Machine Learning / AI

In case of severe class imbalance, how do precision, recall, and PR-AUC behave?

What metrics did you use, how much data was there, and what was the class balance?

Tell us about yourself and the tasks you have performed.

If you can only communicate with a limited number of clients, how do you evaluate such a scoring model?

Was this a classification or time series task? How was time considered during training and testing?

What do you know about ROC-AUC?

How is the ROC curve calculated iteratively after ranking objects by score?

Tell us about one of the recent classification tasks and describe its pipeline.

What does PR-AUC show?

Do you know what a scoring model is?

Was there a checklist that the operator had to follow during the call, or was the check performed without it?

Tell us about an NLP task you performed.

Tell us about a clustering task.

Find all clients in January who had at least one sale after communication within 30 days.

You have a target and a verified feature store with about 5,000 features. How would you test it on this task?

Is a precision of 700 sales per million selected clients good in a rare conversion task?

Find all clients in January who had at least one sale after communication within 30 days. sales id customer_id amount sale_date [phone] communications id customer_id type comm_date partition_date 1 1 mail [phone] call [phone] call [phone] email [phone] sms [phone] #There is a list, you need to write a function that returns two lists, one with normal values, and the other with anomalies. Let's assume for now that an anomaly is a value that exceeds the average data = [100, 200, 150, 5000, 1300, 120, 700, 2000, 200]

How many features will you take from the feature store and how will you select them if all 5000 features are valid and without leaks?