Machine Learning (24%)

DY0-001 · 72 questions

• A logistics model of missed deliveries reports high accuracy. Missed deliveries are rare, and the model never predicts a miss. Why can that accuracy look excellent?
• A retail clothing chain fits a weekly demand model that matches its training weeks almost perfectly, yet error jumps on a random holdout drawn from the same season. Which learning problem does that gap indicate?
• A retail grocery chain fits a linear model of weekly sales on many promo flags and puts an L1 penalty on the coefficients. What can that penalty do that an L2 penalty does not?
• A utility outage model stays far from its training weeks and equally far from a later holdout month. The form is a straight line through a clearly curved load pattern. Which problem is that?
• A logistics team scores a delivery-time model with k-fold validation on one historical table of completed routes. What does that procedure do with the rows?
• A media lab standardizes each viewing-hours feature using the mean and spread of the entire table, then splits the rows into training and holdout. What went wrong?
• An electric transmission company adds many weakly related sensor features to a nearest-neighbor model of transformer faults, and distances between units start to look almost the same. Which description names that problem?
• A regional power cooperative fits a demand model until the error on the weeks used to fit it is tiny, but a later week the model never saw shows a much larger error. Which figure should decide whether the model is ready to use?
• A hospital will deploy a single readmission model. On held-out discharges, a short list of clinical fields ties a model whose many extra fields did not improve that holdout score. Which model should go into production?
• A home-improvement retailer trained a demand model on markdowns from small discounts up to a moderate cap. A buyer wants predicted demand at a markdown deeper than every discount in that training data. What is that prediction?
• A parcel network must tune many hyperparameters of a delivery-time model, but a full grid of every combination is too large to finish and only a limited number of trials will fit in the schedule. Why is random search the better use of those trials?
• A mortgage lender's training file has far more loans that stayed current than loans that charged off. Before fitting a classifier, an analyst applies SMOTE to the numeric features. What does SMOTE do with the charge-off cases?
• A utility's outage model still sees the same mix of weather and load features as last year, but those features no longer line up with outage length the way they did. Storms that used to predict long outages now predict short ones after the grid was rebuilt. What changed?
• A factory keeps two scrap models. Shift leads can read the first model's risk factors directly from its coefficients. The second model is a black box, and a separate method explains individual scores only after that model has produced them. How do the two explanation styles differ?
• A retailer's return model was trained when most orders were in-store pickup. Ship-to-home orders have since become the common case, so the feature mix being scored no longer matches training, while the link from a given basket to a return has not been shown to change. What kind of drift is described?
• A logistics marketplace recommends spare carriers to shippers. The score for a shipper comes from which carriers similar shippers have booked, not from that shipper's fleet size, region, or contract type alone. Which method is the marketplace using?
• A hospital risk model flags one patient for sepsis review, and the chart note lists the few vital signs that pushed that single score upward. A quality officer treats the note as a description of how the model behaves for every patient on the ward. What does the note actually provide?
• A retailer's recommender suggests accessories from past co-purchases. A brand-new jacket is listed with no orders, no clicks, and no co-purchase history, so the co-purchase score has nothing to read. What problem is this?
• A hospital lab fits a line through assay readings. Some readings come from instruments whose measurement error is known to be very small, and others come from instruments whose measurement error is known to be large. The team wants the more precise readings to pull harder on the fitted line. Which method does that?
• A retail model must estimate the probability that a shopper returns, and the outcome is coded yes or no. A straight-line fit on that flag produces some fitted values below zero and some above one. Which model is the right replacement?
• A hospital team fits ordinary least squares for length of stay on age, admission type, and a severity score. Against fitted values, the residuals stay centered near zero, with a tight vertical band at low fitted values and a much wider fan at high fitted values. Which ordinary least squares assumption does that pattern violate?
• A utility demand model includes many weather features and calendar features. The team wants the coefficients pulled toward zero, and it wants every one of those features to remain in the model. Which penalty matches that goal?
• A utility models whether a customer misses a bill, and the outcome is yes or no. One candidate maps the linear score to a probability with a logistic link, and the other maps that score with a standard-normal cumulative link. Which statement describes the two models?
