Solving the accuracy-diversity dilemma via directed random walks

TL;DR

This paper introduces a directed random walk approach in collaborative filtering that enhances recommendation diversity without sacrificing accuracy by adjusting the similarity measurement direction.

Contribution

It reveals the impact of random walk direction on user similarity and proposes a new algorithm that improves both accuracy and diversity in recommendations.

Findings

Outperforms state-of-the-art CF methods in accuracy and diversity

Reveals the importance of random walk direction in similarity measurement

Achieves accurate and diverse recommendations without context-specific info

Abstract

Random walks have been successfully used to measure user or object similarities in collaborative filtering (CF) recommender systems, which is of high accuracy but low diversity. A key challenge of CF system is that the reliably accurate results are obtained with the help of peers' recommendation, but the most useful individual recommendations are hard to be found among diverse niche objects. In this paper we investigate the direction effect of the random walk on user similarity measurements and find that the user similarity, calculated by directed random walks, is reverse to the initial node's degree. Since the ratio of small-degree users to large-degree users is very large in real data sets, the large-degree users' selections are recommended extensively by traditional CF algorithms. By tuning the user similarity direction from neighbors to the target user, we introduce a new algorithm specifically to address the challenge of diversity of CF and show how it can be used to solve the accuracy-diversity dilemma. Without relying on any context-specific information, we are able to obtain accurate and diverse recommendations, which outperforms the state-of-the-art CF methods. This work suggests that the random walk direction is an important factor to improve the personalized recommendation performance.