[Paper Review] A Consensus-Focused Group Recommender System

Stratis Ioannidis, S. Muthukrishnan+1 more|arXiv (Cornell University)|Dec 26, 2013

Recommender Systems and Techniques11 references3 citations

TL;DR

This paper proposes a consensus-focused group recommender system that improves group event recommendations by modeling social influence through asynchronous voting and feedback. It introduces a decision cascading model to estimate how individual votes are influenced by prior group decisions, achieving significantly better prediction accuracy (AUC of 0.85 vs. 0.61) than individual preference-only models.

ABSTRACT

In many cases, recommendations are consumed by groups of users rather than individuals. In this paper, we present a system which recommends social events to groups. The system helps groups to organize a joint activity and collectively select which activity to perform among several possible options. We also facilitate the consensus making, following the principle of group consensus decision making. Our system allows users to asynchronously vote, add and comment on alternatives. We observe social influence within groups through post-recommendation feedback during the group decision making process. We propose a decision cascading model and estimate such social influence, which can be used to improve the performance of group recommendation. We conduct experiments to measure the prediction performance of our model. The result shows that the model achieves better results than that of independent decision making model.

Motivation & Objective

• To address the gap in group recommender systems that overlook post-recommendation feedback during group decision-making.
• To model social influence in group consensus processes where individual decisions are affected by others’ prior votes.
• To improve group recommendation accuracy by incorporating observed feedback and voting dynamics into a decision cascading model.
• To design a system that supports asynchronous voting, commenting, and dynamic recommendation updates for group events.
• To evaluate the impact of social influence on prediction performance using real-world group event data.

Proposed method

• The system uses a decision cascading model where each user's vote depends on their individual preference and the weighted influence of prior voters.
• The model uses a probabilistic formulation: Pr(y_u,i=1) = 1 - (1-p(u|i)) * ∏(1-p(v|u)) for all v who voted before u.
• Pairwise influence p(v|u) and individual preference p(u|i) are estimated via maximum likelihood using a geometric programming formulation.
• The model treats individual preference as a known constant and estimates influence parameters from observed voting sequences.
• The system collects feedback through an interface supporting asynchronous voting, event creation, friend recommendations, and comment threads.
• Parameter estimation is performed using a geometric program with constraints ensuring valid probability bounds and binding inequalities.

Experimental results

Research questions

• RQ1 How does social influence during group decision-making affect individual voting behavior in group recommendations?
• RQ2 Can a decision cascading model that captures sequential voting behavior improve prediction accuracy over individual preference models?
• RQ3 To what extent does observed post-recommendation feedback enhance the performance of group recommender systems?
• RQ4 How does the inclusion of social influence impact the AUC and accuracy of predicting group consensus?
• RQ5 What is the effect of asynchronous voting and feedback on consensus formation in group event selection?

Key findings

• The influence-based decision cascading model achieved an AUC of 0.85 on the test set, significantly outperforming the baseline logistic regression model (AUC 0.61).
• The model showed a higher true positive rate (0.77 vs. 0.52) and better accuracy (0.70 vs. 0.65) than the baseline, indicating improved prediction of positive votes.
• The false positive rate was higher (0.60 vs. 0.35), likely due to the presence of undecided users who did not provide a 'like' or 'dislike' vote.
• The system collected 4,398 valid votes across 277 group events with average group size of 2, 4, or 8 members, and 19 individuals participated on average in 54 events.
• The model’s performance improvement was consistent across different group sizes, demonstrating robustness to group dynamics.
• The geometric programming formulation enabled efficient and stable parameter estimation for the influence and preference parameters.