Performance analysis for a sequential recommendation algorithm

Abstract

In recent years, recommender systems have become a popular topic in research and many applications have been developed. Among the many recommendation tasks, next item recommendation is a task that predicts what items users will interact with in the next time. However, most systems give the recommendation based on users’ general preference, missing the opportunity to recommend items based on users’ sequential pattern. The authors in [1] aims to use convolutional sequence embedding recommendation (Caser) model to solve this problem. The objective of this project is to analyse multiple benchmark datasets to observe the performance and accuracy of this model. Top-N sequential recommendation recommends a sequence of items instead of a set of items. It captures a sequential pattern where next item or action is more likely depending on user’s recent actions. Sequential pattern represents user’s short term and dynamic behaviours that more recent items in a sequence affect the next item more, whereas general preference refers to user’s static and long-term behaviours. The Caser model solves two limitations of previous work: (1) Fail to model union- level sequential patterns; (2) Fail to allow skip behaviours. Specifically, Caser model leverages Convolutional Neural Network in capturing local features for image recognition and natural language processing. It has the following advantages: (1) Caser is able to capture sequential patterns at point-level, union level and of skip behaviours; (2) Caser considers both general preference and sequential pattern of users; (3) Caser has better performance compared to the state-of-the-art methods in top-N sequential recommendation topic. In this project, some real-life datasets are used to make analysis from different perspectives. In addition to reproducing the experiments in the paper [1], additional steps are made when processing the data in order to observe the performance of Caser model by categories, seasons, user training instances, and time interval. Based on the observed results, further improvement and types of dataset of the Caser model can be summarized.