] In the definition of DDPM, both satisfy... Furthermore, to ensure that the distribution of the embedding matrix is ​​consistent with the assumptions of the diffusion process, a noise figure is set. retention factor Therefore, the embedding reconstruction formula for the denoising diffusion process can be strictly expressed as the following formula (6): , (6); In equation (6), The embedding matrix represents the ideal state diagram reconstructed through the denoising diffusion process, which can also be regarded as the learning result of the embedding learning stage. This represents the prediction result for the noise components contained in the embedding matrix.

[0055] S204 integrates the embedding matrices of each data item in the target entity data to obtain the target feature embedding matrix.

[0056] In this step, in order to construct the feature embedding matrix, multiple feature transformation matrices obtained from users and services are fused during the embedding learning stage, thereby integrating multiple feature information related to users and services.

[0057] Furthermore, in step S204 above, the embedding matrices of each data item in the target entity data are integrated to obtain the target feature embedding matrix, including: The mean embedding matrix of each data item in the target entity data is obtained by averaging the embedding matrix of the target entity data. The mean embedding matrix of the target entity data is normalized to obtain the target feature embedding matrix.

[0058] Here, when the target entity data is user entity data, the target feature embedding matrix is ​​the user feature embedding matrix, as shown in the following formula (7): (7); In equation (7), Represents the user feature embedding matrix; Represents the first in the user entity data Embedding matrix of data; This indicates normalization, which aims to stabilize the numerical distribution of the embedding matrix and thus improve the generalization performance of the model.

[0059] When the target entity data is the service entity data, the target feature embedding matrix is ​​the service feature embedding matrix, as shown in the following formula (8): (8); In equation (8), Represents the service feature embedding matrix; Indicates the first in the service entity data An embedding matrix of data.

[0060] Therefore, this implementation method has multiple advantages over traditional GNN-based methods in both theory and practice: First, it incorporates a Transformer-based attention mechanism during the diffusion process, enabling each node to access the complete set of index embeddings. This global receptive field characteristic cannot be easily achieved by shallow GNNs. Second, unlike GNNs that heavily rely on structural adjacency relationships for inferring similarity, this implementation method, based on DDPM, directly learns the aforementioned similarities in the embedding space. This is better suited to heterogeneous and context-rich data scenarios, such as those common in user-service interaction scenarios. This flexibility can effectively capture complex relationships that are not strictly local or topologically significant. Third, graph convolutional networks perform poorly when dealing with low-degree nodes due to limited neighborhood information, resulting in weak representation capabilities or insufficient training. In contrast, this implementation method utilizes the global statistical properties of the entire graph during the diffusion-based embedding reconstruction stage to generate information-rich prior representations for each node. This preprocessing method ensures that even sparsely connected nodes retain sufficient discriminative features, thereby improving robustness in cold-start and sparse data scenarios.

[0061] In some implementations, refer to Figure 3 In user-service interaction modeling, traditional matrix factorization (MF) methods map users and services to a low-dimensional latent space, capturing user preference relationships in the form of matrix inner products. However, MF methods only focus on simple linear relationships between users and services, ignoring potential complex nonlinear interactions and making it difficult to accurately capture the intrinsic characteristics of interactions. Furthermore, CF-based methods achieve recommendation and prediction by analyzing the similarity between users or services. While performing well in some scenarios, they are highly dependent on the density of user interaction data, making them difficult to handle sparse data environments and susceptible to data noise and cold start problems.

[0062] In contrast, this implementation provides an attention-based modeling method in step S103, which can more flexibly and accurately capture the attention of user entities to different service entities and the attention of service entities to different user entities by adaptively learning the user-service interaction weight distribution, thereby effectively expressing the nonlinear dynamic interaction relationship between users and services. Particularly noteworthy is that the attention mechanism can clearly highlight key features that contribute to QoS prediction, effectively alleviate data sparsity and cold start problems, and demonstrate significant advantages in modeling complex interaction relationships.

