Embedded Files
Explainable Ai (XAI)-
Team: Tiffany Berry, Milan Ta, Devin Chen, Jennifer Tsai, & Melissa Yee
Faculty: Prof. Sanghamitra Dutta
Graduate Students: Faisal Hamman
I4C Teaching Assistant: Sayee Naresh Kumar
Project Overview
Project Overview
Our Question:
What is features are contributing to a model's prediction or answer? How can we visualize it?
What is Explainable AI?
The process of taking machine learning algorithms to then visualize and explain what contributes to the resulting model's prediction or answer.
Brief Summary of Chosen Dataset:
A supervised dataset that categorizes an adult's annual income based on the parameters of whether or not one makes over or under 50K. Examples of features include work class, education, relationship status, and occupation.
Building and Testing the Model
Building and Testing the Model
Step 1: Process the Data
Some data, such is the case with Strings, are incredibly difficult if not impossible in some cases for neural networks to process. Using methods such as dummy variables can transform unusable data into a numerical form, allowing for use in the model.
Step 2: Build and Train the Model
After processing the data, we can use sklearn library to build several models (regression, neural network, decision tree), and train the model on a training set.
Step 3: Check the Accuracy
Using our test data, the neural network was able to classify the adults with 83% accuracy, 94% precision, and 85% recall.
Applying explainability techniques
Applying explainability techniques
*example data
Lime (Local Interpretable model-agnostic explanations)
Creates a linear approximation of the decision boundary around a datapoint then applies it to a visual model.
***example data
Shap (SHapley Additive exPlanations)
By applying the black box model concept, Shap transforms the data into a more easily understandable visual representation.
Fairness of Our Model
Fairness of Our Model
Models may exhibit bias toward some specific groups based on gender, race, etc.
In order to eliminate bias from AI, it requires explainability.
Accuracy Rate
Accuracy Rate
How well did the model accurately predict the true outcome? In other words, whether or not a person make more or less than 50K.
Selection Rate
Selection Rate
For each group, what are the chances they were "selected"? In other words, every group should have an equal chance at selection. Ex: College Admissions
Future Impact / Conclusions
Future Impact / Conclusions
Assessing models associated with higher risks and implications - ex: detecting disease
Identifying and correcting possible unfairness or bias across specific groups in Machine Learning.
Takeaway: The importance of understanding what is actually happening behind the scenes.
Page updated
Report abuse