Day 1 - Algorithm Selection: Mapping by Problem Type

📌 Key Takeaways

• Week 6 is the starting point of Domain 3 (Modeling). The exam is built so that you narrow down in this order: ML problem type → algorithm family → SageMaker built-in.
• There are exactly two primary criteria: output form (category / number / group / recommendation / anomaly / topic) and presence of labels (supervised / unsupervised).
• The secondary filter is data form. If you see the words "text, image, time series, sparse," plain XGBoost drops out of the candidates.
• On imbalanced data, accuracy is a trap. Switch to precision, recall, F1, and PR-AUC.
• Even with the same type, data size, sparsity, interpretability, and training cost narrow the final candidate down to one.

Translating a Business Problem into an ML Type

The first decision in modeling is "which algorithm to use," and before it there must always come the translation of the ML problem type. Exam passages almost always give the problem in business language, so converting it into a standard ML type is step 1.

"Identify customers who will churn in advance"      → binary classification (churn yes/no)
"Sort email into spam / normal / promotional"       → multiclass classification
"Predict tomorrow's sales amount"                   → regression
"Automatically discover groups of similar users"    → clustering (unsupervised)
"Recommend products never seen before"              → recommendation / collaborative filtering
"Detect transactions that differ from the usual"    → anomaly detection
"Discover topics in a pile of documents"            → topic modeling (unsupervised)

There are three discriminating signals.

• Are there labels (answers)? If yes, supervised learning; if no, unsupervised.
• Is the output a category or a number? Category = classification, continuous number = regression.
• Is it "discovering groups/patterns" or "predicting"? Discovering = clustering/topics, predicting = classification/regression.

The boundary between supervised, unsupervised, and reinforcement is drawn by "the existence of answer labels." Classification and regression are supervised learning that learns input-output pairs, while clustering, dimensionality reduction, topic modeling, and (mostly) anomaly detection are unsupervised learning that finds structure without labels. Recommendation deals with a user-item interaction matrix and is closer to unsupervised from the collaborative filtering viewpoint, but if you use explicit ratings as the target it can also be viewed as supervised (regression). Reinforcement learning is a separate category that learns a policy from a reward signal (SageMaker RL).

Algorithm Selection Decision Tree

                        [Business problem passage]
                                │
                  ┌─────── Are there labels? ───────┐
                Yes (supervised)                  No (unsupervised)
                  │                                │
      ┌───── Output form? ─────┐          ┌──── What do you want to do? ────┐
   Category    Number    Rank/preference  Find groups  Anomaly  Reduce dim  Find topics
     │           │           │              │          │        │        │
 Classification Regression Recommendation K-Means     RCF /     PCA    LDA / NTM
     │           │           │                      IP Insights
  ┌──┴──┐    ┌───┴───┐   Factorization
Tabular Text Tabular Time series  Machines
  │    │      │       │
XGBoost │  XGBoost  DeepAR
Linear  │  Linear
KNN  BlazingText
  │
Image → Image Classification

Classification

When the output is a discrete category. It splits into binary, multiclass, and multilabel.

• Accuracy alone is often insufficient. For imbalanced data (1% fraud), look at precision, recall, F1, and AUC.
• The binary/multiclass distinction is specified by predictor_type (Linear Learner) or objective (XGBoost).

Regression

When the output is a continuous number. Predicting sales amount, price, demand quantity, and so on.

• Candidates: XGBoost (objective=reg:squarederror), Linear Learner (predictor_type=regressor).
• When the time axis is the core, as in time-series demand forecasting, DeepAR (Day 3) or Amazon Forecast is more suitable than forcing it into regression.
• Metrics: RMSE, MAE, R². If you are sensitive to outliers, look at MAE as well.

Clustering

Groups data into natural clusters without labels. Customer segmentation, document grouping.

• Candidate: K-Means (SageMaker built-in). You specify the number of clusters k in advance.
• Evaluation: since there is no external answer, choose k with internal metrics such as the silhouette coefficient and the elbow (WCSS).

Recommendation

Predicts preferences from user-item interactions.

• Candidates: Factorization Machines (strong on sparse high-dimensional interactions), neural-network-based recommendation.
• Cold start and sparsity (most cells are empty) are the key difficulties.

Anomaly Detection

Finds rare events that deviate from normal. It is commonly approached as an unsupervised problem with almost no labels.

