Abstract
As the ecosystem of Web APIs expands, efficient discovery and categorization become increasingly difficult due to the inherent noise and redundancy in user-generated tags. This paper introduces Ontology Extended Labeled LDA (OEL-LDA), a framework that enhances standard supervised topic modeling by integrating a semi-automatic ontology-based pre-processing step. By aggregating specific, redundant tags into high-level "ancestor concepts," OEL-LDA achieves a significant improvement in topic coherence and interpretability.
Introduction
With the number of Web APIs exceeding 8,000, automated categorization is essential for service retrieval. While Latent Dirichlet Allocation (LDA) is a standard tool for this task, "Standard Labeled LDA" (SL-LDA) often fails when applied to real-world data. User-provided tags are frequently redundant (e.g., "blog" vs. "blogging") or vary wildly in granularity (e.g., "GPS" vs. "Map"). When a model treats every unique tag as a distinct topic, the resulting categories are fragmented and difficult for humans to navigate. OEL-LDA addresses this by bridging the gap between messy folksonomies and formal taxonomies.
Methodology
The core innovation of OEL-LDA lies in its semantic pre-aggregation pipeline, which acts as a "denoiser" for input labels before they are fed into the topic model.
- Ontology Construction: The system builds a tag-to-tag co-occurrence matrix from API data and applies hierarchical clustering to identify natural groupings.
- Semantic Mapping: These clusters are refined using formal ontologies (such as SUMO and DBPedia) to establish "subclass-of" relationships.
- Pre-aggregation: The original 133 tags are mapped to 26 broad "ancestor concepts." For instance, specific tags like "hotel" and "booking" are rolled up into the "Travel" concept.
- Model Training: These cleaned, high-level concepts serve as the supervised labels for the LDA model, ensuring that the generated topics are semantically distinct.
Experimental Results
The model was evaluated using a dataset of 4,815 multi-labeled APIs crawled from ProgrammableWeb. The primary metric for success was topic coherence, which measures how logically related the top words in a topic are to one another.
OEL-LDA outperformed standard SL-LDA with an average gain of 7.0% in topic coherence. Furthermore, the model successfully reduced the dimensionality of the label space from 133 noisy tags to 26 interpretable topics. Validation via cosine similarity heat maps confirmed that the specific topics from the original data mapped with high precision to their respective ancestor concepts in the OEL-LDA model.
Conclusion
OEL-LDA demonstrates that improving the quality of input labels through ontological aggregation is as critical as the underlying algorithm for generating meaningful topics. By reducing noise and redundancy, the model provides a more robust framework for Web API discovery. Future work will explore the use of Linked Open Data (LOD) to further automate ontology construction and support more complex API chaining.
Summary updated