Recommending VideoLectures with Linear Regression

Discovery Challenge Workshop 2011.

工业技术;计算机科学

Martin Mozina ; Aleksander Sadikov ; Ivan Bratko

Others  :  http://ceur-ws.org/Vol-770/paper7.pdf
PID  :  42381

来源: CEUR

【 摘 要 】

This paper describes our approach to the task 1 of the ECML PKDD 2011 VideoLectures.Net Recommender System Challenge. The task was to select a set of lectures to be recommended to a visitor of the VideoLecture.Net homepage after already seeing another lecture. Our proposed approach is a hybrid recommender system combining content and collaborative approaches. The core of the system is a linear regression model for predicting the rank of a lecture, whereas by rank we mean the lecture’s position in the list of all lectures ordered by the interest of the visitor. Due to the complexity of the problem, the model could not be learned by a classical approach - instead, we had to employ the stochastic gradient descent optimization. The present paper furthermore, through evaluation, identifies and describes some interesting properties of the domain and of the algorithm that were crucial to achieve a higher prediction accuracy. The final accuracy of the model was enough to take the third place in the competition.