Rethinking Conversational Recommendations: Is Decision Tree All You Need?

Conversation Recommendation System (CRS) is becoming a topic of interest in information retrieval. Traditional recommender systems rely on historical user interaction records and do not consider the users’ current preferences. Alternatively, conversational recommender systems dynamically obtain the users’ preferences via multi-turn question and answer. The existing works in conversational recommender systems mostly rely on reinforcement policy learning. Also, the current methods use pre-trained user embeddings for recommendations. However, this approach becomes ineffective when a new user enters the system (the cold start problem). This reduces the effectiveness of CRS for cold-start
users, even though it is the main motivation for CRS in the first place.

In this study, we propose two things. Firstly, we challenge the necessity of using reinforcement learning in CRS. There are 3 main challenges in multi-turn CRS: 1) What questions to ask 2) When to recommend 3) How to improve when a user rejects recommendations. We show that supervised machine learning is sufficient to face these 3 challenges. Our supervised learning model is a simple decision tree-based model, namely FACT-CRS (stands for Factorization Tree-based Conversational Recommender System).
Secondly, we show that by taking the user-item interaction into account, we can learn a rule-based method that is effective for users who are new to the system. Extensive experiments on two benchmark CRS datasets (LastFM and Yelp) show that supervised learning is, in fact, sufficient to face the challenges in multi-turn CRS. On the LastFM dataset, our model is 27.42% more successful in recommending the desired item than the best of the baselines. Our model also achieves a 2.05% better success rate than the baselines on the Yelp dataset, which includes more than 1M interactions. Empirical results also suggest that our proposed model can successfully recommend target items to users by asking a fewer number of questions.