Recommender Systems using factored MDP
Data:
The log data is in the following format:
Item_ID ItemFamily_ID Brand Color CategoryGroup1 CategoryGroup2 CategoryGroup3 CategoryGroup4 CategoryGroup5 Cookie_ID Session_ID Action Count PriceLeve
The lines are sorted based on Session_ID preserving the original order, therefore, the sequence of clicks in every session is available.
Config: In the config file, the first line is an integer 'm' defining the number of attributes used from the log data. And the next m lines are the column indexes of attributes in the input file.
The last line is the length of history which here is fixed for all the attributes.
Code: The models in both "Topic Detection" and "Recommendation" are the same, and only the evaluation part is different. In Topic Detection, the topics for every attribute are chosen according to a threshold and saved into separate files.
In Recommendation, the rank of items from the test data is computed from obtained Q-values.
Citation:
@inproceedings{tavakol2014factored,
title={Factored MDPs for detecting topics of user sessions},
author={Tavakol, Maryam and Brefeld, Ulf},
booktitle={Proceedings of the 8th ACM Conference on Recommender Systems},
pages={33--40},
year={2014},
organization={ACM}
}