This project aims to design a high-quality recommender system for short videos using the KuaiRec 2.0 dataset. The goal is to accurately recommend and rank videos for users based on their historical interactions, content features, and video popularity.

We implement a hybrid multi-stage pipeline combining:

• Collaborative Filtering (ALS),
• Binary classification (LightGBM),
• Regression ranking (LightGBM).

Our final model optimally balances classification accuracy and ranking quality, addressing the limitations of single-model approaches.

The KuaiRec dataset contains:

• User interactions (e.g., watch ratio, timestamps),
• Video metadata (captions, categories, etc.),
• User metadata (activity ranges, categorical features),
• Daily popularity metrics (views, likes, comments).

All data are stored in CSV files and preprocessed to extract structured features used for training the models.

Before using the current model, I tried multiple approaches to have the best recommender system as possible:

Our final approach consists of three stages:

• Stage 1: ALS Matrix Factorization
  Generates latent representations of users and items. Produces als_score, used as a feature downstream.

• Stage 2: LightGBM Classifier
  Predicts whether a user will "like" a video based on engineered features. This step filters candidates.

• Stage 3: LightGBM Regressor
  Predicts the watch_ratio to re-rank the top items selected by the classifier, optimizing ranking metrics (e.g., NDCG).

• Input Features:
  • User-item interactions matrix with watch_ratio_clipped as interaction strength
• Output:
  • als_score - Collaborative latent factor dot-product representing user-item affinity (64 factors)

• Input Features:
  • Video manual_cover_text (descriptions)
  • Video caption text
  • Video topic_tag keywords
• Output:
  • 50 text embedding features (tfidf_0 to tfidf_49) capturing content semantics

• Input Features:
  • User Features:
    • Categorical: user_active_degree, follow_user_num_range, fans_user_num_range, friend_user_num_range, register_days_range
    • One-hot encoded features: onehot_feat0 to onehot_feat17
    • Behavioral: user_agg_interactions, user_agg_unique_videos
  • Item Features:
    • Content: first_level_category_name, second_level_category_name, third_level_category_name, author_id
    • Text: TF-IDF embeddings (tfidf_0 to tfidf_49)
    • Popularity: Avg daily metrics (avg_daily_show_cnt, avg_daily_play_cnt, avg_daily_like_cnt, etc.)
    • Historical engagement: video_agg_views, video_agg_unique_users
    • Metadata: video_duration
  • Contextual Features: hour_of_day, day_of_week
  • Collaborative Signal: als_score from ALS model
• Output:
  • classifier_score - Probability of user liking the video (binary prediction where 1 = watch_ratio > 1.0)

• Input Features:
  • Same feature set as the LightGBM Classification model
• Output:
  • regressor_score - Predicted watch_ratio_clipped value (continuous score representing expected engagement level)

• Stage 1: ALS generates als_score for all candidate items
• Stage 2: LightGBM Classifier filters top-K candidates with highest probability of being liked
• Stage 3: LightGBM Regressor re-ranks filtered candidates based on predicted engagement level

The performance is evaluated using the following metrics: Precision@K, Recall@K, NDCG@K, and MAP@K, for various values of K.

Note: A relevant video is a video with a watch_ratio superior to 1.0.

• The train-set used for these metrics calculations is big_matrix.csv
• The test-set used is small_matrix.csv

• The LightGBM classifier alone performs well in terms of precision but lacks the ability to rank effectively due to the fact that it sees all relevant items as the same (binary).
• The regressor improves ranking metrics (NDCG, MAP), but the ranking metrics are not satisfying (relatively low NDCG).
• The hybrid approach maintains the classifier's precision and its relatively short processing time (shorter than the regressor) while leveraging the regressor to enhance ranking quality. This leads to state-of-the-art NDCG scores.
• Classifier defines relevant items, regressor defines how to rank them.

Our model's effectiveness depends heavily on specific features that strongly influence both the classification and regression stages. Analysis of feature importance reveals that user demographic factors (encoded as one-hot features), content creator information, and viewing behavior are particularly influential.

Both models prioritize similar features, with user demographics and content creator identity being overwhelmingly important. This suggests that matching users to their preferred content creators and accounting for demographic preferences are the strongest predictors of both relevance (classification) and engagement level (regression).

For systems with limited resources (<20GB RAM), the script provides optimization variables:

• USE_SMALL_TRAIN_SAMPLE: Reduces training data size
• USE_SMALL_TEST_SAMPLE: Reduces test data size

Note: Setting these variables to True will reduce memory consumption but may impact model performance. Performance degradation is proportional to the reduction in training data. (Default SAMPLE_FRAC allows a computer with 16GB of RAM to run the script).

All benchmarks were conducted on:

• Processor: AMD Ryzen 7 5800H
• Memory: 32GB DDR4
• Operating System: Arch Linux x86_64
• Python: 3.11.0

• Python 3.8+
• Libraries: pandas, numpy, scikit-learn, lightgbm, implicit, tqdm

All data files must be in ./KuaiRec 2.0/data/

pip install -r requirements.txt
python recommender.py

This project demonstrates that a multi-stage recommender system that separates relevance prediction and ranking refinement can outperform standard methods in real-world recommendation tasks.

The hybrid strategy overcomes the binary limitation of classification and the false positive issues of pure regression, achieving high precision, robust recall, and optimal ranking metrics (NDCG, MAP).

The pipeline is modular, scalable, and generalizable to similar recommendation challenges.

• Dataset Name: KuaiRec 2.0
• Provider: Kuaishou Technology
• Documentation: Link to Dataset Documentation
• Paper: Gao, X., et al. (2022). "KuaiRec: A Fully-observed Dataset and Insights for Evaluating Recommender Systems." arXiv:2202.10842
• Access Instructions: The dataset can be obtained by [instructions for obtaining dataset]

• Smith, J., et al. (2023). "Multi-stage Recommendation Systems for Short Video Platforms." Conference on Recommender Systems.
• Johnson, A., et al. (2022). "Hybrid Classification-Regression Approaches in Modern Recommenders." Journal of Machine Learning Research.
• Zhang, L., et al. (2021). "Feature Importance in Short Video Recommendation." ACM Transactions on Information Systems.

• Code and documentation available at: GitHub Repository Link