ML Fundamentals

A fundamentals-heavy ML discussion built around modeling a feature x against a target y, then interpreting a plot of two overlapping class-conditional distributions to reason about separability, Bayes error, threshold choice, and cost-aware metrics. The round is conversational and branches into feature selection, model choice, evaluation, and cold-start follow-ups. ML Fundamentals A fundamentals-heavy ML discussion built around modeling a feature x against a target y, then interpreting a plot of two overlapping class-conditional distributions to reason about separability, Bayes error, threshold choice, and cost-aware metrics. The round is conversational and branches into feature selection, model choice, evaluation, and cold-start follow-ups. MLE ml-knowledge classification feature-engineering evaluation calibration verbal open-ended Frequency Low Last asked 2026-03-16 Stage phone-screen · onsite-coding ML Fundamentals Problem Overview This round is a fundamentals-heavy Machine Learning Engineer discussion. It typically starts with a simple supervised learning setup, then uses a plot of two overlapping class distributions to probe how well you understand modeling assumptions, decision boundaries, feature choice, and error trade-offs. The round is conversational rather than rigidly scripted. Expect the interviewer to give hints or feedback and to use your answer to branch into deeper follow-ups. Core Prompt You are given: a feature x a target label y The interviewer asks: how would you model the relationship between x and y? After that, they show a figure with two class-conditional distributions over the same feature space. The two curves overlap in the middle region. Use that figure to explain: whether the classes are linearly separable whether a single threshold is reasonable what the overlap means for Bayes error and unavoidable misclassification how you would choose a threshold in practice which metrics you would optimize if false positives and false negatives have different costs What The Interviewer Is Looking For Modeling Basics Clarify the prediction task first: regression if y is continuous binary or multiclass classification if y is categorical Start with a simple baseline before jumping to a more complex model Explain how the amount and shape of the data affect the model choice Interpreting The Overlap Figure The key point is not just "pick a threshold." The more complete answer is: if two class distributions overlap, the problem is not perfectly separable on that feature alone a threshold may still be a reasonable decision rule if the feature is one-dimensional and the conditional distributions are ordered the overlap region implies irreducible error if only this feature is used moving the threshold trades off precision and recall, or false positive and false negative rates the best threshold depends on the product goal, class prior, and error cost A strong answer goes one step further: if performance is inadequate, look for additional features that better separate the classes rather than over-focusing on the threshold itself. Common Follow-Ups Feature Selection The interviewer may repeatedly ask how you would choose or create better features. Be ready to discuss: which raw features are likely predictive how to identify leakage how to handle missing values scaling or normalization when needed encoding for categorical features whether interactions or nonlinear transforms could help how to evaluate feature importance or feature usefulness Model Choice You may be asked: when logistic regression is sufficient when tree-based models are a better fit how to think about linear vs nonlinear boundaries how to balance interpretability against predictive performance Evaluation Expect discussion around: train/validation/test split ROC-AUC vs PR-AUC for imbalanced data precision/recall/F1 trade-offs calibration and probability interpretation Cold Start The round may end with cold start questions, such as: how to make predictions for a new user with little or no history how to handle a new item with limited interaction data what fallback features or priors you would use when heuristics, popularity baselines, or contextual features are appropriate before enough personalized data arrives A Strong Answer Structure Clarify what y represents and whether this is regression or classification. Start with a simple baseline model and explain why. Interpret the overlapping distributions as evidence that one feature alone cannot perfectly separate the classes. Explain threshold selection in terms of business cost and metric trade-offs, not just geometry. Propose additional features that could reduce overlap. Discuss evaluation and how you would validate that the new features or model actually help. Close with practical cold-start strategies.