Skip to content

Latest commit

 

History

History
46 lines (39 loc) · 4.17 KB

README.md

File metadata and controls

46 lines (39 loc) · 4.17 KB

Improving Collaborative Filtering Techniques by the use of Co-Training in Recommender Systems.

This repository holds the implementation, the datasets and results for the thesis project: Improving Collaborative Filtering Techniques by the use of Co-Training in Recommender Systems.

This project was done under a thesis research for Fernando Benjamín Pérez Maurera, under supervision of Professor Paolo Cremonesi and Engineer Maurizio Ferrari, at Politecnico di Milano.

Project organization.

The project is organized as follows:

  • Datasets: A folder where the datasets are located.
    • ml10m: folder containing the Movielens10M dataset.
      • ratings.csv: ratings file.
  • Implementation: RecPy module where the recommenders and helper classes are.
  • read-results: folder where bash scripts are, and are used to read the results output by Co-Training. There are for each recommender combination. The cases inside are ItemKNN/FunkSVD, ItemKNN/SLIM, ItemKNN/SLIMBPR and ItemKNN/BPRMF.
  • run-examples: folder where bash scripts to run Co-Training are. There are several recommenders combinations as: ItemKNN/FunkSVD, ItemKNN/SLIM, ItemKNN/SLIMBPR and ItemKNN/BPRMF.
  • Results: A folder generated when running run-knn.sh and results-knn/sh scripts. In this folder the results for each test-case will be put.
  • scripts: Folder where its located the two main Python files, holdout.py, which makes a holdout@k of the dataset, runs Co-Training and evaluates the recommenders, and read-results.py, which reads the results of each output file and generates new plots.
  • README.md: This file.
  • requirements.txt: File for Conda or PIP that has the libraries and modules required to run the code.
  • results-knn.sh: main bash script to read the results that each Co-Training process outputs for each test case inside read-results.
  • run-knn.sh: main bash script to run the Co-Training process for each test case inside run-examples.

Project installation.

Requirements.

  • Python 3.6+.
  • C++ Compiler.
  • On Linux, ensure that you have packages libc6-dev and build-essentials

Installation instructions

  1. [On Linux] Install Linux packages: apt-get install -y libc6-dev build-essentials.

  2. Install Miniconda for Python 3.6+ here.

  3. Create the virtual environment: conda create -n cotraining --file requirements.txt

  4. Activate the virtual environment: source activate cotraining.

  5. [Installation and run separately] Install the project: cd Configuration/ ; sh install.sh ; cd ..

  6. [Installation and run separately] Run one of the examples: * cd run-examples/ ; sh knn-funksvd.sh -p <p-most positive> -n <n-most negative> -u <size of U'>; cd .. * cd run-examples/ ; sh knn-slim.sh -p <p-most positive> -n <n-most negative> -u <size of U'>; cd .. * cd run-examples/ ; sh knn-bprmf.sh -p <p-most positive> -n <n-most negative> -u <size of U'>; cd .. * cd run-examples/MyMediaLite/bin/ ; sh knn-slimbpr.sh -p <p-most positive> -n <n-most negative> -u <size of U'>; cd ..

  7. [Installation and run integrated] Run the run-knn.sh script: sh run-knn.sh -p <p-most positive> -n <n-most negative> -u <size of U'>

  8. [Only to generate new plots] Run the results-knn.sh script: sh results-knn.sh -p <p-most positive> -n <n-most negative> -u <size of U'>

Results.

The test cases included with the project are ItemKNN/FunkSVD, ItemKNN/SLIM, ItemKNN/SLIMBPR and ItemKNN/BPRMF. The dataset used is Movielens10M, a holdout technique at 20% was used. A top-10 recommendation list was generated at evaluation time for each user. The items were divided into 10 bins based on their popularity, where the least popular is bin_0 and the most popular is bin_9. When running the test cases, the Results folder will be created, and a subfolder for each test case will be created.

At the moment, the project generates output files for: evaluation of RMSE, MAP, ROC-AUC, Precision, Recall, NDCG and MRR, a file containing the number of p-most and n-most items rated at each iteration, the agreement between the recommenders, and the popularity of the items recommended.