• An agency models grant-processing days with ordinary least squares, using staff count, backlog size, and a complexity score as features. The team must state the assumption about the mean of processing days given those features. Which statement is that mean assumption?
• A parcel carrier studies the rule that a fragile mark implies extra padding. Among shipments already marked fragile, the planner wants how often extra padding is present too. Which association measure is that?
• A hardware retailer studies the basket rule that paint implies brushes. The team wants the share of all baskets in which paint and brushes occur together, not how often brushes follow paint. Which association measure is that share?
• A freight-fraud screen scores shipments from several package attributes that move together. The model multiplies a likelihood for each attribute and does not estimate how those attributes covary inside a class. Which assumption is the model using?
• A grocery chain separates shoppers into two loyalty tiers from the same purchase features. The analyst assumes both tiers share one covariance, so only the typical shopper in each tier sits in a different place. Which classifier is built on that assumption?
• A news site mines the rule that watching a trailer implies starting a subscription. Confidence looks strong, but subscriptions are already common among all visitors. The team wants to know whether the rule beats that base rate. Which measure answers that?
• An electric utility wants a classifier that estimates a separate covariance for storm-related outages and for equipment-related outages, because spread and correlation differ by class. Which method estimates those separate covariances?
• A bank fraud desk builds a naive Bayes classifier on flags that are clearly related, such as an overseas swipe and a high amount on the same card. A reviewer says the ranking is useless because independence is only approximate. Which conclusion is sound?
• A retail grocer grows one decision tree until every training leaf is pure, including leaves that hold a single receipt. Training accuracy is far above holdout accuracy. What did that very deep tree do?
• A retail pharmacy chain fits a linear regression of weekly sales on many promotion flags and wants some coefficients driven exactly to zero so those flags drop out of the equation. Which penalty does that inside the regression?
• A logistics railroad wants a linear model of late arrivals that shrinks coefficients and can also set some of them exactly to zero. Which penalty is that mix?
• A telecom carrier's churn model is a logistic regression, and the positive class is churn. The coefficient on the number of support calls is positive. What does that positive coefficient mean?
• A factory grows a classification tree on defect types. At one node, half the parts are scratches and half are dents. Which quantity is high because that node is still mixed?
• A retail loss-prevention team uses gradient boosting to flag risky transactions, and later trees start fitting one-off tickets that never repeat. Which adjustment addresses that chase?
• A grocery retailer trains a bagged ensemble of trees to forecast daily store demand. A new analyst thinks every tree is fit on the identical full table. What does bagging actually give each tree?
• A hospital team grows a decision tree to flag which admissions are likely to need an intensive care bed. At one node, entropy falls after a split on arrival hour. What does that drop measure?
• A factory tree separates good batches from bad ones by splitting a pressure feature at a threshold. An engineer then applies a strictly increasing rescaling to that feature and fits again. What happens to the rows that split sends to each side?
• A logistics team builds a boosted sequence of trees to predict which shipments will miss the delivery window. How does the next tree in that sequence get its training job?
• A power utility builds a random forest to predict which feeders will overload on hot days. A colleague believes each tree inspects every weather and load feature at every split. What does a random forest actually do when it chooses a split?
• A utility builds an ensemble for feeder-fault risk by training many trees on separate bootstrap samples and averaging their scores. A colleague says each tree is fitting the residual left by the previous tree. Which description of this ensemble is right?
• A city permitting agency compares one deep decision tree with a random forest for flagging applications that need a manual review. Across resamples of the same training pool, the single tree's flagged set jumps around while the forest stays steadier. Why is the forest more stable?
• A hospital adds rounds to a boosted tree model that predicts which outpatient visits will be missed. Training error keeps falling, but error on a held-out week gets worse after a point. What is happening as those later rounds are added?
• A retail model must separate coupon responders from non-responders using two purchase features. Responders sit in two opposite corners of the scatter, so no single straight cut isolates them, and a lone perceptron keeps missing one corner. What limit explains that miss?
• A utility network scores boiler-load deviation with a hidden unit that uses a ReLU activation. The linear sum into that unit is negative on this reading. What does that ReLU unit output?
• A hospital model predicts which patients return within a month from two lab values. Returning patients sit above a diagonal line in the scatter, not on one side of a single vertical or horizontal cut, and a shallow decision tree misses them. Why does that tree need several splits to follow the diagonal?
