Group modeling, recommendation and evaluation in collaborative filtering group-based recommender systems
As the field of recommender systems has grown, more and more attention has been focused on the need forsystems that provide and tailor recommendations to groups of users, as opposed to individuals. We have identified open issues in group-based recommender system along three different dimensions: (1) the group recommendation technique employed to generate recommendations, (2) the group modeling strategy for generating group recommendations, and (3) the group evaluation metrics and procedures used for assessment of the recommendations.Group recommendation presents significant challenges in evolving best practice approaches to group modeling, recommendation, and evaluation. Early research provided more limited, illustrative evaluations for group recommender approaches, but recent work has been exploring more comprehensive evaluative techniques. The main research problem we address is how to improve the prediction accuracy in group-based recommender systems employing a memory-based collaborative filtering technique. We break down this question along the dimensions we identified in group-based systems in this:- How to implement a principled approach to evaluate the prediction accuracy in group-based recommender systems using datasets of individual users' preferences?- Does the evaluation approach affect the results for the prediction accuracy? - Does rating normalization increase the prediction accuracy for the group? - Does incorporating the group-context in the neighborhood selection increase the prediction accuracy? - Does a hybrid group-based strategy increase the prediction accuracy? - Does incorporating the group context in the group modeling increase the prediction accuracy?Group-based recommender systems introduce extra overhead in recruiting groups of users to cooperate toward a common goal at the same time. To overcome this limitation researchers have utilized publicly available large-scale datasets derived from individual-based recommender systems by creating synthesized groups and using them in offline evaluations. The need for a principled approach to evaluations utilizing this technique remains an open issueWe address this problem by developing a group testing framework to evaluate group-based recommenders in this context using data sets from traditional, single-user, collaborative filtering systems. We utilize this group testing framework in conducting comparatively large-scale evaluations of our proposed approaches along the problem dimensions of group-based recommenders. We first show the feasibility of an exact overlap constraint for evaluation. We then compare the prediction accuracy of some of the most commonly adopted group modeling strategies and compare to previous research that utilized synthesized groups with an average ground truth and report on the discrepancy of evaluation results between the two approaches. We also show that the choice modeling the ground truth for synthesized groups affects the evaluation results by comparing the prediction accuracy using various models. To be able to compare any new approach to previous research in this domain we need to utilize the same baseline they have used for evaluation where they model the actual group preference as an average of the individual group members preferences. We utilize the testing framework to create groups and the training and testing datasets for those groups. Since different users provide preferences on different scales research in individual based recommender has shown that normalizing the ratings in the prediction calculation increased the accuracy of predictions for users. This led us to explore the affect of rating normalization on the prediction accuracy for groups of users. We show the conditions where rating normalization would be beneficial for the group. Previous research in this context has mainly adopted recommendation techniques validated for single user recommender systems with out considering the group context in the recommendation technique. We believe that utilizing the group context in the recommendation technique and group modeling would result in predictions with higher accuracy for the group. We evaluate a neighborhood model incorporating the group context in the neighborhood selection with a weighted approach based on neighborhood overlap. Our results show a higher prediction accuracy for the group is realized with this finer-grained neighborhood weighting model based on the group context is applied. We also evaluate a hybrid recommendation technique that incorporates the two group-based strategies. Our results show an increase in the prediction accuracy for groups in general. We also evaluate the performance of this approach along different group contexts and identify that this model is more advantageous for groups with highly similar group members terms of prediction accuracy. To incorporate the group context in the group modeling previous research proposed incorporating a disagreement component in the model using the predicted ratings for that item. This led us to explore ways to base the disagreement model using a concrete preference rather than a predicted one since we believe this would increase the prediction accuracy for the group. Our proposed disagreement model based on item similarity revealed that it might not be feasible to model disagreement by incorporating similar items, rated by all the group members, to the target item in the prediction calculation. Continuing to think along the lines of the group context as a first-class element of the group model we investigated a Case-based Reasoning approach where cases are matched based on the group context. Our approach is based on group-to-group similarity rather than user-to-user case matching where we retrieve whole previous groups as the starting point for predictions. We confirmed the potential benefit for integrating whole-group retrieval CBR approaches into group recommendation across different case-base and group conditions. We also demonstrated that the benefits of a CBR approach may be found even in straightforward implementations, showing the potential for a broad range of deployments and investigation in the space.