advanced

Applied ML

Recommender systems, time-series forecasting and A/B testing — three applied-ML surfaces where the model is the easy part and honest evaluation is what interviews actually probe.

ml-applied

Practice: 11 questions →

Sections

Detailed explanation

Applied ML

Recommender systems, time-series forecasting and A/B testing look like three unrelated topics, but interviews hit them with the same question: how do you know your number is real? In each one the model is comparatively easy and the danger lives in the evaluation, because the natural, obvious way to measure performance quietly leaks the very thing you are trying to estimate.

A recommender scored on its own impression log only rewards repeating yesterday's ranking. A forecast validated with random k-fold has trained on the future. An A/B test watched every morning crosses the significance line by chance long before the effect is real. Each layer below teaches the mental model first, then names the trap that turns a confident offline win into a flat — or negative — result in production.

Recommender systems — collaborative vs content-based filtering, matrix factorization, the two-stage candidate-generation + ranking architecture, implicit feedback, cold start, and why offline ranking metrics disagree with online tests.
Time-series forecasting — stationarity and differencing, why random k-fold leaks the future, rolling-origin backtesting, seasonality, classical models vs GBDT, multi-step strategies, and metrics that survive near zero.
A/B testing — hypotheses and power, sample size and MDE, p-values vs confidence intervals, why peeking inflates false positives, CUPED variance reduction, and validity checks like SRM.

Why it matters

In production ML the modelling is rarely the bottleneck; the trap is an evaluation that quietly leaks the answer. A recommender that scores its own training log, a forecast validated with random k-fold, or an A/B test peeked at until it turns significant all report a win that does not exist. This topic teaches the mental model behind each surface and names the leakage traps interviewers use to separate people who have shipped from people who have only read the paper.

Applied ML

Recommender systems — collaborative vs content-based filtering, matrix factorization, the two-stage candidate-generation + ranking architecture, implicit feedback, cold start, and why offline ranking metrics disagree with online tests.

Time-series forecasting — stationarity and differencing, why random k-fold leaks the future, rolling-origin backtesting, seasonality, classical models vs GBDT, multi-step strategies, and metrics that survive near zero.

A/B testing — hypotheses and power, sample size and MDE, p-values vs confidence intervals, why peeking inflates false positives, CUPED variance reduction, and validity checks like SRM.