Wednesday, October 13, 2010

Personality-based Recommender System

Have you seen the next generation of recommender systems are going to include personality traits?

"Design and User Issues in Personality-based Recommender Systems"

"Recommender systems have emerged as an intelligent information filtering tool to help users effectively identify information items of interest from an overwhelming set of choices and provide personalized services. Studies show that personality influences human decision making process and interests. However, little research has ventured in incorporating it into recommender systems. ... The overall goal is to develop an efficient personality-based recommender system and to arrive at a series of design guidelines from the perspective of human computer interaction"
In that paper, the results on a proposed personality-based music recommender prototype is presented

[ That Personality-based Recommender System is useless at all for serious dating proposals.

That paper calculates personality similarity between users: ".. we treat users' personality characteristics as a vector as rating records. For each user u, his/her personality descriptor Pu is a n-dimension vector. Consequently, the similarity between two user u and v can be computed as the Pearson correlation coefficient of their personality descriptors..."

but think for a minute! If you use those Personality-based Recommender Systems for Online Dating Purposes, then they are ... guess ... Compatibility Matching Algorithms!

If they use the Big5 to assess personality and the Pearson correlation coefficient to calculate similarity, they are nothing new, the same stuff already available.

The Big 5 traits personality model is good for orientative purposes but not good enough for predictive purposes.

"Because the Big Five groups the more specific primary-level factors, feedback organized around the five Global Factor scales is more easily understood. For detailed feedback or predictive purposes, one should assess the more specific primary factors. Research has shown that more specific factors like the primary scales of the 16PF Questionnaire predict actual behavior better than the Big 5 Global Factors. For example, one extravert (a bold, fearless, high-energy type) may differ considerably from another (a sweet, warm, sensitive type), depending on the extraversion-related primary scale score patterns, so deeper analysis is typically warranted."
Extracted from the 16PF5 Manual

Moreover if a visual personality quizz is used to assess personality of users, like the one offered by VisualDNA or Dewey Color System, it adds a lot of distortion to the measurement. Those Personality-based Recommender Systems will perform worse than actual Compatibility Matching Algorithms!


"Design and User Issues in Personality-based Recommender Systems" is related to

"Using Personality Information in Collaborative Filtering for New Users"
(they) compare the performances of rating-based similarity (RBS), personality-based similarity (PBS) and their hybrid (RPBS) in different start-up settings
Figure 1 shows how RPBS and PBS are better models than RBS (in a sparse music dataset).

No comments:

Post a Comment