Boosting Recommendation Systems through an Offline Machine Learning Evaluation Approach

Abstract

Nowadays, recommendation systems are widely deployed to suggest a variety of products and services for target users. Practical examples of recommendation systems that we daily encounter include social, educational, and political services such as academic courses, movies, travel, music, and news feed on the web. Current recommendation systems can be categorized into threefold: knowledge-based, collaborative filtering, and content-based systems. Knowledge-based recommendation systems are based on domain knowledge without user-specific data. Collaborative filtering provides recommendations based on existing user's historical record, which can be a problem when the user is new to the platform. Content-based recommendations build models based on a number of records on product transactions. In this paper, we introduce the three recommendation systems, then suggest a new strategy that integrates domain knowledge for new users into a machine learning based recommendation system for existing users. We also propose an offline/online evaluation strategy to reduce customer churn, which enables the deployment of accurate recommendations for both new and existing users. Using the movie Tweetings dataset, we implement each recommendation system and generate a boosted system which addresses the issues caused by current recommendation systems.