Non-Deep Learning Collaborative Filtering Techniques for Implicit Feedback Data in Music Recommender Systems

Note: this work and theory have been extended in the work / repository: Application of Graph Neural Networks to Music Recommender Systems

• Introduction
• Dataset
• Evaluation
• Repository Structure
• Setup
  • Download the Data
  • Setup the Virtual Environment
    • Setup with conda
    • Setup with pip
• Export requirements.txt

🕵️ Introduction

This repository contains code and resources for the research seminar titled "Non-Deep Learning Collaborative Filtering Techniques for Implicit Feedback Data in Music Recommender Systems." The focus of this seminar is on applying user-based and item-based neighborhood techniques, as well as matrix factorization using alternating least squares (ALS) optimization, to a real-life dataset with binary feedback. By implementing and evaluating these classical collaborative filtering methods, this work aims to establish a baseline for future research in music recommender systems, facilitating the development and comparison of more advanced techniques.

💾 Dataset

• # of customers: 58.747
• # of records: 37.370

#	file name	# rows, (users, items)	sparsity	features
1	user_item_interaction_RAW_ANONYMIZED.txt	23.545.542, (70.309, 37.408)	0.68 %	userID, itemID
2	user_item_interaction_FILTERED_ANONYMIZED.txt	17.665.904, ( 58.747, 37.370 )	0.8 %	userID, itemID

💎 Evaluation

After the hyperparameter search, the best 3 settings for each algorithm were applied to the entire filtered dataset. A Top-10 recommendation was made for a random set of 100 users. This process was repeated 5 times for each algorithm and each parameter setting, and the mean of the best 5 runs is reported in the following table in descending order by the NDCG. The best values are printed in bold and the second largest are underlined:

Algorithm	Pre	MRR	nDCG	MAP	IC	ARP	APLT
ALS-MF	0.116	0.3587	0.4040	0.2643	0.01663	5179	0.00
UserkNN	0.111	0.3288	0.3934	0.2544	0.01540	5831	0.002
UserAsymkNN	0.0980	0.2849	0.3295	0.2105	0.01616	5300	0.016
ItemAsymkNN	0.0642	0.2374	0.2838	0.2035	0.005823	9836	0.0104
MostPop	0,0118	0.04305	0.05797	0.04223	0.001054	12478	0.00

Repository Structure

• data: contains the data used for the recommendation algorithms
  • evaluation: contains the evaluation files and parameter settings
  • processed: contains the filtered dataset
  • raw: contains the raw un-filtered dataset
• notebooks:
  • DataObservation.ipynb: initial observation of the raw dataset
  • Evaluation.ipynb: evaluates the best parameter settings
  • EvaluationParameterSearch.ipynb: evaluates the parameter search
  • ParameterSearchItem.ipynb: the parameter search logic for the item-based algorithms
  • ParameterSearchMF.ipynb: the parameter search logic for the alternativ least squares algorithm
  • ParameterSearchUser.ipynb: the parameter search logic for the user-based algorithms
• src: Enthält Entwicklungscode zum Backend inkl. separater README.md mit Erläuterungen
  • Helper.py: contains often used functions
  • Metrics.py: contains the entire metric and evluation logic
  • MfAlgorithms.py: contains the matrix factorization recommender
  • NeighborhoodAlgorithms.py: contains the neighborhood-based recommender
  • NonPersonalizedAlgorithms.py: contains the non-personalized recommender
• CONDA_Requirements.txt: contains all installed libraries generated via conda list -e CONDA_Requirements.txt
• PIP_Requirements.txt: contains all installed libraries generated via pip3 freeze > PIP_Requirements.txt
• Seminar_Paper.pdf: the corresponding paper "Non-Deep Learning Collaborative Filtering Techniques for Implicit Feedback Data in Music Recommender Systems"

Setup

Download and store the Dataset

The dataset can be found in the Google Drive. Store the containing files as follows:

• user_item_interaction_RAW_ANONYMIZED.txt -> data/raw
• user_item_interaction_FILTERED_ANONYMIZED.txt -> data/processed

Setup Virtual Environment

Setup with Conda (recommended)

Create virtual environment with predefined libraries:

    conda create --name <env> --file requirements.txt

Setup with Pip

0. Use python version between 3.8 and 3.9 (3.9.19 used and recommended)

1. Create virtual environment in project directory (ResearchSeminarMusicRecommender2024):

   python3 -m venv .venv
   

2. Activate virtual environment:

• on Mac / Linux:

    source .venv/bin/activate
  

• on Windows:

    .venv\Scripts\activate
  

3. Upgrade pip

    pip3 install --upgrade pip
   

4. Install Packages

   pip install -r PIP_Requirements.txt
   

Export requirements.txt

Open the terminal, activate the environment, and run:

    pip list --format=freeze > requirements.txt