ADEO Dev x Data x PMS 2024

In search and recommendation engine, ranking products is a big challenge. With everyday new products arrival, catalog grows in vocabulary and diversity, making ranking problem even harder.
Transformers, the famous technology behind Large Language Models (LLM), with their impressive learning power and good inference speed, appears to be an attractive solution.
Transformers efficiently capture the semantic meaning of product designation and description, in addition to other common numerical and categorical product features.

In this talk, we will show how we can train a Transformer model to rank products from a keyword search, based on their related attractiveness score.
Then, unveiling the black box, we will demonstrate how words or features explain product relevancy.

Talk summary:
- Introduction to Transformers, BERT, and sentence-transformers python module
- A click model to compute product attractivity score related to a keyword
- Definition and training of a cross-encoder model to fit product score
- Measuring ranking quality with nDCG
- Understanding the model inner workings with SHAP value extraction

At the end, you will be able to create your own model, adapt it to your business case, drink some coffees during training time and evaluate its results.