Exploring Optimizations in Multi-Task Recommendation via Manipulating Auxiliary Gradients

Abstract

Auxiliary Task Learning has proven to be an effective way to transfer knowledge between tasks. This is the case in many personalized recommendation scenarios, where many auxiliary tasks can boost the performance of the primary task via a shared layer in a multi-task network. However, we often encounter a gradient imbalance issue when updating the parameters of the neural network. For example, the gradient magnitudes for an auxiliary task could vary significantly compared to the target task which leads to imbalanced information transfer between tasks. This could cause a particular auxiliary task to over or under influence the output of the target task. Multiple works have been done to resolve the gradient imbalance issues. Some methods include directly manipulating the gradients with respect to the shared parameters in order to balance gradient magnitudes. Other approaches identify conflicting gradient direction between tasks and dynamically scale auxiliary gradients pending on their direction agreement with the target task. This thesis seeks to understand the baseline benchmark techniques that exist and to combine multiple techniques to improve recommendation accuracy. Current methods of direct gradient manipulation are not combined as they have shown sub-optimal performance. In this work, we evaluate and compare these different methods on a large, public recommendation data set. We explore various ways to combine both methods with the aim of achieving boosted performance. In our analysis, we learn that naively combining gradient balancing methods with gradient direction based methods yield sub-optimal results; however, they outperform the baseline methods which suggests that more refined techniques could improve upon state-of-the-art performance.