Evaluate OutlierHandler Class for Code Quality and Testing
Quick Overview
Evaluates code-review judgment for OutlierHandler and imputer preprocessing classes in ML pipelines. Strong answers explain fit-transform separation, leakage prevention, API quality, edge cases, and testing.
Evaluate OutlierHandler Class for Code Quality and Testing
Company: Capital One
Role: Data Scientist
Category: Machine Learning
Difficulty: medium
Interview Round: Onsite
##### Scenario
Code-review of two Python preprocessing components: an OutlierHandler class and three Imputer classes
##### Question
Give a high-level summary of what the OutlierHandler class does. Why is it beneficial to separate `fit` and `transform` into two methods in this context? Point out any coding-style or maintainability issues you observe in the file (naming, docstrings, magic numbers, etc.). Write the single most critical unit test you would add for the OutlierHandler. For the three imputation classes, describe their overall purpose and identify at least two style problems (e.g., use of `from numpy import *`).
##### Hints
Relate your answers to the scikit-learn transformer API, unit-testing best practices, and PEP-8.
Quick Answer: Evaluates code-review judgment for OutlierHandler and imputer preprocessing classes in ML pipelines. Strong answers explain fit-transform separation, leakage prevention, API quality, edge cases, and testing.
Evaluate OutlierHandler Class for Code Quality and Testing
Capital One
Jul 12, 2025, 6:59 PM
mediumData ScientistOnsiteMachine Learning
66
0
Code Review: OutlierHandler and Imputer Classes
You are given a Python module that implements one OutlierHandler class and three Imputer classes for preprocessing tabular data. The classes appear intended for machine-learning pipelines, but the style and test coverage are mixed.
Assume OutlierHandler detects outliers per feature using rules such as IQR capping or z-score thresholds, and the imputer classes learn statistics during fit and fill missing values during transform.
Constraints & Assumptions
Treat the classes as stateful preprocessing components for train/validation/test pipelines.
Focus on code quality, API design, correctness, leakage prevention, and testing.
Do not assume access to production internals beyond the class behavior described.
Discuss both behavior and maintainability.
Clarifying Questions to Ask Guidance
Should the classes follow scikit-learn's estimator API exactly?
Are inputs NumPy arrays, pandas DataFrames, or both?
Should transforms preserve column names, dtypes, indexes, and missing-value markers?
Are outliers capped, removed, replaced, or flagged?
Part 1 - OutlierHandler Summary
Provide a high-level summary of what the OutlierHandler class does.
What This Part Should Cover Guidance
Explain that it learns per-feature thresholds during
fit
.
Explain that
transform
applies stored thresholds to new data consistently.
Mention strategies such as IQR, z-score, capping, masking, or replacement.
Connect the class to ML preprocessing pipelines.
Part 2 - Fit and Transform Separation
Explain why separating fit and transform into two methods matters.
What This Part Should Cover Guidance
Prevent data leakage by learning statistics only on training data.
Ensure validation, test, and production data are transformed consistently.
Support pipelines, cross-validation, serialization, and reproducibility.
Clarify behavior when
transform
is called before
fit
.
Part 3 - Code Quality and Testing
Evaluate the code quality and propose tests.
What This Part Should Cover Guidance
Review API consistency, input validation, error handling, documentation, naming, type handling, and edge cases.
Test missing values, constant columns, all-null columns, mixed dtypes, unseen categories, extreme values, small samples, and transform-before-fit errors.
Test shape preservation, no mutation of inputs, deterministic output, and parity across train/test.
Include unit tests and integration tests inside a simple ML pipeline.
Follow-up Questions Guidance
How would you make the classes compatible with scikit-learn pipelines?
What bug would you expect if thresholds are recomputed during transform?
How would you test the behavior on a DataFrame with nonnumeric columns?
