Learning Paradigms and Problem Framing

A 1-to-5 star rating is an ordinal target: the values are ordered, but the gaps between them are not guaranteed equal. Regression uses the order and assumes equal spacing; five-class classification keeps the classes but discards the order.

Build one row per time step whose features use only past values — lags such as 1, 7 and 14 steps back, plus rolling means or spreads over trailing windows — and whose label is the value at the forecast horizon.

Because it sets the bar almost for free. Seasonal-naive predicts each future hour with the load observed at the same hour one week earlier, capturing the daily and weekly cycles without any training; a model that cannot beat it is adding nothing.

A discriminative classifier models p(y|x) - the label given the features - directly. A generative classifier models the joint p(x,y), normally as a class prior p(y) times a per-class feature distribution p(x|y), and inverts it to score labels.

Unlabelled data helps only when its shape reveals where the classes separate. Semi-supervised methods assume points in the same dense cluster share a label, so the boundary belongs in a low-density gap. When that assumption is false, unlabelled data hurts.

A parametric model summarises the training data with a fixed number of parameters, decided before fitting, so its size does not change as data grows. A non-parametric model's effective parameter count grows with the training set.

Supervised learning trains on examples that each carry a known target value and learns to predict it. Unsupervised learning has no target; it describes structure in the inputs alone - groups, directions of variation, or where the data is dense.

Batch refitting discards the current model and fits a new one over the whole dataset. Incremental (online) learning keeps the existing parameters and nudges them with each new example or chunk, without revisiting the older rows.

That is a multilabel target: the tags are not mutually exclusive, so an article can carry zero, one or several. Frame it as 20 independent yes/no decisions with a probability per tag, not one distribution over 20 tags summing to one.

Bandit feedback reveals the outcome only for the action actually taken; a supervised label states the correct answer regardless of what the model did. Showing one of six headlines teaches you nothing about the other five.

Self-training adds the model's own confident predictions to its training set as ground truth. Some are wrong, and training on them raises confidence in the same mistakes, so more of them clear the threshold each round.

Pose it as ranking when a fixed review capacity, not a decision rule, consumes the output. Only the relative order near the top then changes outcomes, so absolute score values and a class boundary buy nothing.

Choose one-class modelling when the anomalies you will face are not represented by the anomalies you have labelled, so no boundary learned from past examples transfers. Choose supervised classification when labelled anomalies are plentiful and representative of what comes next.

With far more columns than rows, use a strongly regularized linear family - ridge, lasso or elastic net, or a linear SVM - possibly after dimension reduction. Flexible families such as deep trees have far too little data per decision.

A warm start carries old data in only as a starting point. Once every later update uses recent rows alone, each step moves the parameters further from that start, so an old example's influence decays away.

Recursive forecasting iterates one one-step model, feeding its predictions back as inputs, so errors compound. Direct forecasting fits a separate model per horizon on real observed history: no compounding, but many models and thinner data.

Naive Bayes trades bias for variance: its generative assumptions are usually wrong, putting a floor under its error, but its per-class estimates stabilise after very few labels. Logistic regression has a lower floor and needs more data to reach it.

Uncertainty sampling scores every unlabelled row by how unsure the current model is — one minus the top predicted probability, the top-two gap, or entropy — and buys human labels for the least certain rows, then retrains and repeats.

Flexibility is paid for in rows. With 250 examples a model whose capacity grows with the data builds each local estimate from a handful of points, so it mostly tracks noise. A fixed functional form acts as a stabilising prior.

An unsupervised grouping has no accuracy score, because there is no ground truth. Judge it on stability under resampling, on whether domain experts recognise the groups, and on whether acting on it improves a measurable downstream outcome.

Fit out-of-core: stream the file in fixed-size chunks and update a learner that has a per-chunk update rule, so only one chunk sits in memory. First check whether a sample or fewer columns removes the problem.

Model the number when you need the ETA itself or the cutoff may move; classify the flag directly when the cutoff is fixed and only the decision matters. A regressor optimises accuracy everywhere, not around the 30-minute boundary.

Define features against the last landed observation, not the clock. With a three-hour delay the shortest usable lag is four hours, and training rows must be rebuilt at that cutoff so the model never relies on values serving lacks.

Repayment is observed only for approved applicants, so rejected ones carry no outcome to learn from. The training set is the slice the previous policy approved, and nothing in it says whether a rejection would have repaid.

An undefined target is a definition problem, not a modelling problem. Find a recorded outcome that can stand in for 'high performer', name the bias that proxy encodes, test whether two reviewers agree, price the labelling, and consider a descriptive deliverable.

Only with a correction for how those labels were produced. Reviewers judged only what the incumbent detector surfaced, so a classifier trained on them learns to imitate that detector and stays blind wherever it never looked.

A count is a non-negative integer, usually right-skewed with many zeros and a spread that grows with its own level. Plain squared-error regression predicts unbounded numbers and assumes constant spread, so it emits negative counts and lets busy hours dominate.

Graded labels make the target an ordering among relevant items rather than a set of hits. With perfect, good, marginal and bad grades, ranking a marginal result above a perfect one is an error binary labels cannot express.

An n-gram matrix has tens of thousands of mostly-zero columns holding many weak additive cues. A linear model sums all their weights at once; a tree tests one feature per split, so it needs thousands of trees.