[0063] Specifically, in step S103 above, the user feature embedding matrix and the service feature embedding matrix are used to perform feature interaction to obtain interactive fusion features, including the following steps S301-S303.

[0064] S301, Input processing is performed on the user feature embedding matrix and the service feature embedding matrix to obtain the first input feature and the second input feature.

[0065] It should be noted that the first input feature and the second input feature are features obtained through different input processing.

[0066] In this step, the user feature embedding matrix is ​​processed as input to obtain the first input feature; at the same time, the service feature embedding matrix is ​​processed as input to obtain the second input feature.

[0067] Further, in step S301 above, the user feature embedding matrix and the service feature embedding matrix are processed as input to obtain the first input feature and the second input feature, including: The user feature embedding matrix and the service feature embedding matrix are fused to obtain joint features; The joint features are processed based on the first bias and the first weight to obtain the first input features; The joint features are processed based on the second bias and the second weight to obtain the second input features; Here, the first bias and the second bias are biases with different values, and the first weight and the second weight are weights with different values.

[0068] Here, firstly, the user feature embedding matrix and the service feature embedding matrix are fused to obtain joint features. For example, a multi-head attention mechanism can be used for fusion; alternatively, the two can be directly concatenated. , , The joint features are represented. Then, the joint features are processed differently through two layers of linear transformation and nonlinear activation functions. Specifically, the joint features are processed based on the first bias and the first weight to obtain the first input features, and the joint features are processed based on the second bias and the second weight to obtain the second input features. The aim is to map the joint features to a 256-dimensional hidden space, as shown in the following formula (9): , , (9); In equation (9), Indicates the first input feature; Indicates the second input feature; Indicates the first weight; Indicates the first bias; Indicates the second weight; Indicates the second bias; This represents a non-linear activation function, such as the ReLU activation function, but is not limited to this. It should be understood that both weights and biases are values ​​that can be set according to the specific circumstances.

[0069] S302, perform multi-head attention operation on the first input feature and the second input feature to obtain the first attention feature and the second attention feature.

[0070] It should be noted that the first attention feature and the second attention feature are features obtained through different queries, keys, and values.

[0071] In this step, the first input feature and the second input feature are subjected to a first type of multi-head attention operation to obtain the first attention feature; at the same time, the first input feature and the second input feature are subjected to a second type of multi-head attention operation to obtain the second attention feature.

[0072] Further, in step S302 above, multi-head attention processing is performed on the first input feature and the second input feature to obtain the first attention feature and the second attention feature, including: Using the first input feature as the query and the second input feature as the key and value, multi-head attention operation is performed on the first and second input features to obtain the first attention feature.

[0073] Further, in step S302 above, multi-head attention processing is performed on the first input feature and the second input feature to obtain the first attention feature and the second attention feature, including: Using the second input feature as the query and the first input feature as the key and value, multi-head attention operation is performed on the first and second input features to obtain the second attention feature.

[0074] Here, in order to explicitly model the bidirectional attention dependency between users and services, two multi-head attention mechanisms are defined to characterize the interaction information in two directions: the user entity pays attention to the features of the service entity and the service entity pays attention to the features of the user entity, as shown in the following formula (10): , (10); In equation (10), Indicates the primary attentional feature; Indicates the second attentional feature; This represents the first multi-head attention mechanism; This represents the second multi-head attention mechanism; multi-head attention mechanisms are used to capture dependency information between interacting parties.

[0075] S303, the first attention feature and the second attention feature are fused to obtain the interactive fusion feature.

[0076] In this step, after completing the multi-head attention operation, the first attention feature and the second attention feature are fused together to obtain the interactive fused feature. For example, a multi-head attention mechanism can be used for fusion; alternatively, the two can be directly concatenated. , This indicates interactive fusion features.

