Demystifying machine learning using lime
6 likes2,069 views
The document discusses the challenges of interpretability in machine learning models, particularly focusing on black box models. It introduces LIME (Local Interpretable Model-Agnostic Explanations) as a method to provide interpretable explanations for any classifier's predictions. An example of using LIME for a URL phishing detection classifier is provided, illustrating how it can evaluate and explain phishing probabilities based on website URLs.
Demystifying machine learning using lime
- 1. DEMYSTIFYING MACHINE LEARNING Alejandro Correa Bahnsen 1
- 2. BLACK BOX MODELS Machine learning models are often dismissed on the grounds of lack of interpretability. When using advanced models it is nearly impossible to understand how a model is making a prediction. 2
- 3. MNIST - ACCU VS # PARAMS Notebook to create the plot 3
- 4. LIME stands for Local Interpretable Model- agnostic Explanations, and its objective is to explain the result from any classifier so that a human can understand individual predictions LIME 4
- 5. LIME An interpretable representation is a point in a space whose dimensions can be interpreted by a human. LIME frames the search for an interpretable explanation as an optimization problem. Given a set G of potentially interpretable models, we need a measure L(f,g,x) of how poorly the interpretable model g∈∈G approximates the original model f for point x this is the loss function. We also need some measure Ω(g) of the complexity of the model (e.g. the depth of a decision tree). We then pick a model which minimizes both of these ξ(x) = argmin g∈∈G L(f,g,x)+Ω(g) 5
- 6. LIME 6
- 7. LIME EXAMPLE URL PHISHING DETECTION 7
- 8. URL PHISHING CLASSIFIER Objective: Evaluate phishing probability using only the web site URL 8
- 9. URL PHISHING CLASSIFIER Train a random forest 9
- 10. LIME EXAMPLE Fit lime explainer Explain an instance 10
- 11. LIME EXAMPLE Example Phishing URL Phishing probability 1.0 Url = http://login.paypal.com.convexcentral.com/Update/ab770f624342b07b71e56c1bae5d9bcb/ 11
- 12. LIME EXAMPLE Example Phishing URL Phishing probability 0.0283 Url = ...http://www.redeyechicago.com/entertainment/tv/redeye-banshee-ivana-mili 12
- 13. THANK YOU FULL NOTEBOOK IN HTTPS://GITHUB.COM/ALBAHNSEN/TALK _DEMYSTIFYING_MACHINE_LEARNING 13