i designed music recommendation system. 1st approach was to use item item matrix(co-occurence matrix) . 2nd approach was to make song vs user matrix(with listen count as values) and calculate both song and user features (just like we do in movie rating) using collaborative filtering linear regression.
But the results for both of them on same user are pretty different.So, is it normal or am i doing something wrong?
i checked my 2nd approach and it fitted well my training examples(means it predict listen count which we already have pretty closely)