"... Corpus of Face Video Clips with Affective and Personality Metadata"
"... The corpus presented here has been compiled for the needs of our research work of affective and personality based user modeling in recommender systems.
We had 52 students of a secondary school who participated in the experiment. The average age was 18.3 years (SD = 0.56). There were 15 males and 37 females.
4.1. Affective content based recommender system
Part of the presented corpus was used to develop and evaluate an affective user and item modeling approach in a content based recommender (CBR) system.
4.2. Personality based collaborative recommender system
Another part of the presented corpus was used in an experiment where we evaluated a novel
personality based user similarity measure for collabrative filtering.
A well known drawback of collaborative filtering methods is the new user problem. It occurs when a new user joins the system and the system has little or no knowledge on the user’s preferences.
As the user similarity measures rely on explicit ratings given by the users, when there are little ratings available, the algorithm for finding similar users tend to give bad choices. Consecutively, the predicted ratings have low correlation with real ratings. In order to alleviate the new user problem we introduced an initial questionnaire to assess the personality of each user.
We constructed a user similarity measure as an Euclidian distance in the big five personality space.
(papers: Emotive and Personality Parameters in Multimedia Recommender Systems and Personality based user similarity measure for a collaborative recommender system)
Statistical analysis showed that the proposed personality based user similarity measure yielded significantly better results than the rating based user similarity measure which makes it more suitable not only to alleviate the new user problem but also to use when the new user phase dies away.
4.3. Emotion detection from video clips
Part of our ongoing research work is the detection of emotion from face videoclips. Our goal is to develop a method for detecting emotions in users with two novel properties:
(i) the inclusion of personality parameters as features
(ii) detection in the valence-arousal-dominance (VAD) space (instead of the coarse space of basic
The 2 WEAK POINTS are:
The Big 5 to assess personality of users
the age of the sample, they are too young persons (personality not consolidated yet)
Have you seen the next generation of recommender systems include normative personality traits?
That is the only way to improve recommender systems, to include the personality traits of their users.
Have you seen they need to calculate personality similarity between users?
Have you seen there are different formulas to calculate similarity?
In case you did not notice, recommender systems are morphing to .......... compatibility matching engines!!!