The reinforcing influence of recommendations on global diversification

📝 Abstract

Recommender systems are promising ways to filter the overabundant information in modern society. Their algorithms help individuals to explore decent items, but it is unclear how they allocate popularity among items. In this paper, we simulate successive recommendations and measure their influence on the dispersion of item popularity by Gini coefficient. Our result indicates that local diffusion and collaborative filtering reinforce the popularity of hot items, widening the popularity dispersion. On the other hand, the heat conduction algorithm increases the popularity of the niche items and generates smaller dispersion of item popularity. Simulations are compared to mean-field predictions. Our results suggest that recommender systems have reinforcing influence on global diversification.

💡 Analysis

Recommender systems are promising ways to filter the overabundant information in modern society. Their algorithms help individuals to explore decent items, but it is unclear how they allocate popularity among items. In this paper, we simulate successive recommendations and measure their influence on the dispersion of item popularity by Gini coefficient. Our result indicates that local diffusion and collaborative filtering reinforce the popularity of hot items, widening the popularity dispersion. On the other hand, the heat conduction algorithm increases the popularity of the niche items and generates smaller dispersion of item popularity. Simulations are compared to mean-field predictions. Our results suggest that recommender systems have reinforcing influence on global diversification.

📄 Content

arXiv:1106.0330v1 [physics.soc-ph] 1 Jun 2011 The reinforcing influence of recommendations on global diversification An Zeng1, Chi Ho Yeung1, Mingsheng Shang2, Yi-Cheng Zhang1,2∗ 1Department of Physics, University of Fribourg, Chemin du Mus´ee 3, CH-1700 Fribourg, Switzerland 2Web Sciences Center, School of Computer Science and Engineering, University of Electronic Science and Technology of China Chengdu 610054, P. R. China (Dated: June 2, 2022) Recommender systems are promising ways to filter the overabundant information in modern society. Their algorithms help individuals to explore decent items, but it is unclear how they allocate popularity among items. In this paper, we simulate successive recommendations and measure their influence on the dispersion of item popularity by Gini coefficient. Our result indicates that local diffusion and collaborative filtering reinforce the popularity of hot items, widening the popularity dispersion. On the other hand, the heat conduction algorithm increases the popularity of the niche items and generates smaller dispersion of item popularity. Simulations are compared to mean-field predictions. Our results suggest that recommender systems have reinforcing influence on global diversification. PACS numbers: 89.75.-k, 89.65.-s, 89.20.Ff I. INTRODUCTION. Due to the rapid expanding of the internet, we are overloaded by the unlimited information on the World Wide Web [1]. For instance, one has to choose among millions of candidate commodities to shop online. Com- prehensive exploration is infeasible [2]. As a result, var- ious recommendation approaches have been proposed to help filtering the relevant information [3, 4]. For instance, the popularity-based recommendations (PR), which rec- ommend the most popular items to users, are commonly adopted in online recommender systems. However, such recommendations are not personalized such that identi- cal items are recommended for individuals with far dif- ferent taste. By comparison, the collaborative filtering (CF) makes use of collective data of individual prefer- ence and provides personalized recommendations [5, 6]. So far, CF has been successfully applied to many online applications. Recently, recommendation algorithms have been pro- posed from a physics perspective [7, 8]. For instance, diffusion is applied on the user-item bipartite networks to explore items of potential interest for a user. This mass diffusion (MD) algorithm is shown to outperform CF in the recommendation accuracy [7]. However, a sim- ilar problem as observed in PR is found in MD: diffusion- based recommendations are biased to popular items even individual preferences are considered. In fact, a good recommendation algorithm should recommend items of personal interest and at the same time maximize the di- versity of choices. An alternative approach based on the heat conduction (HC) on the user-item graphs is thus introduced [8]. This method provides users with many novel items and leads to diverse recommendation results among users. How- ∗yi-cheng.zhang@unifr.ch ever, HC has low accuracy compared with MD. The para- dox is eventually solved by combining MD with HC in a hybrid algorithm [9], which can be well-tuned to obtain significant improvement in both recommendation accu- racy and item diversity. Though they are helpful in filtering information, rec- ommendation algorithms may impose reinforcing influ- ence on the system, by guidance to one’s choices which influence subsequent recommendations and hence choices of others. The influence is amplified with successive rec- ommendations. We note that such perspective is em- ployed to explain the evolution movie popularity [10, 11], which yields consistent predictions compared with ob- served data. It is thus interesting to examine such in- fluence on recommender systems. Unlike most existing works which are devoted to improving recommendation accuracy [6], our present study presents a physics per- spective and utilizes microscopic interactions to explain and predict macroscopic behaviors of recommender sys- tems [12, 13]. In this paper, we use the Gini coefficient to measure the dispersion in item popularity [14]. We note that a small dispersion implies similar popularity among items, and hence diverse recommendations for users. We consider various conventional algorithms including the popularity- based, the collaborative filtering, the mass diffusion and the heat conduction algorithms. We focus on the physical aspects and study numerically and theoretically the rein- forcing influence of recommendations on the dispersion of item popularity. The result indicates that MD and CF reinforce the popularity of popular items, as similar to PR. On the other hand, the heat conduction algorithm increases the popularity of the niche items and generates smaller dispersion in item popularity. Our results suggest that recommender systems have reinforcing influence on global diversification. 2 1900 1950 1925 1975 2000 0 0.2 0.4 0.6 0.8 (a) APS Citation

This content is AI-processed based on ArXiv data.