Collaborative Filtering for people-to-people recommendation in online dating: Data analysis and user trial. Issue 76 (April 2015)

• Record Type: Journal Article
• Authors: Krzywicki, A.; Wobcke, W.; Kim, Y.S.; Cai, X.; Bain, M.; Mahidadia, A.; Compton, P.
• Is Part Of: International journal of human-computer studies. Issue 76(2015)
• Page Start: 50
• Page End: 66
• Publication Date: 2015-04
• Subjects: Recommender systems -- Machine learning -- User study; Human-machine systems -- Periodicals; Systems engineering -- Periodicals; Human engineering -- Periodicals; Human engineering; Human-machine systems; Systems engineering; Periodicals; Electronic journals; 004.019
• DOI: 10.1016/j.ijhcs.2014.12.003
• Languages: English
• ISSNs: 1071-5819

Abstract

Abstract: A common perception is that online dating systems "match" people on the basis of profiles containing demographic and psychographic information and/or user interests. In contrast, product recommender systems are typically based on Collaborative Filtering, suggesting purchases not based on "content" but on the purchases of "similar" users. In this paper, we study Collaborative Filtering for people-to-people recommendation in online dating, comparing this approach to a baseline Profile Matching method. Initial data analysis highlights the problem of over-recommending popular users, a standard problem for Collaborative Filtering applied to product recommendation, but more acute in people-to-people recommendation. We address this problem with a two-stage recommender process that employs a Decision Tree derived from interactions data as a "critic" to re-rank candidates generated by Collaborative Filtering. Our baseline Profile Matching method dynamically chooses, for each user, attributes that contribute most significantly to successful interactions with candidates having the best matching attribute value. The key evaluation metric is success rate improvement, the increase in the chance of a user having a successful interaction when acting on recommendations. Our methods were first evaluated on historical data from a large online dating site and then trialled live over a 9 week period providing recommendations via e-mail to a large number of users. The trial confirmed the consistency of the analysis on historical data and the ability of our Collaborative Filtering method to generate suitable candidates over an extended period. Moreover, the Collaborative Filtering method gives a higher success rate improvement than Profile Matching. Abstract : Author-Highlights: Collaborative Filtering is feasible for people-to-people recommendation. Collaborative Filtering improves users׳ success rates more than Profile Matching. A general method is given to avoid over-recommending highly popular users. Collaborative Filtering and Profile Matching have been validated in a live trial. Performance of methods in the live trial setting is consistent over time.