Abstract
Web service recommendation based on the quality of service (QoS) is important for users to find the exact Web service among many functionally similar Web services. Although service recommendations have been recently studied, the performance of the existing ones is unsatisfactory because: 1) the current QoS predicting algorithms still experience data sparsity and cannot predict the QoS values accurately and 2) the previous approaches fail to consider the QoS variance according to the users and services' locations carefully. A Web service recommendation method based on the QoS prediction and hierarchical tensor decomposition is proposed in this paper. The method is called QoSHTD that is based on location clustering and hierarchical tensor decomposition. First, the users and services of the QoSHTD cluster into several local groups based on their location and models local and global triadic tensors for the user-service-time relationship. The hierarchical tensor decomposition is then performed on the local and global triadic tensors. Finally, the predicted QoS value through local and global tensor decomposition is combined as the missing QoS values. The comprehensive experiment shows that the proposed method achieves a high prediction accuracy and recommending quality of Web service, and can partially address data sparsity.
Introduction
In the modern service-oriented architecture (SOA) economy, the proliferation of Web services has led to a "choice overload" for users. Many services offer identical functionalities, such as weather updates or payment processing. To distinguish between them, developers rely on Quality of Service (QoS) metrics, primarily Response Time (RT) and Throughput (TP).
However, building effective recommendation engines faces two critical challenges:
- Data Sparsity: Real-world datasets are often extremely sparse, as users only interact with a tiny fraction of available services.
- Spatiotemporal Variance: QoS is highly dependent on geography and time. A service may perform well in one region at a specific time but poorly elsewhere or at a different hour.
The QoSHTD Framework
To address these issues, this research introduces QoSHTD (QoS Prediction based on Hierarchical Tensor Decomposition). The framework integrates geographic clustering with advanced tensor modeling to capture complex multidimensional relationships.
Location-Based Clustering
The process begins by grouping users and services into "local" neighborhoods using K-means clustering based on IP-derived coordinates. This ensures that the model accounts for network proximity, which is a primary driver of latency.
Triadic Tensor Modeling and Decomposition
Unlike traditional 2D matrix models (User x Service), QoSHTD employs a 3D triadic tensor (User x Service x Time). This structure allows the model to capture temporal fluctuations in performance.
The core of the method is Hierarchical Tensor Decomposition. As illustrated in the architecture below, the system decomposes tensors at two levels: the local level, which captures nuances specific to geographic clusters, and the global level, which captures general performance trends across the entire dataset.
The decomposition is solved using multi-layer neural networks with ReLU activation functions, mapping high-dimensional data into a lower-dimensional latent space to accurately predict missing QoS values.
Hybrid Prediction
Final predictions are generated by combining local and global results using a linear weighting parameter (alpha). This allows the model to prioritize regional data—which is often more relevant—while still benefiting from the broader patterns found in the global dataset.
Experimental Evaluation
The researchers evaluated QoSHTD using the WSDream dataset, which contains QoS records for 142 users and 4,532 Web services over 64 time slots.
Key Findings
- Sparsity Resilience: At a data density of only 1%, QoSHTD significantly outperformed baseline methods like Collaborative Filtering (WSRec) and standard Tensor Decomposition (TD). For Response Time, QoSHTD achieved a Mean Absolute Error (MAE) of approximately 0.54, compared to ~0.60 for WSRec.
- Optimal Weighting: Experiments revealed that an alpha value between 0.7 and 0.8 yields the best results, indicating that while global trends are foundational, local geographic context provides the necessary fine-tuning for high-precision recommendations.
- Ranking Quality: Beyond numerical accuracy, QoSHTD demonstrated superior performance in ranking "top-K" services, as measured by Recall and nDCG (Normalized Discounted Cumulative Gain).
Conclusion
QoSHTD effectively addresses the challenges of data sparsity and spatiotemporal variance in Web service recommendation. By leveraging hierarchical tensor decomposition and geographic clustering, it provides a robust framework for predicting service performance. Future work will explore integrating subjective user preferences and social network data to further refine the "cold-start" performance for new users.
Summary updated