[0077] Therefore, traditional CF-based neural network models generate static interaction weights for each user-service pair using a single global parameter set, which are insensitive to instance-specific contexts. In contrast, this implementation introduces an attention-driven interaction layer that adaptively reallocates weights, selectively amplifying information-rich feature dimensions while suppressing noisy features. Furthermore, the proposed dual-path attention design explicitly models the bidirectional dependencies between user entities and service entities, as well as between service entities and user entities. This capability captures asymmetric behavioral patterns that standard collaborative filtering multilayer perceptrons (MLPs) cannot represent, as the latter only learn symmetric mappings. Finally, the multi-head mechanism provided in this implementation achieves tensor-level reorganization in a single forward propagation, essentially expressing second-order and higher-order feature interactions. To achieve equivalent performance with a classic fully connected multilayer perceptron (CF-MLP), more layers need to be stacked, increasing the number of parameters and exacerbating overfitting and gradient decay. These characteristics collectively endow the attention-based architecture with stronger adaptability, richer representation capabilities, and superior robustness compared to traditional CF models.

[0078] In some implementations, refer to Figure 3 In step S104 above, QoS prediction is performed based on the interaction fusion features to obtain the service quality result of the service entity relative to the user entity, including: By using a multilayer perceptron to perform nonlinear mapping and prediction processing on the interaction fusion features, the service quality results of the service entity relative to the user entity are obtained.

[0079] In this embodiment, the interactive fusion features are input into a three-layer perceptron, thereby performing nonlinear mapping and prediction processing to obtain the service quality result of the service entity relative to the user entity, as shown in the following formula (11): , , (11); In equation (11), Layer Normalization; Indicates the third weight. Indicates the third bias; Indicates the fourth weight. Indicates the fourth bias; Indicates the fifth weight. Indicates the fifth bias; It can be an activation function, such as the Sigmoid activation function, used to output normalized prediction results, but is not limited to this; This represents the service quality result. It should be understood that both weights and biases are values ​​that can be set according to the actual situation.

[0080] In some implementations, the above methods are integrated into a neural network model, which may include an input layer, an embedding learning layer, a feature interaction layer, and an output layer. The input layer performs step S101, the embedding learning layer performs step S102, the feature interaction layer performs step S103, and the output layer performs step S104. The parameters of each layer of the model are learned by optimizing the objective function. To achieve this, a method of minimizing the loss function is used for model training and parameter updates. The loss function can be flexibly set according to the actual situation; for example, the loss function can be the mean squared error (MSE) to measure the difference between the predicted value and the true value, but it is not limited to this. Furthermore, parameter updates can be performed using a backpropagation mechanism combined with optimization algorithms such as Adam. To verify the effectiveness of the method provided in this application, the method provided in this application will be compared with related technologies below.

[0081] Comparative experiments were conducted on the WSDream dataset to test the prediction performance of various methods for non-functional QoS attributes at four different matrix densities: 2.5%, 5%, 7.5%, and 10%. Non-functional QoS attributes include response time (RT) and throughput (TransactionsPerSecond, TP). Experimental results were generated based on three different random seeds and are presented as the mean and standard deviation of mean squared error and root mean square error (RMSE), as shown in Tables 1 and 2. In Tables 1 and 2, "Model" represents the method, where QoSDiff is the method used in this embodiment, and "Density" represents the matrix density.

[0082] Table 1: Performance Comparison of QoS Prediction Models in Response Time Metrics

[0083] Table 2: Performance Comparison of Various QoS Prediction Models in Throughput Metrics

[0084] As can be seen from Tables 1 and 2: UPCC, IPCC, and UIPCC are CF-based methods, which primarily calculate and predict missing quality of service (QoS) values ​​based on the direct similarity between users and services. Results show that these methods perform poorly under low data density and are highly sensitive to data sparsity, with performance significantly degrading, especially at a density of 2.5%.

[0085] PMF and BiasMF are methodologies based on root mean square error (RMSE). RMSE-based methods can capture the latent features of users and services well under high data density conditions. However, these methods are extremely sensitive to the selection of latent feature dimensions, leading to significant performance fluctuations under sparse or complex data conditions. For example, the unstable performance of PMF on the root mean square error metric highlights this shortcoming.

