|
Activity Number:
|
306
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Tuesday, August 5, 2008 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Statistics and Marketing
|
| Abstract - #300620 |
|
Title:
|
Random Forests for Internet Recommendations
|
|
Author(s):
|
Michiel van Wezel and Erik Miedema and Rob Potharst*+
|
|
Companies:
|
Erasmus University and Erasmus University and Erasmus University
|
|
Address:
|
P.O.Box 1738, Econometric Institute, Rotterdam, International, 3000 DR, The Netherlands
|
|
Keywords:
|
recommender system ; random forest ; decision tree ; multiple targets
|
|
Abstract:
|
We consider the use of Random Forests (RF) for internet recommendation: the most interesting items (movies, books, etc) should be recommended to a user. Recommendation problems are typically characterized by high numbers of users and items, and a sparse data matrix. To generate recommendations for large numbers of items, we adapt the standard CART algorithm to deal with multiple targets. Three splitting rules are implemented: choosing the split that minimizes the loss for a random target, choosing the split that minimizes the loss for all targets, and ignoring all targets by splitting on a random value of the split-variable. Nodes in our trees may inherit predictions when data in a target dimension is too sparse to be reliable. Experiments with the well-known Jester dataset show that accuracy and speed of the RF approach compete with the top results reported in the literature.
|
