Hybrid Techniques to Address Cold Start Problems for People to People Recommendation in Social Networks Yang Sok Kim, Alfred Krzywicki, Wayne Wobcke, Ashesh Mahidadia, Paul Compton, Xiongcai Cai and Michael Bain School of Computer Science and Engineering University of New South Wales Sydney NSW 2052, Australia {yskim,alfredk,wobcke,ashesh,compton,xcai,mike}@cse.unsw.edu.au Abstract. We investigate several hybrid approaches to suggesting matches in people to people social recommender systems, paying particular attention to cold start problems, problems of generating recommendations for new users or users without successful interactions. In previous work we showed that interaction- based collaborative filtering (IBCF) works well in this domain, although this ap- proach cannot generate recommendations for new users, whereas a system based on rules constructed using subgroup interaction patterns can generate recommen- dations for new users, but does not perform as effectively for existing users. We propose three hybrid recommenders based on user similarity and two content- boosted recommenders used in conjunction with interaction-based collaborative filtering, and show experimentally that the best hybrid and content-boosted rec- ommenders improve on the IBCF method (when considering user success rates) yet cover almost the whole user base, including new and previously unsuccess- ful users, thus addressing cold start problems in this domain. The best content- boosted method improves user success rates more than the best hybrid method over various “cold start” subgroups, but is less computationally efficient overall. 1 Introduction In recent work we have investigated people to people recommendation with particu- lar application to online dating sites. This problem is particularly interesting since it requires a reciprocal recommender [8], one where “items” are users and hence have their own preferences that must be taken into account when recommending them to others. Reciprocal recommenders must consider user “taste” (how they view potential matches) and “attractiveness” (how they are viewed by others), Cai et al. [3]. In partic- ular, in contrast to typical product recommender systems, it is not generally appropriate to recommend popular users, since popular users are typically highly attractive, hence less likely to respond positively to an expression of interest from the average user. Cold start problems are particularly acute in people to people recommendation in online dating since the user base is highly dynamic, so there are many new users at any point in time. In addition, our previous work has shown that the best information to use for recommendation is the set of positive interactions between users, those interactions initiated by one user where the other replies positively (whether a reply is positive is