Improving Reachability and Navigability in Recommender Systems

TL;DR

This paper analyzes recommendation networks from a graph perspective, evaluating their ability to support navigation and exploration, and proposes methods to enhance their reachability and navigability.

Contribution

It introduces a comprehensive approach to assess and improve the navigability of recommendation networks, extending evaluation from single-click to multi-click scenarios.

Findings

Standard algorithms yield poor navigability in recommendation networks.

Modifying recommendation strategies can significantly improve reachability.

The proposed evaluation framework supports arbitrary recommendation network analysis.

Abstract

In this paper, we investigate recommender systems from a network perspective and investigate recommendation networks, where nodes are items (e.g., movies) and edges are constructed from top-N recommendations (e.g., related movies). In particular, we focus on evaluating the reachability and navigability of recommendation networks and investigate the following questions: (i) How well do recommendation networks support navigation and exploratory search? (ii) What is the influence of parameters, in particular different recommendation algorithms and the number of recommendations shown, on reachability and navigability? and (iii) How can reachability and navigability be improved in these networks? We tackle these questions by first evaluating the reachability of recommendation networks by investigating their structural properties. Second, we evaluate navigability by simulating three different models of information seeking scenarios. We find that with standard algorithms, recommender systems are not well suited to navigation and exploration and propose methods to modify recommendations to improve this. Our work extends from one-click-based evaluations of recommender systems towards multi-click analysis (i.e., sequences of dependent clicks) and presents a general, comprehensive approach to evaluating navigability of arbitrary recommendation networks.