Combining Task Predictors via Enhancing Joint Predictability
This work addresses predictor combination for machine learning practitioners, offering an incremental improvement over existing approaches by better leveraging dependencies among reference predictors.
The paper tackles the problem of improving a target predictor by combining it with reference predictors from related tasks, developing a Bayesian algorithm that jointly assesses all references rather than using pairwise relationships. The method demonstrated significant performance gains across seven real-world datasets in visual attribute ranking and multi-class classification.
Predictor combination aims to improve a (target) predictor of a learning task based on the (reference) predictors of potentially relevant tasks, without having access to the internals of individual predictors. We present a new predictor combination algorithm that improves the target by i) measuring the relevance of references based on their capabilities in predicting the target, and ii) strengthening such estimated relevance. Unlike existing predictor combination approaches that only exploit pairwise relationships between the target and each reference, and thereby ignore potentially useful dependence among references, our algorithm jointly assesses the relevance of all references by adopting a Bayesian framework. This also offers a rigorous way to automatically select only relevant references. Based on experiments on seven real-world datasets from visual attribute ranking and multi-class classification scenarios, we demonstrate that our algorithm offers a significant performance gain and broadens the application range of existing predictor combination approaches.