[0086] CSMF, NFMF, and NCRL are deep learning-based methods. Deep learning-based methods further enhance modeling capabilities by uncovering nonlinear and higher-order interaction relationships. Experimental results show that NFMF, due to its multi-task learning strategy, effectively extracts latent user-service features and outperforms other deep learning methods overall. In contrast, NCRL has lower overall performance and is insensitive to changes in matrix density, indicating that its embedding learning structure is imperfect.

[0087] Graph MF and QoS GNN are graph neural network-based methods. Graph neural network-based methods can explicitly capture the complex structural relationships between users and services, thereby achieving high prediction accuracy. However, their high dependence on explicit graph structures leads to excessive computational complexity, which limits their scalability and efficiency in large-scale scenarios.

[0088] Compared to the aforementioned existing methods, the embedding learning step based on the denoising diffusion mechanism in this application (i.e., step S102) can effectively model service features through implicit relational reasoning. This is particularly suitable for data scenarios lacking explicit graph structures and exhibiting high sparsity, fundamentally alleviating the sensitivity to sparsity and high computational overhead of CF and GNN methods. Simultaneously, the introduced self-attention hybrid interaction mechanism (i.e., step S103) can fully capture complex user-service relationships, effectively overcoming the shortcomings of MF and traditional DL methods in feature dimension selection and overfitting. Overall, the embodiments of this application demonstrate superior prediction performance under different data density conditions, fully validating its significant improvement in robustness and generalization ability.

[0089] To further verify the advantages of the method in QoS prediction tasks according to the embodiments of this application, two sets of ablation experiments were conducted to compare its performance with traditional methods. These ablation experiments were conducted on the WSDream dataset with matrix density conditions set to 2.5%, 5%, 7.5%, and 10%, as well as the EEL dataset with the same conditions. Mean Absolute Error (MAE) was used as the performance evaluation metric, primarily testing response time and throughput.

[0090] First, the embedding learning method of this application (i.e., step S102) is compared with existing embedding learning methods: To verify the effectiveness of the matrix transformation mechanism based on denoising diffusion (i.e., step S102) in capturing high-order relationships between users and services, this embodiment compares the performance of simple embedding representation and embedding representation enhanced by the denoising diffusion framework. The comparison results are as follows: Figure 4 and Figure 5 As shown, Figure 4 This section compares the performance of existing embedding learning methods with the embedding learning method provided in this application in terms of response time. Figure 5 This section compares the performance of existing embedding learning methods with the embedding learning method provided in this application in terms of throughput metrics. Figure 4 and Figure 5 The embedding learning method based on denoising is the embedding learning method provided in this application. It can be seen that this application introduces a denoising diffusion framework in embedding learning and obtains a lower MAE score under all test conditions. This shows that the denoising diffusion framework can effectively capture the dynamic relationship in the embedding, thereby improving the QoS prediction accuracy.

[0091] Then, the feature interaction method of this application (i.e., step S103) is compared with the existing inner product interaction method (MF) and collaborative filtering interaction method (CF): Feature interaction learning aims to capture the potential interaction relationship between users and services through transfer demonstrations. This embodiment examines three interaction strategies: MF, CF, and a hybrid attention mechanism to evaluate their impact on QoS prediction accuracy. The comparison results are as follows: Figure 6 and Figure 7 As shown, Figure 6 This section compares the performance of existing feature interaction methods and the feature interaction method of this application in terms of response time. Figure 7 The comparison of the performance of existing feature interaction methods and the feature interaction method of this application in terms of throughput metrics shows that the feature interaction method of this application achieves lower MAE values ​​under all matrix densities on the RTTP dataset, highlighting its significant advantages in prediction tasks.

