Multi Task Learning: A Survey and Future Directions

Multi-task learning (MTL) is a problem that must be applied in modern recommendation systems and is just as difficult. In the recent e-commerce advertising market, it is necessary to be able to predict not only the probability of users clicking, but also the probability of conversion and purchase. By predicting multi-task, it is possible to increase the accuracy of each task and optimize advertisements for various goals of advertisers. Traditional conversion rate (CVR) prediction models have difficulty learning because the number of conversions is too small compared to the total number of impressions. This problem is called a data sparsity (DS) problem. Another problem is that CVR models trained with samples of clicked impressions infer on samples of all impressions. This problem is called a sample selection bias (SSB) problem. This paper is a summary of the various solutions and current limitations and further directions about solving sample selection bias problem and data sparsity problem.

I. Introduction

Recently, the importance of MTL in recommendation systems has been increasing. MTL means learning multiple tasks together within the same model rather than learning them individually. Of course, it is better to learn individually if the tasks are completely independent, but it is also a good approach to share some parameters when similar tasks or when the goals of the tasks are ordered. Multi-task learning can be applied to a wide range of applications, real-world applications, NLP, audio, computer vision, recommendation systems, etc. However, the most useful place among them is recommendation systems. Why do recommendations systems need MTL?