Introduction
In the early 2000s, human-curated directories like Yahoo! and the Open Directory Project (DMOZ) served as the primary navigational maps for the World Wide Web. However, these directories often employed disparate organizational logic; for instance, one might categorize "Autos" by technical specifications while another focused on "Classic Cars." Integrating these diverse hierarchies is a significant challenge, as traditional methods often flatten the directory tree into a simple list, thereby discarding the rich contextual information embedded in the "is-a" relationships (e.g., "Mustang" under "Ford" under "Automotive").
Traditional machine learning approaches, such as Naive Bayes, typically rely on content analysis. However, text-based classification is prone to ambiguity—a page containing the word "bank" could refer to financial services or river ecology. To address these limitations, this paper introduces Similarity-Based Integration (SBI), a method that leverages instance-based learning to map categories between directories without relying on textual content.
Methodology: Similarity-Based Integration (SBI)
The core innovation of SBI is its reliance on shared instances (URLs) rather than page content. By observing where human editors have placed the same webpage across different directories, the algorithm captures semantic alignment through a statistical lens.
The Kappa Statistic
SBI utilizes the $k$-statistic (Kappa) to measure the agreement between two categorization criteria. Unlike simple overlap measures, Kappa accounts for the probability of agreement occurring by chance. If two categories in separate directories share a significant number of URLs, the $k$-statistic identifies them as a similar category pair.
Hierarchical Processing
Unlike "flattening" methods, SBI operates directly on the directory's hierarchical structure. It assumes a transitive relationship: if a document belongs to a sub-category, it inherently belongs to all ancestor categories. To maintain computational efficiency, the algorithm employs a top-down recursive search. By identifying matches at higher levels of the tree, the system can prune the search space, avoiding exhaustive comparisons across the entire directory. In cases where a direct match is unavailable, a fallback policy allows the system to inherit mappings from parent categories.
Experimental Evaluation
The researchers evaluated SBI using real-world data from the Yahoo! and Google (DMOZ-based) directories across five domains: Autos, Movies, Outdoors, Photography, and Software. The performance was benchmarked against Enhanced Naive Bayes (E-NB), a state-of-the-art text-based classifier.
Results
The experimental results demonstrated that SBI significantly outperforms text-based methods. On average, SBI achieved an accuracy of 78.4% when mapping from Yahoo! to Google, compared to 64.4% for E-NB—an improvement of 14%.
The most notable success occurred in the "Movies" domain, where SBI saw a 35.3% improvement over E-NB. This is attributed to the fact that movie-related pages often contain highly diverse keywords (e.g., actor names, plot summaries, settings) that confuse text-based classifiers but do not impact a structural, instance-based approach.
Conclusion and Future Work
The success of SBI underscores two key principles: human-curated placement is a powerful proxy for semantic meaning, and hierarchical structure is a valuable data feature. By preserving the "is-a" relationships, SBI maintains the context that text-only classifiers lose.
However, the method is currently limited by its dependency on shared links; it cannot bridge directories with zero overlapping URLs. Future research may explore hybrid models that combine SBI's structural mapping with text-based analysis to provide robust integration even in the absence of shared instances.
Summary updated