One of the algorithms to perform better in the KDD CUP contest this year is Koren's time-SVD++ algorithm. You can find more details in a previous post.
To summarize, according to our testing, we got the following validation RMSE on KDD CUP data:
(as lower prediction error, the better..)
1 Neighborhood model Adjusted cosine (AC) similarity 23.34
2 ALS 22.01
3 Weighted ALS 18.87**
4 BPTF 21.84
5 SGD 21.88
6 SVD++ 21.59
7 Time aware neighborhood 22.7
8 time-SVD 21.41
9 time-SVD++ 20.90
10 MFITR 21.30
11 time-MFITR 21.10
12 Random forest 26.0
(Weighted ALS did not perform well on test data).
As can be seen from the above table, time-SVD++ was the best performing single algorithm
on the KDD CUP data. I also got the same impression when talking to Yehuda Koren.
time-SVD++ is now implemented as part of GraphLab's collaborative filtering library.
You are all welcome to try it out!
No comments:
Post a Comment