Ade Idowu - Hands on Intro to Developing Explainability for Recommendation Systems

Discover the importance of explainability in recommendation systems, and learn hands-on methods to develop and evaluate explainable recommendations, including association rules, matrix factorization, and post-hoc explanation methods like LIME.

Key takeaways

The goal of recommendation systems is to compute unknown ratings, and explainability is critical to understand the intuition behind recommendations.
Association rules can be used to explain recommendations by identifying patterns in user behavior.
Matrix factorization is a method to decompose user-item interaction matrices into lower-dimensional spaces, making recommendations more interpretable.
Model-based explainability couples the recommendation engine with an explainability model, providing insights into the predictions made by the model.
Post-hoc explanation methods, such as LIME, can explain black box models by generating explanations based on perturbations around a sample input.
Explainability is crucial in recommendation systems, and it’s no longer acceptable to build models that are opaque.
There are various metrics to evaluate the fidelity of explanations, including precision, recall, and relevance.
NLP-based models can generate text-based explanations, which can be visualized as word clouds or sentences.
Regularization is essential to ensure that the model is generalizable and not overfitting.
Matrix factorization is a fundamental technique in collaborative filtering, and it’s used to predict user ratings.
The first demo shows an example of a movie recommendation system based on user-item interaction data.
The second demo shows how to use the matrix factorization method to compute recommendations.
The third demo demonstrates the use of LIME to explain a black box model by generating explanations based on perturbations around a sample input.
The fourth demo shows how to use a visual style explanation to provide insights into the recommendations made by the model.
The fifth demo demonstrates the use of association rules to explain recommendations.
The sixth demo shows how to use a knowledge-based recommendation system to provide explanations based on external data.

Ade Idowu - Hands on Intro to Developing Explainability for Recommendation Systems

More talks