Enhanced-MovieLens-Recommender

This project extends and refines the PEAGNN (Path Extrapolation and Aggregation Graph Neural Networks) approach for movie recommendations using the MovieLens dataset. Building upon the work from PEAGNN, primary focus is on redefining the criteria for positive and negative samples to improve recommendation accuracy.

Key Contributions

1. Redefined Positive/Negative Sample Criteria:

   • Implemented a new, more nuanced approach to classifying user interactions:
     • Positive samples: Ratings >= 4
     • Negative samples: Ratings <= 2
   • This redefinition aims to capture user preferences more accurately, considering ratings of 4 and above as indicative of positive user experience.

2. Adaptive Sampling Strategy:

   • Developed a method to handle the natural imbalance between positive and negative samples resulting from the new criteria.
   • Introduced a configurable positive-to-negative ratio for flexible dataset creation, allowing for experimentation with different sampling ratios.

3. Modified Model Architecture:

   • Adapted the PEAGAT (Path Extrapolation and Aggregation Graph Attention Network) model to work effectively with the newly defined sample criteria.
   • Updated the forward pass and prediction methods to handle the redefined positive and negative interactions.

4. Refined Evaluation Method:

   • Adjusted the test method to align with the new sample definition, ensuring that the evaluation reflects the updated criteria for positive interactions.
   • Implemented Hit Ratio at 10 (HR@10) as the primary evaluation metric, based on the redefined positive sample criterion.

Implementation Details

• Data Preprocessing: The MovieLens class now includes methods for preparing training data based on the new positive/negative sample criteria.
• Model: The PEAGATRecsysModel class has been updated to work with the redefined interaction types.
• Training and Evaluation: The Solver class now incorporates the new sample definitions in both training and evaluation processes.

Running the Code

python main.py --dataset Movielens --type 25m --hidden_dim 64 --num_layers 2 --num_heads 1 --dropout 0 --channel_aggr att --device cuda --gpu_idx 0 --runs 5 --epochs 30 --batch_size 1024 --lr 0.001 --weight_decay 0.001

Future Work

• Investigate the impact of different rating thresholds for positive/negative sample classification.
• Explore how the redefined criteria affect recommendation diversity and novelty.
• Conduct comparative studies with other sample definition approaches in recommender systems.

Project Basis

This work is based on the PEAGNN project.