A Graphbased Framework for Multi-Task Multi-View Learning

Jingrui He, Rick Lawrence

Introduction

In real-world machine learning applications, data often exhibits "dual heterogeneity." First, there is Feature Heterogeneity (Multi-View), where a single task involves different types of data. For instance, a webpage can be described by its text content, the anchor text of incoming links, and the words in its URL. Second, there is Task Heterogeneity (Multi-Task), where we must solve several related but distinct tasks, such as classifying webpages for four different universities.

Historically, researchers have treated these two forms of heterogeneity separately. Multi-task learning (MTL) leverages related tasks but ignores the internal consistency of multiple views. Conversely, multi-view learning (MVL) ensures different feature sets agree on a label but ignores information from other tasks.

This paper introduces a novel framework for Multi-Task Multi-View ($M^2TV$) learning. By using a graph-based approach to enforce consistency both across views and across tasks, the authors demonstrate significant performance gains, particularly in scenarios where labeled data is scarce.

The GraM2 Framework and IteM2 Algorithm

To bridge the gap between multi-task and multi-view learning, the authors propose a graph-based framework called GraM2 and an accompanying iterative solver, IteM2.

Bipartite Graph Construction

For every task and every view, the framework constructs a bipartite graph. One set of nodes represents the examples, and the other represents the features. The edges connecting them represent the feature values. This structure allows the model to handle tasks with different numbers of views and varying feature dimensions without requiring complex manual alignment.

Bipartite graph structure illustrating shared and task-specific views across multiple tasks.

Dual Consistency Constraints

The GraM2 framework relies on two primary consistency constraints:

  1. Intra-task Consistency: The model requires that all views of a single task produce the same classification function ($g_i$).
  2. Inter-task Consistency: For views shared across different tasks, the model imposes a similarity constraint on the feature-level functions ($f_{ik}$).

The authors provide a theoretical foundation by showing that this graph-based objective function naturally connects to Reproducing Kernel Hilbert Spaces (RKHS). This connection reduces a complex graph problem to a standard supervised learning optimization. The proposed IteM2 Algorithm iteratively updates the classification functions for examples and the functions for features, scaling efficiently and converging to an optimal solution.

Key Insights

  • Feature Polarity: Unlike many models that only assign labels to examples, GraM2 defines a function $f(\cdot)$ on the features themselves. A positive value indicates that the feature supports a positive class (e.g., the word "excellent" in sentiment analysis), providing an added layer of interpretability.
  • The Power of Shared Views: The framework's most significant advantage is its ability to transfer knowledge through shared features. Even if Task B has zero labeled data, it can "learn" from Task A's labels because their shared views act as a conduit for knowledge transfer.

Experimental Results

The framework was evaluated on three distinct real-world datasets:

  • 20 Newsgroups: Text classification where tasks share a common vocabulary but have unique, task-specific vocabularies.
  • WebKB: Classifying university webpages into "course" or "non-course" across 4 universities (tasks) using page text, anchor text, and titles (views).
  • ECML 2006 (Spam Detection): Classifying emails as spam or ham for multiple users, using a common vocabulary as a shared view and user-specific folders/contacts as task-specific views.
Classification error on the 20 Newsgroups dataset demonstrating IteM2's superiority, especially with scarce labeled data.

Superiority in Data Scarcity

The $M^2TV$ framework consistently outperformed competitors like SVM-2K (which only uses multi-view learning) and SMTL/CASO (which only use multi-task learning). The performance gap was most dramatic when the number of labeled examples per task was very low (e.g., under 20 examples).

Robustness and Scalability

The parameter $b$, which controls how much tasks should influence each other, proved to be highly robust. Even a small amount of cross-task information significantly boosted performance compared to treating tasks entirely independently ($b=0$). Furthermore, theoretical analysis showed the algorithm's time complexity is linear with respect to the number of examples and features, making it highly viable for large-scale applications.

Limitations and Future Work

While powerful, the current framework assumes features are non-negative (such as word counts or pixel intensities). Extending the model to handle negative feature values remains a primary direction for future research. Additionally, the views in this study were pre-defined by the dataset structure; future work may explore automatically identifying "latent views" within a single high-dimensional feature set. Finally, while the model is robust, it relies on hyperparameters that may require cross-validation in highly sensitive applications.

Summary updated

Abstract

Many real-world problems exhibit dualheterogeneity. A single learning task might have features in multiple views (i.e., feature heterogeneity); multiple learning tasks might be related with each other through one or more shared views (i.e., task heterogeneity). Existing multi-task learning or multi-view learning algorithms only capture one type of heterogeneity. In this paper, we introduce Multi-Task Multi-View (M 2 T V) learning for such complicated learning problems with both feature heterogeneity and task heterogeneity. We propose a graph-based framework (GraM 2) to take full advantage of the dual-heterogeneous nature. Our framework has a natural connection to Reproducing Kernel Hilbert Space (RKHS). Furthermore, we propose an iterative algorithm (IteM 2) for GraM 2 framework, and analyze its optimality, convergence and time complexity. Experimental results on various real data sets demonstrate its effectiveness. 1.
BibTeX
@inproceedings{he2011graphbased,
  author = {Jingrui He and Rick Lawrence},
  title = {{A Graphbased Framework for Multi-Task Multi-View Learning}},
  booktitle = {International Conference on Machine Learning},
  year = {2011},
  url = {https://openreview.net/forum?id=Hy-5ri-dWB},
}
Figure