Abstract
Web page classification is a complex task due to the multi-view nature of web data, the scarcity of labeled examples, and high-dimensional feature spaces. This paper proposes a novel feature extraction method, Uncorrelated Semi-supervised Intra-view and Inter-view Manifold Discriminant (USI$^2$MD) learning. USI$^2$MD integrates discriminant information from both within and across different data views while enforcing an uncorrelation constraint to minimize redundancy. By leveraging manifold learning for unlabeled data, the method achieves significant improvements in classification accuracy and feature discriminability compared to state-of-the-art baselines.
Introduction
The modern web presents a "multi-view" challenge: a single page is defined by its textual content, its hyperlink context, and its associated media. While the volume of web data is immense, the cost of manual labeling remains high, leading to a semi-supervised learning environment. Furthermore, raw web features are often high-dimensional and redundant. Existing feature extraction methods frequently overlook the "inter-view" relationships—how different views of the same object relate to its class—and fail to address the high correlation between views, which can degrade classifier performance.
The USI$^2$MD Methodology
The proposed USI$^2$MD approach addresses these challenges through two primary components: the SI$^2$MD schema and an uncorrelation constraint.
SI$^2$MD Schema
The core of the method is the Semi-supervised Intra-view and Inter-view Manifold Discriminant (SI$^2$MD) schema. Using graph-based modeling, it constructs:
- Intrinsic Graphs: These graphs aim to keep samples of the same class, as well as nearby unlabeled samples, close together in the projected feature space.
- Penalty Graphs: These graphs work to maximize the distance between samples of different classes.
Crucially, this modeling occurs both within individual views (intra-view) and across different views (inter-view), capturing a richer set of discriminative features than traditional single-view or consistency-based multi-view methods.
Uncorrelation Constraint
To combat redundancy, USI$^2$MD introduces a mathematical constraint that forces extracted features from different views to be uncorrelated. This ensures that each view provides unique information to the classifier, effectively filtering out the "noise" of overlapping data and leading to a more robust representation.
Experimental Evaluation
The method was evaluated on two benchmark datasets:
- WebKB: Contains web pages from computer science departments (Course vs. Non-Course) with page and link views.
- Internet Advertisements (AD): Focuses on identifying advertisement images using Base URL, Destination URL, and Image URL views.
USI$^2$MD was compared against seven state-of-the-art methods, including SSGCA, MVML-GL, VRKHS, and MHL.
Results and Discussion
Experimental results demonstrate that USI$^2$MD consistently outperforms baseline methods in terms of classification accuracy across varying ratios of labeled data.
As shown in the visualizations of baseline methods, existing approaches like SSGCA and MHL often result in feature spaces where different classes remain significantly "mingled," as highlighted by the green circles in the scatter plots.
In contrast, the USI$^2$MD method is reported to produce much clearer, distinct clusters, facilitating easier classification. Statistical analysis using McNemar’s test confirmed that the performance gains of USI$^2$MD are statistically significant (p < 0.05).
Conclusion
USI$^2$MD represents a significant step forward in multi-view semi-supervised learning. By explicitly modeling inter-view discriminative information and enforcing feature uncorrelation, it provides a more effective way to handle the complexities of web page data. Future work may explore the application of this framework to other high-dimensional domains such as bioinformatics and video analysis.
Summary updated