Reinforcement Learning Based Recommendation System - Part 1

Recommender System

Every day, we always face with the huge amount of information available on many means leading to information overloading problem, which make people feel difficult to make right decision.

When we surf a shopping website, we have to pass through many more items in the main page. The more items in the list, the harder it become to select among them. Understanding the demand, and with the development of many information platform such as YouTube, Amazon, Netflix or e-commerce, Recommendation System has been established and improved with the development of machine learning and artificial intelligent.

You can also check posts in series:

• Part 2: Reinforcement Learning in Recommendation system
• Part 3: Reinforcement Learning based Recommendation System

Recommendation System

Definition

Recommender System (RS) are software tool and algorithms that have been developed to help users find their interested items based on the prediction of their item preferences or their rating on the items. RS is established to deal with the “long tail problem” that when users just only have attention on the highlighted items in a specific domain, they left behind many items that may be useful for them and the number of those item is very huge. This situation happens frequently in supermarket or in e-commercial website.

Method

The method for recommender system is divided into three type as the taxonomy below: content-based methods, collaborative filtering methods and hybrid methods which is the combination of two previous method.

Content-based methods: These methods make the recommendations based on a description of the item, a profile of the user’s preferences and the relation between them. This method can be used to a large number of user because it doesn’t any data from other users and each recommendation is specific user. However, this method needs a huge knowledge about a domain (for example, information about rating of user for an items), which mean that we need to handle knowledge-crawling by hand.

Collaborative filtering methods: There are two types of this methods: memory-based and model-based. While the memory-based method is based on the similarities between users or items in the interaction observation (user-user or item-item), model-based method uses some algorithms such a Clustering algorithms, Matrix Factorization-based algorithm and Deep-Learning methods to learn the users/items behavior from the interaction matrix. Different with the above methods, this model no need a huge domain knowledge and it can help use discover more about their interest by recommending an item from a similar user interested with it. However, this model cannot deal with the new items, or cold-start problem, but there are some ways to address this problem or using content-based method. Other problem is that this system is hard to include some side-feature for items that can help the recommender system is more trustworthy.

Hybrid methods: The combination between Content-based methods, and collaborative filtering methods, which takes advantage of the advantages of each method to minimize the drawbacks of them.

Useful resources:

• Introduction to recommender systems | by Baptiste Rocca
• Recommender Systems with Python series | by Nikita Sharma