Handle missing and unavailable predictive features

Scenario

You are building a model to predict whether a user will successfully file taxes (binary label success) for a TurboTax-like product.

One of the most predictive features is:

• session_count = cumulative number of sessions a user has had in the product.

However:

• In the training dataset, session_count has many values that are 0 and many that are missing.
• In production, stakeholders claim that session_count is not available at scoring time (i.e., when you need to make the prediction), even though it appears in the schema.
• Exploratory analysis shows session_count is negatively correlated with success.

Questions

1. Data quality / missingness: How would you investigate why session_count is often 0 or missing, and how would you treat these cases during modeling?
2. Training-serving skew: If session_count is not available at inference time, what are your options? How do you decide whether to (a) drop it, (b) engineer a proxy, or (c) change the prediction timing / problem definition?
3. Interpretation: Provide at least two plausible explanations for the negative correlation between session_count and success (including an “opposite viewpoint”), and describe what additional data or analyses you would use to validate/refute each explanation.

Constraints / expectations

• Assume you have standard product event logs available in principle (page views, step completions, timestamps), but instrumentation may be imperfect.
• Your answer should cover: leakage risk, feature availability, and how you would communicate tradeoffs to stakeholders.