MovieLens-Recommender

MovieLens-Recommender is a pure Python implement of Collaborative Filtering. Which contains User Based Collaborative Filtering(UserCF) and Item Based Collaborative Filtering(ItemCF). As comparisons, Random Based Recommendation and Most-Popular Based Recommendation are also included. The famous Latent Factor Model(LFM) is added in this Repo,too.

The buildin-datasets are Movielens-1M and Movielens-100k. But of course, you can use other custom datasets.

Besides, there are two models named UserCF-IIF and ItemCF-IUF, which have improvement to UseCF and ItemCF. They eliminate the influence of very popular users or items.

Overview

The book 《推荐系统实践》 written by Xiang Liang is quite wonderful for those people who don't have much knowledge about Recommendation System. But the book only offers each function's implement of Collaborative Filtering. A good architecture project with datasets-build and model-validation process are required.

So I made MovieLens-Recommender project, which is a pure Python implement of Collaborative Filtering based on the ideas of the book.

This repository is based on MovieLens-RecSys, which is also a good implement of Collaborative Filtering. But its efficiency is so damn poor!

Besides, Surprise is a very popular Python scikit building and analyzing recommender systems. So, I Mix the advantages of these two projects, and here comes MovieLens-Recommender.

My Recommendation System contains four steps:

Create trainset and testset
Train a recommender model
Give recommendations
Evaluate results

At the end of a recommendation process, four numbers are given to measure the recommendation model, which are:

Precision
Recall
Coverage
Popularity

No python extensions(e.g. Numpy/pandas) are needed!

Getting started

1. Download

Git is awesome~

git clone https://github.com/fuxuemingzhu/MovieLens-Recommender.git

Movielens-1M and Movielens-100k datasets are under the data/ folder.

2. Run

The configures are in main.py. Pleas choose the dataset and model you want to use and set the proper test_size. The default values in main.py are shown below:

dataset_name = 'ml-100k'
# dataset_name = 'ml-1m'
# model_type = 'UserCF'
# model_type = 'UserCF-IIF'
# model_type = 'ItemCF'
# model_type = 'Random'
# model_type = 'MostPopular'
model_type = 'ItemCF-IUF'
# model_type = 'LFM'
test_size = 0.1

Then run python main.py in your command line. There will be a recommendation model built on the dataset you choose above.

Note: my code only tested on python3, so python3 is prefer.

Python main.py
#Python3 main.py

if you are using Linux, this command will redirect the whole output into a file.

Python main.py > run.log 2>&1 &
#Python3 main.py > run.log 2>&1 &

This command will run in background. You can wait for the result, or use tail -f run.log to see the real time result.

All model will be saved to model/ fold, which means the time will be cut down in your next run.

3. Output

Here is a example run result of ItemCF model trained on ml-1m with test_size = 0.10. No mater which model are chosen, the output log will like this.

**********************************************************************
    This is ItemCF model trained on ml-1m with test_size = 0.10
**********************************************************************

ItemBasedCF start...

No model saved before.
Train a new model...
counting movies number and popularity...
counting movies number and popularity success.
total movie number = 3693
generate items co-rated similarity matrix...
 steps(0), 0.00 seconds have spent..
 steps(1000), 18.50 seconds have spent..
 steps(2000), 46.39 seconds have spent..
 steps(3000), 63.52 seconds have spent..
 steps(4000), 87.37 seconds have spent..
 steps(5000), 111.83 seconds have spent..
 steps(6000), 132.71 seconds have spent..
generate items co-rated similarity matrix success.
total  step number is 6040
total 133.61 seconds have spent

calculate item-item similarity matrix...
 steps(0), 0.00 seconds have spent..
 steps(1000), 1.77 seconds have spent..
 steps(2000), 3.47 seconds have spent..
 steps(3000), 5.01 seconds have spent..
calculate item-item similarity matrix success.
total  step number is 3693
total 5.67 seconds have spent

Train a new model success.
The new model has saved success.

recommend for userid = 1:
['1196', '364', '1265', '318', '2081', '1282', '1198', '2716', '1', '2096']

recommend for userid = 100:
['2916', '1580', '457', '1240', '589', '1291', '780', '1036', '1610', '1214']

recommend for userid = 233:
['1022', '594', '1282', '2087', '2078', '1196', '608', '2081', '593', '1393']

recommend for userid = 666:
['296', '1704', '593', '356', '1196', '589', '1580', '50', '1393', '1']

recommend for userid = 888:
['2916', '457', '480', '2628', '1265', '1610', '1198', '1573', '2762', '1527']

Test recommendation system start...
 steps(0), 0.10 seconds have spent..
 steps(1000), 291.42 seconds have spent..
 steps(2000), 627.60 seconds have spent..
 steps(3000), 898.21 seconds have spent..
 steps(4000), 1219.94 seconds have spent..
 steps(5000), 1523.29 seconds have spent..
 steps(6000), 1817.46 seconds have spent..
Test recommendation system success.
total  step number is 6040
total 1829.26 seconds have spent

precision=0.1900	recall=0.1147	coverage=0.1673	popularity=7.3911

total Main Function step number is 0
total 1972.49 seconds have spent

Benchmarks

Here are four models' benchmarks over Precision、Recall、Coverage、Popularity. The testsize is 0.1.

These results are nearly same with Xiang Liang's book, which proves that my algorithms are right.

Movielens 1M:

Movielens 1M	Precision	Recall	Coverage	Popularity
UserCF	19.84%	11.97%	28.16%	7.2023
ItemCF	19.00%	11.47%	16.73%	7.3911
UserCF-IIF	19.77%	11.93%	29.62%	7.1660
ItemCF-IUF	18.71%	11.29%	15.03%	7.4748
LFM	/	/	/	/
Random	0.54%	0.33%	100.00%	4.4075
Most Popular	10.59%	6.39%	2.76%	7.7462

Movielens 100k:

Movielens 100k	Precision	Recall	Coverage	Popularity
UserCF	19.69%	18.50%	22.20%	5.4928
ItemCF	17.89%	16.80%	13.23%	5.6202
UserCF-IIF	19.57%	18.38%	22.74%	5.4716
ItemCF-IUF	20.38%	12.30%	17.30%	7.3643
LFM	20.29%	19.06%	27.41%	4.9983
Random	0.82%	0.77%	99.64%	3.0332
Most Popular	10.54%	9.90%	4.07%	5.9578

Notice

UserCF is faser than ItemCF. Using ml-100k instead of ml-1m will speed up the predict process.

Caculating similarity matrix is quite slow. Please wait for the result patiently.

LFM will make negative samples when running. And when the ratio of Neg./Pos. goes to larger, the performance goes to better.

LFM has more parameters to tune, and I don't spend much time to do this. I believe you will do quite better!

Licence

Apache License.