Scalable Deep Q-Learning for Session-Based Slate Recommendation

What is it about?

This work is related to scaling application of slate recommendation using Reinforcement Learning paradigm. In this study we investigate the SlateQ algorithm proposed by google to be scalable while using in the production level. The idea is how to reduce the number of iterations of the Q functions which learns the Q-values of the slate.

Featured Image

Why is it important?

It is important because whenever we talk of session based recommendation we try to capture long term user satisfaction. In this case using RL becomes intuitive. But then in the RL setting having action as slate increases the action space combinatorially on addition of an item. Although SlateQ breaks this problem to a linear scale by decomposing the value of the slate to the value of the item it still has to iterate the costly Q-function proportional to the number of items during the serving time of the algorithm.

Perspectives