An empirical study of natural noise management in group recommendation systems

Highlights

• Improving the performance of Group Recommender Systems

• Management of Natural Noise in Group Recommender Systems

• Deployment of Group Recommender Systems with Natural Noise Management

Abstract

Group recommender systems (GRSs) filter relevant items to groups of users in overloaded search spaces using information about their preferences. When the feedback is explicitly given by the users, inconsistencies may be introduced due to various factors, known as natural noise. Previous research on individual recommendation has demonstrated that natural noise negatively influences the recommendation accuracy, whilst it improves when noise is managed. GRSs also employ explicit ratings given by several users as ground truth, hence the recommendation process is also affected by natural noise. However, the natural noise problem has not been addressed on GRSs. The aim of this paper is to develop and test a model to diminish its negative effect in GRSs. A case study will evaluate the results of different approaches, showing that managing the natural noise at different rating levels reduces prediction error. Eventually, the deployment of a GRS with natural noise management is analysed.

Introduction

A recommender system (RS) focuses on providing personalised access to items (or information in general) in an overloaded search space. With this purpose in mind, RSs explore users' interests to find out which items are the most suitable. RSs have been successfully applied to support users trying to overcome the information overload problem in several domains [15], such as e-commerce [7], financial investment [20], e-learning [16], [28], e-government [14], and e-tourism [22], among others.

RSs employ different approaches depending on the information they rely on, but the most widespread ones are content-based and collaborative filtering approaches [1]. A content-based recommender system (CBRS) relies on users' preferences and information about items, such as features or a textual description, to be able to make recommendations [10]. On the other hand, collaborative filtering recommender systems (CFRSs) rely on users' preferences to be able to make recommendations [21]. Therefore, CFRSs are able to generate effective recommendations using only users' preferences (rating values) [25] hence the quality of the ratings influences the quality of the recommendations.

The ratings can be gathered implicitly [6] or explicitly, this article focuses on the latter case. Prior research has explored how noisy preferences intentionally inserted by users affect RSs [8], [13], so-called malicious noise. However, the noisy ratings introduced unintentionally by users, so-called natural noise, has recently attracted the attention of researchers. Several proposals investigate the detection and correction of such natural noise. Some proposals perform these tasks by exploiting the items' attributes [24], either by taking advantage of user's interaction [3], or by using knowledge extracted from the ratings themselves [29]. The main benefit of these proposals is their positive impact on the recommendations.

So far, natural noise has been studied only in RSs for individuals. However, Group Recommender Systems (GRSs) [11] play an important role in many social activities that require recommendations to be delivered to a group of users, such as watching TV with family, sightseeing with others, or going to the cinema with friends.

GRS approaches extend individual RS for recommending to groups by aggregating user individual information [18]. Therefore, GRSs use explicit ratings, and natural noise is also present biasing the group recommendation. Consequently, its management may play an important role in the quality of the group recommendations. This paper aims at researching the natural noise management (NNM) in GRS to study its influence in group recommendation.

In group recommendation, unlike traditional recommendation scenarios, the ratings dataset has different levels of information from the group viewpoint. Therefore, the direct application of NNM methodologies is not adequate in this context [29]. This fact makes necessary to develop new methodologies that enable the application of NNM across the different levels of information in GRSs. This is the objective of the current paper, which will be reached by the study of different hypotheses.

Taking into account that group members' ratings are key data in GRSs for computing the recommendations provided to the group by their aggregation [11], their quality influences the recommendation accuracy. It leads to the formulation of the first hypothesis:

• H1. NNM using only the group ratings would improve the group recommendation.

On the other hand, NNM for individual RSs has shown clear improvements [29]. Due to the fact that, GRSs approaches are supported by individual RSs [11], a new hypothesis is raised:

• H2. NNM in the entire ratings database, disregarding the groups, would improve the group recommendation.

Eventually, if both levels (group ratings, all users ratings) are considered, some ratings tagged as noisy at the group level could be not noisy at the global level, and vice versa. Therefore, the NNM at both levels could lead to a better recommendation accuracy. Hence, H3 is formulated as follows:

• H3. managing natural noise in the entire ratings database and, after that, adding a second step that manages natural noise in the group ratings, would improve the results as compared to a single step of NNM.

Different approaches which apply these NNM processes to GRSs are presented and a case study to verify the hypotheses is performed.

The remaining of the paper is organised as follows. Section 2 provides a background on CFRSs, GRSs, and natural noise. Section 3 presents four methods of natural noise management for GRSs. Section 4 develops the case study and analyses the results of the proposals, leading to the acceptance of H2 and H3. Finally, Section 5 points out the conclusions and discusses future research.