Introduction
Multi-Task Learning (MTL) aims to improve the performance of multiple related tasks by sharing information across them. Traditional MTL methods often operate in batch mode or assume fixed, pre-defined task relationships. However, in real-world streaming environments, task correlations are often unknown a priori and may even be negative. This paper introduces an Online Multi-Task Learning (OMTL) framework that adaptively learns both task weights and their underlying relationships on the fly. By utilizing Bregman divergence minimization, the model captures complex correlations, leading to improved accuracy and more efficient active learning.
Methodology
The proposed framework treats MTL as an alternating optimization problem, simultaneously updating task weight vectors ($\mathbf{w}$) and a task-relationship matrix ($\mathbf{A}$).
- Weight Update: Upon a prediction mistake, weights are updated using a Perceptron-like rule that incorporates the current estimate of task relationships, effectively "sharing" the update across related tasks.
- Relationship Update: The matrix $\mathbf{A}$ is updated to minimize its Bregman divergence from the previous iteration, constrained by new information from the weight updates. Two specific divergences are explored: LogDet Divergence (OMTLLog), which yields an inverse-based update, and von-Neumann Divergence (OMTLVon), which results in a matrix exponential-based update.
A critical component of the methodology is the "Priming Duration" (Epoch). To ensure stability, the model initially treats tasks as independent. Relationship learning is only activated after a specific data threshold is reached, preventing the model from being misled by early, noisy weight estimates.
Experimental Setup
The OMTL framework was evaluated on four diverse datasets:
- Synthetic: Designed with specific positive, negative, and zero correlations to test relationship discovery.
- 20newsgroups: Approximately 11,000 messages across 10 newsgroup categories.
- Sentiment: Amazon user reviews across 8 product domains (e.g., books, electronics).
- Spam: Email classification for 3 different users, where each user represents a task.
Results and Discussion
The OMTL variants, particularly OMTLLog and OMTLCov, consistently outperformed independent learning (IPL) and fixed-relationship models (CMTL).
Negative Correlation Discovery
A key advantage of OMTL is its ability to identify negative correlations. While previous methods like CMTL often assume positive relationships, OMTL correctly identified a -0.9 correlation in synthetic tests where CMTL incorrectly predicted a positive relationship.
Active Learning Efficiency
The framework was extended to an active learning setting, where learned task similarities guide the selection of informative examples for human labeling. This approach significantly reduced the labeling burden.
As shown in the results, OMTL variants achieved label reductions of 42% to 67% compared to standard baselines while maintaining or improving classification accuracy. The LogDet update (OMTLLog) also demonstrated superior numerical stability in streaming scenarios compared to the von-Neumann variant.
Conclusion and Future Work
The OMTL framework provides a robust method for dynamic task relationship learning in online settings. However, challenges remain, such as developing a data-driven approach to determine the optimal "priming duration" rather than relying on a fixed threshold. Additionally, future research could address the computational complexity of matrix updates for very large task sets and seek formal theoretical mistake-bound proofs for these adaptive update rules.
Summary updated