[0092] Regarding hyperparameters, this application adjusts the embedding dimension to map users and service nodes to a high-dimensional feature space, thereby enhancing the model's ability to characterize complex latent patterns. Experimental results are as follows: Figure 8 As shown, the model's predictive performance exhibits a trend of first increasing and then stabilizing with the increasing embedding dimension. When the embedding dimension reaches 256 or higher, the model achieves optimal performance in both mean error and root mean square error. Therefore, in subsequent experiments and system implementation, it is recommended to uniformly adopt 256-dimensional embedding as the default configuration to balance performance and computational efficiency.

[0093] Furthermore, the multi-head attention mechanism introduced in this application can capture different noise features in parallel during the diffusion process, thereby improving the model's ability to represent and predict the noise distribution in the diffusion transformation matrix. Experimental results are as follows: Figure 9 As shown, the model performance gradually decreases with the increase of the number of attention heads, indicating that too many attention channels may lead to redundant feature representations and introduce noise interference. Considering both the prediction accuracy and computational complexity of the model, this application fixes the number of attention heads to 1 in the main experiments and deployments to achieve the best balance between performance and stability. In addition to the methods mentioned above, refer to Figure 10 This application also provides a QoS prediction device, including: Module 401 is used to acquire user entity data and service entity data. The first processing module 402 is used to perform embedding learning on user entity data and service entity data to obtain user feature embedding matrix and service feature embedding matrix. The second processing module 403 is used to perform feature interaction between the user feature embedding matrix and the service feature embedding matrix to obtain interactive fusion features. The third processing module 404 is used to perform QoS prediction based on the interaction fusion characteristics to obtain the service quality result of the service entity relative to the user entity.

[0094] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0095] In summary, this application addresses the core bottlenecks of existing QoS prediction methods in terms of scalability, implicit relationship modeling, and sparse interaction, and proposes a service quality prediction framework based on a denoising diffusion probability model and bidirectional hybrid self-attention.

[0096] Existing methods heavily rely on constructing explicit user-service bipartite graphs. As the service ecosystem expands, the construction and reasoning of graph structures face scalability challenges due to high computational complexity and storage overhead. Furthermore, graph neural network-based methods struggle to effectively capture latent implicit relationships in the absence of explicit topological connections (such as in cross-domain service composition scenarios), limiting their modeling capabilities. In addition, traditional collaborative filtering or matrix factorization methods are highly sensitive to data sparsity and struggle to characterize complex nonlinear high-order interactions between users and services, resulting in insufficient prediction accuracy and stability in real-world noisy data environments.

[0097] To systematically address the aforementioned issues, this application aims to overcome the limitations of explicit graph structures. Through a denoising diffusion-driven implicit embedding learning mechanism, robust user and service representations are directly reconstructed from the data distribution without constructing a massive physical graph, fundamentally improving the model's scalability and adaptability to sparse data. Furthermore, a bidirectional hybrid self-attention interaction mechanism is introduced to replace traditional unidirectional or shallow interaction modeling. This dynamically captures fine-grained, asymmetric semantic dependencies from both user-to-service and service-to-user perspectives, enabling deep expression and accurate prediction of complex interaction relationships. Finally, by integrating the inherent noise resistance of the denoising diffusion model into the entire embedding learning process, the model's performance on large-scale real-world datasets is significantly improved, ensuring the reliability and stability of prediction results in complex real-world environments. These innovations effectively enhance QoS prediction accuracy.

[0098] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0099] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0100] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several programs to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit a program for use by or in conjunction with a program execution system, apparatus, or device.

[0102] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0103] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable program execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0104] In the foregoing description of this specification, the references to terms such as "one embodiment," "another embodiment," or "some embodiments," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0105] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0106] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A QoS prediction method, characterized in that, Includes the following steps: Obtain user entity data and service entity data; Embedding learning is performed on the user entity data and the service entity data to obtain the user feature embedding matrix and the service feature embedding matrix; The user feature embedding matrix and the service feature embedding matrix are subjected to feature interaction to obtain interactive fusion features; Based on the interaction fusion features, QoS prediction is performed to obtain the service quality result of the service entity relative to the user entity.

