Job recommendation system

Abstract

Recommender systems are being used in all types of web based services. As the amount of data available in each service is huge it is not possible to search all the relevant and useful information. So for this purpose we need a recommender system, as it will take into consideration the user preferences for recommending the most relevant items. Many different recommendation algorithms have been used for this purpose. Collaborative filtering, matrix factorization techniques, content based methods have been extensively researched upon. Also significant amount of work has been done by using deep learning approaches in the existing recommendation frameworks. However these methods have suffered from the problem of overspecialization, also these approaches have not been very successful in the cold start scenario. The methods also do not effectively capture the change in user preferences. Recent developments in the field of deep reinforcement learning in games have shown to outperform the human players. This technique has largely remained unexplored in the field of recommendation. Reinforcement learning can very effectively handle the cold start issue as it is supposed to work in unknown state as well. Also as it is continuously learning from its interaction with the environment, so it can be used for capturing the changing user preferences. Based on which we can handle the recommendations which is personalized as per the user preferences. In this work, we describe a reinforcement learning based job recommendation system in which we model the recommendation problem as a Markov Decision Process(MDP) which is a 5- tuple (S, A, P, R, g) including states, actions, reward, a transition probability, and a discount factor. Reinforcement Learning is concerned with how the recommender agent takes actions in an environment so as to maximize the cumulative reward. We have also implemented, the collaborative filtering approach proposed in [7], the baseline approach suggested in ACM Recommender Systems Challenge 2017 [1] and the content-based approach proposed in [13]. We have finally shown that reinforcement learning method outperforms the other three.