• An apparel retailer's recommendation network memorizes training clicks. The team wants a regularizer that randomly turns some hidden units off on each training step. Which method is that?
• A logistics network's hidden activations change scale from one mini-batch of shipment scans to the next. The team wants each mini-batch recentered and rescaled from that batch's own mean and spread. Which method does that?
• A utility classifier assigns each outage one of several mutually exclusive causes and must turn the raw scores into shares that add to one. Which output behavior does that?
• A media captioning model keeps lowering its training loss, but validation loss has risen for several checks in a row. The team wants training to halt when validation loss stops improving. Which practice is that?
• A logistics network with several layers scores delivery exceptions, and the team needs the loss to assign blame to an early weight by passing back through the later layers. Which procedure does that?
• A grocery chain trains a deep network to score whether a shopper will redeem a coupon, and the hidden layers use a sigmoid. After several layers, updates to the earliest weights shrink to almost nothing. What is the sigmoid doing to those updates?
• A hospital teaching lab wants synthetic chest films. One network proposes images and a second network judges whether each image looks real or fake, and the two train against each other. Which architecture is this?
• A utility forecasts hourly load from a long stretch of meter readings. A plain recurrent net updates one hidden state each hour and has lost the early months by the time winter arrives. Which design keeps a gated memory across that long sequence?
• A retail chat assistant reads a long return-policy thread. A recurrent model walks that thread one token at a time and slows as the thread grows. The team switches to a design that relates each token to the others in one pass. Which design is that?
• A logistics team trains a network that scores late-delivery risk. Plain stochastic gradient descent uses one learning rate for every parameter. The team switches to a method that adapts each step from the gradient's first and second moments. Which method is that?
• A utility wants a short code for each feeder's daily load curve and trains a network so the original curve can be rebuilt from that code. Which model is this?
• A bank reviews photos of paper checks to flag a small altered-ink pattern that can sit anywhere on the check. The same local mark should be recognized in the corner or in the middle. Which network design matches that job?
• A logistics team runs k-means on delivery stops for many cluster counts and plots within-cluster scatter against that count. The scatter falls quickly and then flattens. What does the elbow on this plot mark?
• A retailer clusters its stores and computes a silhouette score for each store. The score is high for some stores and low for others. What does a high silhouette indicate about a store?
• A utility inspector has one labeled photo of a new insulator crack and many unlabeled tower photos. The team wants the model to flag that crack after seeing only that single labeled photo. Which learning setup matches this request?
• A grocery chain runs k-means on shoppers using annual spend in dollars and visits per month. Spend ranges in the thousands while visits stay in the dozens, and one spender group dwarfs the others. Which limitation of k-means fits this run?
• A factory engineer wants to group defect photos without choosing how many groups to use at the start, then cut that grouping at a level the crews can staff. Which method fits this plan?
• A hospital imaging team must handle a rare implant the model has never seen. They can describe that implant in text with no example images, or they can supply a handful of labeled images of it. Which pairing of those two setups is correct?
• A retail team holds a large matrix with one row per store and one column per product. The team wants to factor that matrix into orthogonal pieces plus a diagonal that records how strong each piece is, then keep the strongest pieces. Which method is that factorization?
• A hospital infection-control team clusters caregiver movement traces. True groups follow long curved corridors rather than tight round blobs, and a few isolated traces should stay unlabeled instead of joining a group. Which method matches both requirements?
• An agency stores dozens of correlated economic indicators and wants fewer uncorrelated directions that preserve as much of the spread as possible, then use those directions as inputs to a linear model. Which method produces the directions?
• A telecom desk receives a new cell-site reading that has no label yet. Historical readings close to it in feature space are already marked either as an alarm or as normal operation. Which rule should assign the new reading's label?
• A hospital radiology group uses t-SNE to plot study embeddings so similar scans sit near each other in two dimensions. A colleague wants to save those two coordinates and feed them to a diagnostic model as if they were stable inputs. What is the sound use of that plot?
• A utility plots feeder sensors with UMAP so operators can see which outage neighborhoods sit near each other. An engineer wants to store those coordinates as permanent model inputs beside a linear component already saved for reuse. How should the UMAP coordinates be treated?