2. The method according to claim 1, characterized in that, The embedding learning of the user entity data and the service entity data to obtain the user feature embedding matrix and the service feature embedding matrix includes: Assign corresponding indexes to each data item in the target entity data; The indexes of each data item in the target entity data are embedded into a continuous vector space to obtain the transformation matrix of each data item in the target entity data; The transformation matrix of each data item in the target entity data is optimized to obtain the embedding matrix of each data item in the target entity data; The embedding matrices of each data item in the target entity data are integrated and processed to obtain the target feature embedding matrix; Wherein, the target entity data is either the user entity data or the service entity data. When the target entity data is the user entity data, the target feature embedding matrix is ​​the user feature embedding matrix. When the target entity data is the service entity data, the target feature embedding matrix is ​​the service feature embedding matrix.

3. The method according to claim 2, characterized in that, The optimization process of the transformation matrix of each data item in the target entity data to obtain the embedding matrix of each data item in the target entity data includes: Noise prediction is performed on the transformation matrix of each data item in the target entity data to obtain the noise prediction component of each data item in the target entity data; Using the noise prediction components of each data item in the target entity data, the transformation matrix of each data item in the target entity data is subjected to denoising diffusion reconstruction processing to obtain the embedding matrix of each data item in the target entity data.

4. The method according to claim 2, characterized in that, The process of integrating the embedding matrices of each data item in the target entity data to obtain the target feature embedding matrix includes: The embedding matrix of each data item in the target entity data is averaged to obtain the mean embedding matrix of the target entity data. The mean embedding matrix of the target entity data is normalized to obtain the target feature embedding matrix.

5. The method according to claim 1, characterized in that, The step of performing feature interaction between the user feature embedding matrix and the service feature embedding matrix to obtain interactive fusion features includes: The user feature embedding matrix and the service feature embedding matrix are processed as inputs to obtain a first input feature and a second input feature; wherein the first input feature and the second input feature are features obtained through different input processes; Multi-head attention processing is performed on the first input feature and the second input feature to obtain a first attention feature and a second attention feature; wherein the first attention feature and the second attention feature are features obtained through different queries, keys and values; The first attention feature and the second attention feature are fused together to obtain the interaction fusion feature.

6. The method according to claim 5, characterized in that, The step of processing the user feature embedding matrix and the service feature embedding matrix to obtain the first input feature and the second input feature includes: The user feature embedding matrix and the service feature embedding matrix are fused to obtain joint features; The joint features are processed based on the first bias and the first weight to obtain the first input features; The joint features are processed based on the second bias and the second weight to obtain the second input features; Wherein, the first bias and the second bias are biases with different values, and the first weight and the second weight are weights with different values.

7. The method according to claim 5, characterized in that, The step of performing multi-head attention operation on the first input feature and the second input feature to obtain the first attention feature and the second attention feature includes: Using the first input feature as the query and the second input feature as the key and value, a multi-head attention operation is performed on the first input feature and the second input feature to obtain the first attention feature.

8. The method according to claim 5, characterized in that, The step of performing multi-head attention operation on the first input feature and the second input feature to obtain the first attention feature and the second attention feature includes: Using the second input feature as the query and the first input feature as the key and value, a multi-head attention operation is performed on the first input feature and the second input feature to obtain the second attention feature.

9. The method according to claim 1, characterized in that, The step of performing QoS prediction based on the interaction fusion features to obtain the service quality result of the service entity relative to the user entity includes: The interaction fusion features are nonlinearly mapped and predicted using a multilayer perceptron to obtain the service quality result of the service entity relative to the user entity.

10. A QoS prediction device, characterized in that, include: The acquisition module is used to acquire user entity data and service entity data. The first processing module is used to perform embedding learning on the user entity data and the service entity data to obtain a user feature embedding matrix and a service feature embedding matrix. The second processing module is used to perform feature interaction between the user feature embedding matrix and the service feature embedding matrix to obtain interactive fusion features. The third processing module is used to perform QoS prediction based on the interaction fusion features to obtain the service quality result of the service entity relative to the user entity.

Images