topic-pred

Solution

Content-to-Topic prediction system for K-12 educational materials using semantic embeddings.

Overview of Approach

This solution uses a semantic embedding similarity approach to match content to topics:

1. Text Preprocessing: Content titles and descriptions are concatenated and cleaned
2. Embedding Generation: Uses paraphrase-multilingual-MiniLM-L12-v2 sentence transformer to encode both content and topics into dense vector representations
3. Similarity Matching: Computes cosine similarity between content and topic embeddings
4. Filtering & Ranking: Returns top-k topics (default k=3) with similarity scores above threshold (default 0.3)

Rationale: Embedding-based approaches capture semantic meaning effectively and generalize well to new content without requiring extensive training data. The multilingual model ensures robustness across varied educational content.

Metrics: since the topics were duplicated, and sometimes we might predict the correct topic but not the correct id. I opted for establish a match if the cosine similarity is greater than 0.9

Code Access Point

• First you need to download the data in the 'data' folder (find them in https://www.kaggle.com/competitions/learning-equality-curriculum-recommendations.). 3 files are needed:

  • content.csv
  • correlations.csv
  • topics.csv

• Run the notebook topic_prediction_training_refactored.ipynb

  • Notebook contains full pipeline: data loading, preprocessing, model training, and artifact generation
  • Run cells sequentially from top to bottom to regenerate artifacts

For predictions:

• predict_template.py → TopicPredictor class (it is only using title and description fields)
• Uses pre-generated artifacts in artifacts/topic_predictor_direct_model.pkl

Supporting utilities:

• preprocess_methods.py: Text preprocessing and embedding generation
• predict_method.py: Core prediction logic using cosine similarity
• io.py: Data loading helpers
• defaults.py: Configuration constants

Metrics

Metric	Score
Recall	0.2638
Precision	0.1655
F1-Score	0.2034

Note: Metrics computed on validation set using correlations.csv as ground truth. Remember that a match is the combined_text similarity greater than 0.9 between pred and true values

What Would You Have Done With More Time?

1. Hybrid Approach: Combine semantic embeddings with metadata features (content_kind, topic_category, keywords) using a learned weighting scheme
2. Fine-tuning: Domain-adapt the sentence transformer on K-12 educational corpus
3. Evaluation: More comprehensive metrics including NDCG, MAP, and per-category performance analysis
4. Data Augmentation: Leverage topic hierarchies and content language fields for improved matching

Quick start

1. Create a virtual environment and install dependencies (python <=3.12):

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Compatibility note

Some binary packages (notably PyTorch) are compiled against specific NumPy ABI versions. If you see errors like

RuntimeError: Numpy is not available

or messages about a module compiled with NumPy 1.x not running with NumPy 2.x, pin NumPy to a 1.x release before installing other packages:

# from project root
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install "numpy<2,>=1.21"
python -m pip install -r requirements.txt

This ensures binary compatibility between PyTorch and NumPy. The requirements.txt in this repository already pins NumPy to <2.