Musical preferences prediction by classification algorithm

In this paper, we use several supervised classification algorithms to predict musical preference of a person. From psychological point of view, although personal emotion is an important feature that has an influence on selecting music, there are some other significant factors such as age, sex, education and district that might have an impact on our musical choices. In this paper, we first collected our data based on an observation method called stratified sampling. In this model, we collected 2000 cases that were grouped into strata (as district in our data feature), then simple random sampling was employed within each stratum. We partitioned our original dataset into two classes, 60% of which we were used to train our models and 40% of which we were held back as a validation dataset. The dataset contains five features as follows: four features named sex, age, education and district as explanatory variables and one feature named music known as response or target variable. The response variable has two different levels, namely traditional and non-traditional so we were dealing with a binary classification. The dataset that we created is called MPD. Moreover, we calculated some important statistical measures such as accuracy, specificity, precision, sensitivity and F-measure. Finally, we examined four different algorithms using R which were a nice mixture of nonlinear (cart, knn) and complex nonlinear methods (rf) and the result in random forest had the highest accuracy with 86.8%. We also observed that the highest F-measure is gained by cart algorithm with 44.7% score. As we have not considered the person's emotion as an influential factor on musical choices, we could expect the accuracy of learning algorithms would not react at very high performance. Our results proved this claim.