This criteria distinguishes Whether Interpret-ability is achieved by restricting the complexity of the machine learning model (intrinsic) or by applying methods that analyze the model after training (post hoc)

Permutation feature importance is, for example, a post hoc interpretation method

Contrastive explanation: explaining only the abnormalities Rather than explaining all features, explaining only 3 or 4. Social context –> who are you going to give answers to

Whether explanation is required or not is also an imp question.

An explanation can be portable: applicable from one to other or not.

We tend to ask questions with why this answer and why not that one? That kind of explanation.

People need explanation about results to know how to get different results.

Mostly, the events that have the lower chance of occurence will be more significant.

Fidelity of any explanation is also significant since for e.g. if model tells adding a balcony increases house’s price, it should be applicable all the time.

Sometimes the model needs to be homoscedastic for all data points meaning have same variance for all predictions : may not be useful though since houses with higher prices can have more variance compared to house with low prices.

R squared, ModifiedR squared are used to explain the effect of categorical variables and their variance. Takes value from 0 to 1.

Feature’s weight plot is used to show significance of each feature and their variance. It can be told from the feature wt plot which features don’t impact the result, which does positively or negatively. Which ones are more random in using for prediction and have lower confidence and so on.

Effects Plot will have an additional contribution multiplied to weights. Similar to normalization Uses a box plot

Instance Plot: Now we calculate each feature’s contribution and multiply with the instance specific data and find total contribution to the result. Now the discrepancies in the actual result explanations and the current predictions help us understand how single feature can impact the total. P.S. We assumed all other features being constant while explaining each feature’s contribution in the previous case. So, it’s important to do this to understand what kind of features directly impact presence or absence of other features. e.g. If a day is rainy , then other good features being true might not contribute much to the original result that is negative.

Encoding of categorical features:

Several techniques

Techniques determine the significance shown by each category

Treatment coding