Apriori vs Word2Vec: Comparative Analysis of Product Recommendation Systems

This repository contains a comparative analysis of traditional association rule mining (Apriori algorithm) with neural embedding techniques (Word2Vec) for product recommendation systems in retail environments.

Overview

The project analyzes transaction data to generate product recommendations using two distinct approaches:

1. Apriori Algorithm: A classic market basket analysis technique that identifies frequent itemsets and generates association rules to discover relationships between products purchased together.

2. Word2Vec: A neural network-based approach that represents products as vector embeddings, capturing semantic relationships between products based on their co-occurrence in transactions.

The goal is to quantitatively and qualitatively evaluate the differences in recommendation quality, performance, and practical applications of both techniques.

Dataset

The analysis uses retail transaction data containing:

• Transaction IDs (BillNo)
• Product names (Itemname)
• Categories (Manually added)
• Dates
• Quantities
• Customer IDs

The dataset is analyzed for key characteristics including unique products, transaction sizes, purchase patterns, and category distributions to provide context for the recommendation system evaluation.

Product Categorization

Products are organized into a hierarchical category structure:

Main Categories:

• Home & Living
• Personal Items
• Entertainment & Hobbies
• Seasonal & Gifts
• Office & Craft

Each main category is further divided into subcategories. For example:

• Home & Living: Home Decor, Lighting, Kitchen & Dining, Textiles & Cozy Items, etc.
• Personal Items: Fashion & Accessories, Electronics & Gadgets
• Entertainment & Hobbies: Kids & Toys, Books & Magazines, Games & Puzzles, etc.

Products were classified through both automated methods and manual review for accuracy (There are more details regarding this in the thesis "3.1.1 Dataset Overview" section)

Repository Structure

├── Dataset/
│   ├── Analysis_plots/
│   ├── Kaggle_Dataset_Pre_Processing/                    # The original Kaggle Dataset without preprocessing
│   ├── distinct_products_with_categories.xlsx            # Unique products category mapping
│   ├── full_validated_dataset.xlsx                       # Final dataset that is validated and sent to the algorithms
│   ├── not_validated_dataset_with_category.xlsx          # Contains the full dataset with category mapping for each product coming from the assignment of gpt 3.5 turbo
│   ├── Original_Dataset.xlsx
│   ├── unique_items.xlsx                                 # Unique products
│   └── validated_distinct_products_with_categories.xlsx  # Manually validated category mapping dataset (row by row manual vaildation)
│
├── Models/
│   ├── Models_Evaluator.ipynb                            # Complete evaluation code
│   └── Models_Parameters_Grid_Search.ipynb               # Parameter optimization code
│
├── Results/                                              # Generated analysis outputs
│   ├── whole_dataset/                                    # Full dataset results
│   ├── first_third/                                      # analysis (first 1/3)
│   └── last_two_thirds/                                  # analysis (last 2/3)
│
│
├── Finding_Categories/                                   # Categorization methodology
│   ├── high_level_categories.xlsx
│   ├── low_level_openai_category_products_mapping.ipynb
│   └── *.ipynb                                           # Category assignment notebooks
├── Dataset-Analysis.ipynb                                # Full data explorationn for the third thesis captial

Requirements

• Python 3.8+
• pandas
• numpy
• matplotlib
• seaborn
• scikit-learn
• mlxtend (for Apriori implementation)
• gensim (for Word2Vec implementation)
• Jupyter Notebook

Install required packages:

pip install pandas numpy matplotlib seaborn scikit-learn mlxtend gensim jupyter

Usage

1. Clone the repository:

   git clone https://github.com/Mohammed87FS/Apriori_VS_Word2Vec.git
   cd Apriori_VS_Word2Vec

2. Run the Jupyter notebooks in sequence:

   jupyter notebook

   • First, run Dataset-Analysis.ipynb to understand the data
   • Then, explore the implementations in Models_Evaluator.ipynb
   • Finally, check comparative analysis in Models_Evaluator.ipynb

Methodology

The project follows these key steps:

1. Data Preprocessing:

   • Cleaning transaction data (handling missing values, duplicates)
   • Organizing products into the hierarchical category structure
   • Splitting data into training and evaluation sets

2. Parameter Optimization: Both algorithms use grid-search optimized parameters to find the configuration where both models have the highst agreement percentage regarding their recommendations to have a better comparasion foundation.

   Apriori Configuration:

   • Minimum Support: 0.01 (1% of transactions)
   • Minimum Confidence: 0.1 (10% confidence threshold)

   Word2Vec Configuration:

   • Vector Size: 50 dimensions
   • Window Size: 10 (context window)
   • Training Epochs: 10
   • Min Count: 20 (minimum word frequency)
   • Similarity Threshold: 0.5 for recommendations
   • Architecture: Skip-gram model

3. Apriori Implementation:

   • Converts transactions to appropriate basket format
   • Measure performance and memory usage across different dataset partions

4. Word2Vec Implementation:

   • Formats transactions as "sentences" of products for training
   • Trains product embeddings using Skip-gram architecture
   • Generates similar product recommendations based on vector similarity

5. Evaluation Metrics:

   • Coverage (percentage of products that can receive recommendations)
   • Execution time for model training and recommendation generation
   • Memory efficiency during training and storage
   • Scalability with increasing dataset size
   • Relevance of recommendations (within-category vs cross-category)

Results

The comparative analysis reveals distinct strengths and weaknesses of each approach:

• Apriori:

  • Provides highly interpretable rules with clear confidence metrics
  • Excellent for identifying direct product relationships
  • Struggles with scalability on large product catalogs
  • High memory usage grows

• Word2Vec:

  • Offers superior recommendation coverage across the product catalog
  • Discovers more product relationships
  • Scales better with increasing dataset size
  • Less interpretable recommendations requiring additional validation

Detailed performance metrics, visualizations, and recommendation examples can be found in the evaluation notebook.

Contact

• Author: Mohammed Alhamadani
• Email: mohammedamaar165@gmail.com
• LinkedIn: https://www.linkedin.com/in/mohammed-ammar-burhan-al-hamadani-a88518302/
• Institution: FH St.Pölten

Acknowledgments

• Thanks to my supervsior FH-Prof. Dipl.-Ing. Dr. Markus Seidl, Bakk. for his guidance and support