Abstract
This paper introduces CoNMF (Co-regularized Non-negative Matrix Factorization), a framework designed to enhance the clustering of Web 2.0 items, such as music and videos, by utilizing user comments as a rich, multi-dimensional data source. By jointly factorizing item metadata, comment text, and commenter identities, CoNMF achieves significant performance improvements—up to 14% in F1 score—over traditional single-view clustering methods.
Introduction
In the Web 2.0 era, platforms are inundated with user-generated content, making effective organization through semantic clustering vital for search and recommendation systems. Traditionally, clustering has relied on intrinsic features, such as an item's title or description. However, these features are often sparse, noisy, or insufficient to capture the full semantic context of an item.
The authors propose that user comments offer a valuable "extrinsic" information source. Comments provide not only textual content but also reveal the identities of interested users. The primary challenge lies in integrating these heterogeneous data sources (text vs. user IDs), which vary significantly in scale and quality. Simple concatenation often fails as low-quality views can obscure high-quality signals.
The CoNMF Framework
The proposed CoNMF framework extends standard Non-negative Matrix Factorization (NMF) into a multi-view setting. It considers three distinct views for each item:
- Description View: Intrinsic metadata (e.g., artist biographies).
- Comment Words View: The textual content of user comments.
- Users View: The identities of the users who have commented on the item.
Co-regularization and Normalization
CoNMF employs a co-regularization term to encourage similarity between the cluster assignment matrices of different views without forcing a rigid consensus. Two paradigms are explored: Pair-wise CoNMF (CoNMF-P), which minimizes distances between pairs of views, and Cluster-wise CoNMF (CoNMF-C), which aligns cluster-to-cluster similarity matrices.
A key technical insight is the necessity of $L_2$ normalization during iterative updates. Because feature vector lengths vary drastically across views (e.g., a short bio vs. thousands of comments), normalization prevents the model from becoming biased toward views with higher feature density.
Experimental Evaluation
The researchers evaluated CoNMF on two real-world datasets: Last.fm (9,694 music artists across 21 genres) and Yelp (2,624 businesses across 7 categories).
Key Results
The experiments demonstrated that CoNMF consistently outperforms both single-view baselines and state-of-the-art multi-view methods. On the Last.fm dataset, CoNMF improved the F1 score by 14% over standard $k$-means.
One of the most striking findings is that user identity is often a more predictive signal than the text of the comments themselves. Users typically exhibit niche interests, making their "commenting footprint" a highly effective feature for identifying item categories.
Conclusion
CoNMF provides a robust solution for multi-view clustering by effectively balancing information from diverse sources. The use of soft constraints via pair-wise regularization allows high-quality views to guide the clustering process without being compromised by noisier data. Future work may involve addressing fixed cluster counts through Matrix Tri-factorization and incorporating temporal dynamics to account for changing user interests over time.
Summary updated