Description:

Description Interpreting a machine learning model and its outputs is crucial for data-driven products This talk will focus on explaining the outcomes of the machine learning model using InterpretML

Objective This talk will explain how interpretability can be achieved and it will also discuss one of the open source libraries.

Abstract

Part 1: What's the problem? : (2-3minutes)

Understanding the result of a model prediction is a diffi cult task for Stakeholders Stakeholders are not only interested in what is predicted but also in rationale behind the prediction. Thus it becomes critical to interpret the model and associated outcome as it helps in : a. Debugging the model b. Detect bias c. Build a trust between humans and models d. Sanity check for high risk predictions.

Part 2: How do we solve it? (5 minutes): Understanding techniques available for model explainability such as SHAP, LIME, Partial dependence plots.

Part 3: Associated Libraries (2 minutes): This section will reference all available open source libraries for interpretability such as InterpretML, Alibi etc.

Part 4: Understanding InterpretML (15-18 minutes) This section will cover InterpretML and its associated benefits, such as :-

1. InterpretML is a simpler ,easy to understand library managed by Microsoft Research team
2. It comes inbuilt with Azure SDK
3. It supports Glassbox models and Blackbox Explainers
4. Glassbox models are models within InterpretML which makes interpretability easier (example of the models: EBM, Decision Tree etc.)
5. These models make general additive models explainable and accurate.
6. Blackbox explainers explains the models from external library using only input and output values.
7. They provide explainers such as SHAP,LIME, Partial Dependence Plots and Morris Sensitivity

Part 5: Q & A (3-5 minutes)

Prerequisites:

Basic Understanding of Python and Machine learning

Video URL:

https://www.youtube.com/watch?v=4sjPL5ILNK4

Content URLs:

Content :

1. https://github.com/vedpd/interpretML/blob/main/InterpretML%20docs_050823.pdf

Code example for demonstration :

2. https://github.com/vedpd/interpretML/tree/main

Code repo interpretML ibrary :

3. https://github.com/interpretml/interpret

Speaker Info:

Ved Prakash Dwivedi is an NIT Jamshedpur graduate. He is currently a Senior Technical Lead Data Scientist in the fintech industry, specifically at Paytm. Prior to Paytm, he worked at big 4 consulting firms. Throughout his extensive ten years of experience, he has tackled various projects within the financial field. He has worked across banking(credit risk modeling), insurance, and human capital-related projects as part of his career.

Speaker Links:

Blogs :

1. https://medium.com/analytics-vidhya/building-pivot-table-on-excel-pandas-sql-e5f97053bbec
2. https://vedprakash-nitjsr.medium.com/demystifying-spark-session-configurations-unleashing-the-power-of-apache-spark-2b7141ff8540

Medium profile :

• https://medium.com/@vedprakash-nitjsr

LinkedIn : https://www.linkedin.com/in/vedprakashdwivedi/