|
A lot of recommendation models have a static model of the user. However, depending upon my mood, I have different preferences. Can you propose a simple model of mood, that makes it possible for the recommendation system to quickly and accurately infer a user's mood, and make recommendations accordingly? Is there recommended prior work on this topic? |
|
If your system is bayesian, "mood" could be a latent variable that acts as a regularizer. I.e., moods are global patterns of behavior (so you have per-mood item baselines instead of global item baselines). Assuming your user is giving you lots of data you can then infer what's his mood now by seeing which mood better predicts his past short-term behavior and recommend accordingly. I'm not sure how to make this fast, however, and I didn't find any references on collaborative filtering that incorporated this sort of model (but then I'm not very familiar with the area and could be missing something obvious). Right, I was thinking about modeling this as a latent variable (hence the tag ;). What kind of latent variable? Multinomial?
(Sep 16 '10 at 14:19)
Joseph Turian ♦♦
I'd use a discrete (multinomial with only one sample), yes, so you can have a discrete number of moods.
(Sep 16 '10 at 14:23)
Alexandre Passos ♦
|
