Abstract
Web service discovery often struggles with the semantic gap between technical WSDL descriptions and user intent. This paper presents WTO, a framework that leverages user-provided tags through a spectral co-clustering approach. By integrating automated tag recommendation to mitigate data sparsity and simultaneously clustering services and descriptive terms, WTO significantly enhances the accuracy of functional service grouping compared to traditional keyword-based methods.
Introduction: The Discovery Challenge
As the ecosystem of Web services expands, traditional discovery mechanisms like UDDI registries have become obsolete, leaving keyword-based search as the primary tool. However, keyword search suffers from the "vocabulary problem"—it requires exact matches between user queries and service descriptions. While clustering services into functional categories can aid discovery, standard approaches relying solely on Web Service Description Language (WSDL) files are often insufficient. WSDL files are technical, inconsistent, and lack the domain-specific context that human users naturally provide through tags.
The Problem of Tag Sparsity
While user tags offer valuable contextual "glue," they are notoriously sparse and noisy. Most services have very few tags, creating a "long tail" distribution that makes traditional statistical analysis difficult. This sparsity necessitates a pre-processing step to "densify" the data before effective clustering can occur.
The WTO Framework
To address these challenges, the authors propose the Web service Tagging-based co-clustering (WTO) framework. WTO moves beyond one-way clustering (like k-means) by employing a bipartite graph model to cluster services and words simultaneously.
WCCluster: Simultaneous Co-Clustering
The core engine, WCCluster, treats the relationship between services ($ws$), words ($w$), and tags ($t$) as a bipartite graph. The goal is to find an optimal "minimum cut" that partitions the graph into clusters where nodes within a cluster are densely connected. This approach captures the dual nature of service data: similar services are described by similar words, and similar words are used to describe similar services.
Data Integration and Recommendation
WTO employs three key pillars to refine the input data:
- Tag Recommendation: To solve the sparsity problem, the framework uses Jaccard similarity and aggregation methods (Vote and Sum) to suggest relevant tags for services with few or no tags.
- Integration Strategies: The authors found that treating tags as "new words" in the service's vocabulary was the most effective way to combine technical WSDL data with crowdsourced tags.
- Spectral Partitioning: The framework uses Singular Value Decomposition (SVD) to project the high-dimensional service-term matrix into a lower-dimensional space before applying k-means, making the clustering more robust to noise.
Key Results and Insights
The researchers evaluated WTO using datasets crawled from the Titan search engine. The results demonstrated a clear advantage for the co-clustering approach:
- Accuracy: On a 5-category dataset, WCCluster improved precision by 30.8% over standard k-means.
- The WTO² Advantage: The version of the framework that included recommended tags (WTO²) consistently outperformed the version using only original tags (WTO¹), proving that automated tag expansion is a vital pre-processing step.
- Domain Context: Analysis showed that tags often describe the domain (e.g., "Research," "Business") while WSDLs describe the mechanics (e.g., "get," "request"). Combining them provides a holistic view of the service.
Conclusion
The WTO framework demonstrates that user-generated content, despite its noise and sparsity, is a powerful asset for Web service discovery when handled with sophisticated co-clustering and recommendation techniques. Future work will focus on scaling these algorithms using distributed frameworks like Hadoop to manage the ever-growing global catalog of Web services.
Summary updated