Curriculum Recommender

Overview

Developed a machine learning solution for the Kaggle "Learning Equality - Curriculum Recommendations" competition. The goal was to match educational content with curriculum topics to help students in underserved communities access relevant learning materials.

Stage 1: Retrieval

Data Strategy

• Used source category as part of training data for all folds, excluded from validation to prevent data leakage
• Applied StratifiedGroupKFold on non-source category topics (group=topic, target=language)
• Split correlations rather than directly splitting topics/content for train/validation

Model & Training

• Base Model: sentence-transformers/all-MiniLM-L6-v2 (compact size for faster experimentation)
• Training Approach: SimCSE contrastive learning
• Grouped by language then shuffled to ensure batches contain single language (small recall improvement)
• Hard negative mining significantly improved model recall
• Multi-round boosting with non-source data improved top50/100 recall for 2 rounds

Model Ensemble

Combined predictions using voting + average cosine similarity ranking for topic-content pairs.

Stage 2: Ranking

• Binary classification model (softmax + CrossEntropyLoss) for re-ranking
• Used threshold tuning to handle sample imbalance
• Tokenizer with pair format to ensure consistent text1/text2 lengths, reusing Stage 1 prompts
• Added recall order as text feature (small improvement observed)
• CV strategy aligned with LB evaluation strategy
• Larger models showed clear improvements

Key Insights

• Non-source Topics Critical: Distribution matches LB test set, essential for good generalization
• Hard Negative Mining: Most effective strategy for improving retrieval performance
• Multi-round Boosting: Iterative training with hard examples from previous models
• Model Size Trade-off: Increasing max_len/batch_size beyond 128/192 showed diminishing returns; larger models might have been more effective
• CV-LB Alignment: Consistent validation strategy crucial for reliable model selection

Results

• Achieved Bronze Medal with 76th place out of 1000+ teams
• Developed a scalable two-stage retrieve-and-rank pipeline
• Gained experience in handling multi-lingual educational content matching

Tech Stack