Music recommendation model for computer users based on personality, cognition, affect, and music preferences

This research aimed to build a model that is capable of recommending music that are appropriate for the users personality current academic emotion as well as music preference. This required investigating the effects of personality traits, using the OCEAN model, to music preference, and how these factors impacted the building of the model, to music preference, and how these factors impacted the building of the model. Furthermore, the proponents also studied how academic emotions and their valence influence music preference. During data gathering, the academic emotions of users were recorded using an annotation tool while the user was studying in a naturalistic setting. These emotions were studied according to the personality traits of the user. The corpus for the model consists of the primary features, namely personality, academic emotions, and valence acquired from a total of seventy-six (76) data subjects. A number of different models were built using the raw data set consisting of four hundred fifty-four (454) instances. Frustration instances were removed due to small amount of instances to allow for better accuracy of the model. The datasets were also divided by gender due to personality studies suggesting that personalities are very different between male and females. Thus, the divided dataset contained two hundred sixty-one (261) instances for male and one hundred seventy (170) instances for female. This divided dataset was balanced to provide equal learning instances to allow a less biased dataset to increase performance. The final dataset contained one hundred eight (108) male instances and ninety-three (93) female instances in which both have an equal distribution of flow, confusion and boredom. The two final music recommendation models, which were based from the gender classification, were built using three classification techniques, namely J48, JRip and PART on both the female and male datasets. J48 provided the best results for the performance of both models with an accuracy of 75.9% and a kappa statistic of 0.6389 when the balanced dataset of the male subjects were used. On the other hand, the results for the balanced female dataset using the J48 algorithm were 54.8% for accuracy and 0.3226 for kappa statistic