MLE CodeSignal OA (MCQ + NN + Classical-ML Coding)

MLE candidates get a 70-minute CodeSignal bundle: 6 multi-select ML conceptual questions, 1 numerical fill-in (forward-pass through a small NN), and 3 coding problems mixing one warm-up algorithm with two from-scratch classical-ML implementations (bagging, k-means, decision tree). Webcam on; screen-share recorded; the AI assistant is not the explicit collaborator here. Requirements Format (70 minutes total, ~10 items): Multi-select ML concepts (6 questions) Calibration / Precision / Recall / FPR trade-offs (e.g. "which thresholds satisfy recall ≥ X and FPR ≤ Y"). Choosing the loss for a logistic regression model. How to simultaneously increase bias and reduce variance. Correct considerations when picking ensemble learning. PCA vs LDA — when to use which. One additional rotating item. Multi-select means all correct answers must be selected; partial-credit is not awarded. Numerical fill-in (1 question, 4-decimal answer) Compute the output of a small 2-layer neural network given weights, biases, and an input. Activations are typically ReLU and Sigmoid; do the arithmetic by hand or with a scratch calculator. Coding (3 problems) Warm-up: array / string algorithm (e.g. longest consecutive run of identical characters returning the rightmost when tied; find local maxima). From-scratch bagging classifier: given a base classifier interface, implement bootstrap resampling, fit, and predict. From-scratch k-means: distance function is provided; implement assignment + centroid update + convergence loop. Decision-tree variant: implement a purity function (formula provided), information gain, and predict. Each problem allows multiple submissions; the harness shows how many of its (~200) hidden tests pass. Notes MCQ items are unforgiving because they are multi-select. Read every option; eliminate-then-confirm. The PCA-vs-LDA item in particular has historically tripped people who only remember the headline trade-off. NN fill-in: lay out the matmul on scratch paper, mind the activation order. Off-by-one on ReLU vs Sigmoid placement is the most common cause of a wrong 4-decimal answer. Bagging / k-means / decision tree are mechanically simple but interface-fiddly. Pre-write a template for each in your interview language so the round becomes a copy-tweak exercise. CodeSignal's Classification Algorithms and Metrics course includes a Decision Tree implementation walkthrough that matches the test interface closely. The 70-minute budget is workable if (and only if) you do not spend more than 25–30 minutes on the MCQ + NN block. Move on from any MCQ you are uncertain on after 90 seconds. Preparation Drill the canonical from-scratch implementations: bagging (bootstrap → fit base → majority-vote predict), k-means (init → assign → update → repeat-until-stable), decision-tree (best-split by information gain → recurse → predict by path). Refresh PCA vs LDA, ensemble-bias-variance, and calibration in a 20-minute review block before the round. Pre-build a 10-line scratchpad function that does dense matmul + activation in your language so the NN fill-in is mechanical.