• Candidates: Random Cut Forest (RCF), IP Insights (entity-IP anomalies), and for streaming, Kinesis Data Analytics RCF.
• The imbalance is so extreme (0.1% anomalies) that it is often hard to solve as classification.

There is a trap: "Can't we solve anomaly detection as classification?" If there are enough anomaly labels and the patterns are stable, supervised classification is also possible, but in reality anomalies are extremely few and keep changing into new forms. In that case, unsupervised anomaly detection (RCF and the like), which learns only normal behavior and measures deviation, is more robust. On the exam, if you see the clue "almost no labels / new forms of anomaly," the RCF family is more likely to be the right answer than classification.

At-a-Glance Mapping Table

Metric Interpretation Table: Once the Type Is Set, the Metric Is Set

⚠️ Pitfall: If the passage says "fraud is 1% of the total," any answer choice that includes accuracy is wrong. Even guessing everything as normal yields 99%. In this case the answer shifts to recall (we cannot afford to miss) or PR-AUC.

Preprocessing Requirements by Algorithm

Even with the same problem type, preprocessing requirements differ. The exam repeatedly asks "is scaling needed?"

Trees split by thresholds within a single feature, so the result is the same as long as only the order of values is preserved (invariance to monotonic transformations). Distance-based and linear models, on the other hand, feed feature values directly into distance and slope calculations, so scale differences distort the result directly. This single line separates every scaling question.

Additional Considerations When Selecting an Algorithm

Selection does not end with the problem type.

Use the Built-in Directly, or Go with a Managed AI Service?

If the passage says "no ML specialists," "launch quickly," or "we do not want to train and manage models ourselves," the answer is not a built-in algorithm but a higher-level AI service.

⚠️ Pitfall: There is a trap in the opposite direction too. If the passage says "we need to blend in our domain-specific features and tune it precisely," choosing a managed AI service is wrong. The required level of control separates the two paths.

Question Clue → Answer Mapping Table

The fastest exam tip is to memorize a dictionary of "passage phrase → candidate algorithm."

If you pick an algorithm name before confirming the output form, you will be wrong. "Customer segmentation" looks like clustering, but if segment labels already exist in the passage, multiclass classification is the answer. The presence of labels flips supervised and unsupervised.

The 5-Step Order for Solving Scenarios

Specialty passages stack several constraints. Reading in the order below narrows most of them to one.

1. Confirm the output form: Is it category, number, group, recommendation, anomaly, or topic?
2. Check for labels: If there are none, every supervised-learning option is eliminated.
3. Check the data form: Is it tabular, text, image, time series, or sparse? Here the built-ins narrow to almost one.
4. Apply constraints: data size, interpretability, GPU and cost, real-time requirement.
5. Check preprocessing and metric consistency: See whether scaling requirements and imbalance metrics contradict the options.

In passages where constraints conflict, the answer is not the flashiest technology but the choice that satisfies all the constraints at once. If it violates even one constraint (for example, supervised classification when there are no labels), it is wrong regardless of its other merits.

Starting tomorrow, we dig deeply into each built-in algorithm in this mapping table (XGBoost, Linear Learner, K-Means, KNN), one at a time.

📖 Glossary

• Supervised learning: Learning a mapping from pairs of inputs and answer labels. Classification and regression belong here.
• Unsupervised learning: Learning that finds groups, anomalies, and topics from the structure of the data alone, without labels.
• Multilabel: A problem in which multiple answers are attached to one sample simultaneously. Several tags on one photo.
• Cold start: A situation, such as a new user, product, or store, where there is almost no history and prediction is difficult.
• Sparsity: A state in which most cells of a matrix are empty. The user-item rating matrix is the typical case.
• Elbow (WCSS): A method of choosing k by increasing the number of clusters and finding the point where the decrease in within-cluster sum of squares bends.
• Silhouette coefficient: A clustering quality metric expressing, from -1 to 1, how well a point fits in its own cluster.
• PR-AUC: The area under the precision-recall curve. More honest than ROC-AUC on data where positives are extremely rare.
• Monotonic transformation invariance: The property that the result is the same as long as only the order of the magnitudes is preserved. Tree models have this, so scaling is unnecessary.
• SageMaker built-in algorithm: A collection of algorithms AWS provides as containers that can be trained with hyperparameters alone, without code.