Sunday, March 28, 2010

Recommendation Engines

paper: "Recommender System for Online Dating Service"

"User-User and Item-Item collaborative filtering recommenders significantly outperform global algorithms that are currently used by dating sites [offering only Browsing / Searching Options, Powerful Searching Engine but not Compatibility Matching Algorithms]."
"A blind experiment with real users [at a proprietary site named ColFi - exclusively designed for the experiment - where 111 users rated 150 photo-profiles, then two recommendation lists of top 10 profiles were generated] also confirmed that users prefer collaborative filtering based recommendations to global popularity recommendations [of 2 Czech online dating sites: ChceteMe (no longer exists now) and LibimSeTi]."
"Recommendations can be further improved by hybrid algorithms. These algorithms are combining the collaborative filtering approach with content information. Another problem specific to dating is that A_likes_B does not imply B_likes_A. Therefore each user should be probably presented with recommendations of such users, who are also interested in him/her. There is a need for reciprocal matching algorithms."
"User interface may introduce bias in the sense that users instead of providing their personal preference try to guess the global preference. This reduces the usefulness of ratings provided."

By my own experience I know that the proprietary Bidirectional Recommendation Engine (Matching based on Self-Reported Data by personal preferences & likes and dislikes) actually in use at Match ("the Daily5"at USA site) is in the range of 3 to 4 persons recommended per 1,000 persons screened, in other words any member receives on average 3 or 4 prospective mates as recommended for dating purposes per 1,000 (one thousand) members screened in Match’s big database
Many online dating sites had been using Behavioural Bidirectional Recommendation Engines for years, like PlentyOfFish, and they could not outperform compatibility Matching Methods based on personality profiling.

Recommendation Engines do not take into account the new discovery uncovered by Eastwick and Finkel 2008; also Kurzban and Weeden, 2007; Todd, Penke, Fasolo, and Lenton, 2007 who found that people often report partner preferences that are not compatible with their choices in real life.

No comments:

Post a Comment