A Survey on Negative Transfer

Transfer learning (TL) utilizes data or knowledge from one more source domains to facilitate in a target domain. It is particularly useful when the domain has very few no labeled data, due annotation expense, privacy concerns, etc. Unfortunately, effectiveness of TL not always guaranteed. Negative transfer (NT), i.e., leveraging data/knowledge undesirably reduces performance domain, and been long-standing challenging problem TL. Various approaches have proposed literature address this issue. However, there does exist systematic survey. This paper fills gap, by first introducing definition NT its causes, reviewing over fifty representative for overcoming NT, which fall into three categories: similarity estimation, safe transfer, mitigation. Many areas, including computer vision, bioinformatics, natural language processing, recommender systems, robotics, that use mitigation strategies positive transfers, are also reviewed. Finally, we give guidelines on task construction baseline algorithms, benchmark existing NT-specific datasets, point out challenges future research directions. To ensure reproducibility, our code publicized at https://github.com/chamwen/NT-Benchmark.