Transfer learning

Using source domain knowledge to improve target domain recommandation results . This project is implemented by tensroflow and libMF

Environment

Mac os x, tensorflow 0.11.0

Usage

compile the libmf

$ cd libmf-2.01
$ make
$ cd ..

split the validation set from training data this python code will create a directory with validation data (./valid)

$ python split_valid.py [data_directory] [cross_validation_size]

if the test.txt in the directory contain answer, then valid = 1 (default is 0)

$ python mapping.py --data_dir [data_directory] --valid [1/0]

use the mapping and the source data to learn a matrix factorization

$ python mf.py --data_dir [data_directory] --valid [1/0] --iter [iterations]

example:

$ python split_valid.py ./test1 5 
$ python mapping.py --data_dir ./valid --valid 1
$ python mf.py --data_dir ./valid --valid 1 --iter 30

Performance

5-cross_validation	MF	Transferred MF
test1	RMSE = 0.18502	RMSE = 0.18008
test2	RMSE = 0.18548	RMSE = 0.17934
test3	RMSE = 1.35124	RMSE = 1.25096

Data

1. test1

   target and source data are exactly same item and user but with different id mapping

2. test2

   disjoin user and item (items in the same domain)

3. test3

   disjoin user and item (totaly different domain)