Fundamentals & Definitions
Explain the concept of outliers and their impact on a dataset.
What is outlier detection and why is it important in data analysis or machine learning?
What are outliers in a dataset, and how do they differ from noise or anomalies?
Why do outliers appear in data, and what are the common sources of outliers?
What is anomaly detection?
What are sentinel or placeholder values like 999 or -1, and why do they show up as outliers?
How does anomaly detection differ from noise removal?
What is the definition of an outlier and why is there no single formal definition?
What is the role of statistics in anomaly detection?
Can an outlier be the most interesting signal in the data rather than an error, and how do you decide whether a flagged point is a mistake, a rare-but-real event, or the thing you actually care about?
Types & Taxonomy Of Outliers
What is a distance-based outlier?
Can you explain the three main types of outliers — point (global), contextual (conditional), and collective — and give an example of each?
What is the difference between a univariate and a multivariate outlier, and can a point be an outlier multivariately even if it's not univariately?
Why should you detect outliers per group or segment rather than globally, and how does this relate to contextual outliers?
What is the difference between an outlier in the outcome, a high-leverage point in the predictors, and an influential point?
Statistical Methods
How do you detect outliers in a dataset?
How do you detect outliers using box plots, z-scores, or interquartile range (IQR)?
What is a boxplot and how does it detect outliers?
What is the 68-95-99.7 rule, and how is it applied in outlier detection for normally distributed data?
What are some common statistical methods for outlier detection?
What are quartiles and the Interquartile Range (IQR)?
How does the Interquartile Range (IQR) method, also known as Tukey fences, work for outlier detection, and what are its advantages and disadvantages compared to the Z-score method?
Why use IQR for outlier detection rather than another method?
Where does the 1.5 multiplier in Tukey's fences come from, and roughly what false-flag rate does it produce under a normal distribution?
Explain the concept of a Z-score and how it is used to detect outliers, including its assumptions, limitations, and the masking effect.
Explain the concept of a modified Z-score using the median and Median Absolute Deviation (MAD), and why it is considered more robust than the standard Z-score.
How do you choose among Z-score, modified Z-score, and IQR rules based on sample size, skew, tail weight, and the expected number of outliers?
Distribution & Normality Assumptions
Can you explain outlier detection methods that rely on assumptions of a normal distribution, such as the Z-score?
Can you discuss the impact of normality assumptions and methods for detecting normality when identifying outliers?
Why do heavy tails and skew break sigma-based outlier rules, and how does a log or Box-Cox transformation help?
Why is the IQR rule skew-sensitive, and what is the adjusted boxplot or medcouple approach for skewed data?
How do you tell an outlier apart from a legitimately heavy-tailed distribution where extreme values are expected?
Formal Statistical Tests
How does the Grubbs test detect an outlier, and what are its assumptions and limitations?
What is Dixon's Q test and when is it appropriate to use it?
What is Chauvenet's criterion and how does it decide whether to reject a data point?
Explain the generalised ESD test and how it differs from Grubbs when you expect more than one outlier.
When you apply an outlier test to every point in a dataset, what multiple-comparison problem arises and how do you account for it?
Distance & Density Based Methods
Why must you scale features before applying any distance-based outlier method?
How can DBSCAN be used for outlier detection, and what role do noise points play in this context?
Explain how k-Nearest Neighbors (k-NN) can be adapted for anomaly detection.
How do you interpret a LOF score, and what does a value near 1 versus well above 1 tell you?
Explain the working principle of Local Outlier Factor (LOF), and why is the concept of local density crucial for it?
Explain the concept of Mahalanobis distance and why it is preferred over Euclidean distance for multivariate outlier detection when features are correlated.
What does the chi-square cutoff mean when using Mahalanobis distance to flag multivariate outliers?
Mahalanobis distance has its own masking problem — how do the Minimum Covariance Determinant and elliptic envelope address it?
Model Based & Learning Approaches
What is the difference between supervised, unsupervised, and semi-supervised (novelty detection) approaches to outlier detection, and when would you use each?
When would you use One-Class SVM for outlier detection?
What does the contamination parameter control in detectors like Isolation Forest, and what happens if you set it wrong?
What is novelty detection, and how does fitting a detector on known-clean data differ from unsupervised outlier detection on contaminated data?
Describe the intuition behind Isolation Forest for detecting outliers, and why anomalies are typically isolated in fewer splits than normal data points.
Conceptually, how can PCA or autoencoder reconstruction error be used as an outlier detector?
Explain how One-Class SVM works for anomaly detection and its underlying principle for identifying novelties.
Robust Statistics
When would you use robust statistical methods like median over mean or MAE over MSE when dealing with outliers?
How do masking and swamping affect outlier detection, and how can they be mitigated?
Explain the concept of robust statistics and why robust methods are important in the context of outlier detection and modeling.
What is the breakdown point of a robust estimator like the Median Absolute Deviation (MAD)?
Handling & Treatment
Once you've identified outliers, what are the different strategies for handling or treating them, and what are the pros and cons of removing, capping/winsorizing, transforming, imputing, or keeping them and using robust methods?
Why is it generally recommended to investigate outliers first rather than immediately removing them?
What is percentile or quantile capping/trimming, and when would you use it for outlier treatment?
What is the importance of domain and business rules in identifying and handling outliers?
What are the analytic and ethical risks — including survivorship bias — of deleting inconvenient data points?
You find outliers in your dataset — walk me through what you do, step by step.
Evaluation Thresholds & Validation
How does the asymmetric cost of a false alarm versus a missed outlier influence where you set your threshold?
Given that outliers are often rare and unlabelled, how would you evaluate the performance of an outlier detection model, and why is accuracy often a misleading metric?
How do you choose an appropriate threshold or contamination rate for an outlier detector?
Why is precision-at-k, recall, or PR-AUC preferred over ROC-AUC when evaluating an outlier detector under extreme class imbalance?
How can you validate an outlier detector using injected or synthetic anomalies when you have no labels?
Why and how would you ensemble multiple outlier detectors, and what is the role of score normalisation when combining them?
Why must you fit an outlier rule or threshold on the training set and apply it to test, and how does re-fitting on test cause leakage?
Method Selection & Data Challenges
Can you list and explain different conceptual methods for outlier detection, such as statistical, proximity-based, and clustering-based approaches?
How can a single outlier create or destroy an apparent correlation between two variables?
What are the key trade-offs to consider when choosing among different outlier detection methods (statistical, distance-based, density-based, model-based)?
How does the curse of dimensionality affect distance-based outlier detection methods?
How would you approach detecting outliers in time-series data, considering temporal dependencies?
How do you handle high-dimensional data in anomaly detection?
What is distance concentration, and why do subspace or feature-bagging methods help detect outliers in high dimensions?
Why does an IID outlier rule misfire on trending or seasonal time-series data, and what temporal methods address this?