For our assignment, we used the "hetrec" last.fm dataset.
Return true if this artist is one of the top 50 artists in our training set.
All of the models differ only in the fourth case (the else) of the if statement in m*_predict. This is the case in which we have seen a user and an artist in training, but have not seen a prior listen between them. Each of the descriptions below will only discuss that case.
All of the models take as input a userID (as uid) and an artistID (as aid).
Below, acc refers to the best accuracy we found on the validation set (by tuning the hyperparameters).
Return true if any of this user's friends have listened to this artist.
Return true if this user is similar to any of the users that have listened to this artist. ("Similar" means their Jaccard similarity [based on the artists they've listened to] exceeds the threshold of 0.4, which we found experimentally via tuning the model based on results on the validation set).
Return true if the any of the user's top tags are in the artist's tags (user top tags = the set of tags attributed to any artist by this user, reverse sorted, and sliced). Best accuracy was observed when taking the top 16 of the user's tags.
In this model, we sorted tags by the number of times they were used for a given user or artist.
Same as above, except in this model, we sorted tags by the number of listens associated with them (for a given user/artist pair, we had both the number of listens as well as a collection of tags for the pair; increment the tag's frequency by the number of listens).
Same as Model 2, but for this model, we computed similarity based on the tags each user has ever used. Again, we tuned the threshold based on validation set performance, and ofund 0.06 to be the best.
Same as Model 1, but checks if any of this user's friends of friends have listened to the artist.
Same as Model 2, but for this model, instead of checking similarity with all the users that listened to this artist, we checked similarity with the intersection of that set and the user's friends.
Same as Model 4, but additionally check if any of the user's friend's top tags are shared by the artist.
Same as Model 2, but the similarity score for this model included both the artist similarity and the tag similarity between the given user and each of the users that listened to this artist. We tuned the Jaccard threshold based on validation set performance and found the best to be 0.05 again.
Same as Model 9, but instead of using all the user's tags in the similarity calculation, we only used the top tags by frequency (like in Model 3). We got the best performance on the validation set when using the top 22 tags.
Same as Model 2, but for the similarity calculation, instead of using all of the artists from each user, we used the top n artists by listens for each user. We found that n = 21 gave the best performance on the validation set.