Decompose Matrix A(m x n) with user factor matrix W(m x k) and item factor matrix H(n x k) where A = W * H'. A is usally a sparse matrix and k is much smaller than m and n. We implement an asynchronous, distributed version of matrix factorization in Paracel.
- Enter Paracel's home directory
cd paracel; - Generate dataset
python ./tool/datagen.py -m mf -o netflix.dat - Set up link library path:
export LD_LIBRARY_PATH=your_paracel_install_path/lib - Create a json file named
cfg.json, see example in Parameters section below. - Run (100 workers, 20 servers, mesos mode in the following example)
./prun.py -w 100 -p 20 -c cfg.json -m mesos --ppn 10 --mem_limit 1000 your_paracel_install_path/bin/mf
Default parameters are set in a JSON format file. For example, we create a cfg.json as below(modify your_paracel_install_path):
{
"input" : "netflix.dat",
"predict_input" : "netflix.dat.predict",
"output" : "./mf_result/",
"update_file" : "your_paracel_install_path/lib/libmf_update.so",
"update_functions" : ["cnt_updater", "mf_fac_updater", "mf_bias_updater"],
"filter_file" : "your_paracel_install_path/lib/libmf_filter.so",
"filter_functions" : ["mf_ubias_filter", "mf_ibias_filter", "mf_W_filter", "mf_H_filter"],
"k" : 100,
"rounds" : 5,
"alpha" : 0.001,
"beta" : 0.001,
"debug" : false,
"ssp_switch" : true,
"limit_s" : 3
}
update_file, update_functions, filter_file and filter_functions store the information of registry function used in the implementation of matrix factorization program. krefers to the factor dimension of matrix W and matrix H. rounds refers to the number of training iterations. alpha is the learning rate of sgd algorithm and beta is the regularization parameter. If you set ssp_swith with true value, this means the training process will be running as asynchrounous mode. limit_s is also used for asynchrounous training. For example, if it equals to 3, it means that the fastest worker will lead no more than three iterations than the slowest worker. If it equals to 0, it is the classic BSP training model.
netflix.dat: training dataset from netflix movie rating data, each line presents a tetrad with "user_id movie_id date rating".
netflix.dat.predict: data to be predicted, each line presents a pair with "user_id,movie_id"
miu_0: record rating size and global mean value.
ubias_0: record user bias values.
W_0: record user factors with k dimension.
ibias_0: record item bias values.
H_0: record item factors with k dimension.
pred_v_x: predict rating value of specified users and movies which are specified in netflix.dat.predict.
Koren, Yehuda, Robert Bell, and Chris Volinsky. "Matrix factorization techniques for recommender systems." Computer 8 (2009): 30-37.