Supervised, Unsupervised and Reinforcement Learning in Snowflake ML

1.Does each training row already have the answer?

Before you choose an algorithm or a Snowflake feature, ask one question about your training data: does each row already include the answer you want the model to produce? The answer decides which of the two paradigms in the Snowflake documentation you are working in.

In supervised learning, each training row pairs some inputs (features) with a known outcome. Snowflake's Classification function makes this explicit. Its training data must include "a target column representing the labeled class of each data point" and at least one feature column. The model learns how the features relate to that label. It then applies what it learned to rows whose label is unknown. The documentation gives churn prediction, credit card fraud detection and spam detection as typical uses. The Auto ML skill in Snowflake describes the same kind of task as "predict a category".

In unsupervised learning, the data has no answer column. The model has to find structure without help, such as values outside the expected range or rows that fall naturally into groups. The Auto ML skill describes clustering as a way to "segment data into groups", for example customer segmentation. Snowflake's anomaly detection function can also be trained with no labels at all.

Reinforcement learning is the third paradigm in the exam guide. The documentation behind this lesson does not cover it. The last section explains what that means for your preparation.

Sources12

2.Supervised learning in practice: training on labelled rows

Classification "uses machine learning algorithms to sort data into different classes using patterns detected in training data." It supports binary classification (two classes) and multi-class classification (more than two). Classification is one of two common supervised task types. The Auto ML skill lists the other as regression: "predict a number", such as revenue, price or demand. The difference lies in the target: classification predicts a category, regression predicts a number. In SQL, you train a classifier by creating a model object and passing it a reference to the training data. You name the label column with the TARGET_COLNAME argument. The training data also needs at least one feature column. Fill in the argument that tells the model which column is the label:

CREATE OR REPLACE SNOWFLAKE.ML.CLASSIFICATION model_binary(
    INPUT_DATA => SYSTEM$REFERENCE('view', 'binary_classification_view'),
     ?  => 'label'
);

Creating the object also fits the model: "The model is fitted to the provided training data." After that you call the model's PREDICT method on new rows. This step, applying a trained model to unlabelled data, is called inference. Inference only works if the new rows look like the training rows. "Inference data must have the same feature names and types as training data." Within that rule, the function tolerates some variation. A categorical feature can contain a value the model never saw during training. Columns that were not in the training data are ignored.

The label column has rules of its own. It needs at least two classes, because a model cannot learn to tell classes apart if every row has the same one. It also cannot have a distinct value on every row, and it may hold at most 255 distinct classes. Numeric features are treated as continuous. String and Boolean features are treated as categorical. To make the model treat a numeric code as a category, cast it to a string.

Sources12

3.Labels make evaluation possible, and parameters are learned

Because a supervised model learns from known answers, you can measure how often it gets them right. The useful test is whether it gets them right on rows it has not seen. Snowflake's Classification function does this automatically. During evaluation, "an additional model is trained on the original data but with some data points withheld." It then runs inference on the withheld rows and compares its predicted classes with the actual ones. You read the results through the model's evaluation methods:

CALL <model_name>!SHOW_EVALUATION_METRICS();
CALL <model_name>!SHOW_GLOBAL_EVALUATION_METRICS();
CALL <model_name>!SHOW_THRESHOLD_METRICS();
CALL <model_name>!SHOW_CONFUSION_MATRIX();

A model has two kinds of settings. Parameters are the values the model learns from the data during training. Hyperparameters are values a person chooses before training starts. The ML Functions hide both kinds from you. For Classification, "You cannot choose or modify the classification algorithm", and "Model parameters cannot be manually specified or adjusted." If you write your own training code on Snowflake's Container Runtime, you control them. The PyTorch loop below is from the Snowflake training guide. The learning rate (lr=1e-3) and EPOCHS = 5 are set before the loop runs. Inside the loop, criterion(logits, yb) compares the model's output with the labels yb, and opt.step() updates the learned weights. This is supervised learning in its most basic form. Snowflake ML can also "Accelerate hyperparameter tuning" with distributed HPO (hyperparameter optimisation).

for epoch in range(1, EPOCHS + 1):
    model.train()
    for xb, yb in train_loader:
        xb, yb = xb.to(DEVICE), yb.to(DEVICE)
        logits = model(xb)
        loss = criterion(logits, yb)
        opt.zero_grad()
        loss.backward()
        opt.step()
    acc = evaluate(val_loader)
    print(f"epoch {epoch} val_acc={acc:.3f}")

Sources13

4.Unsupervised learning: finding outliers and groups without labels

"Anomaly detection is the process of identifying outliers in data." Outliers are data points that fall outside the expected range. The function works on time series and needs two columns: a timestamp, and a target column holding "some quantity of interest at each timestamp". The word *target* is easy to misread here. In Classification, the target is the label to predict. In anomaly detection, it is the metric being watched, such as daily sales. What makes anomaly detection supervised or unsupervised is a separate argument, LABEL_COLNAME. In the documentation's basic example, the label argument is an empty string: "Because this example uses unsupervised training, you do not need to use the label column." If you do know which past rows were anomalies, a Boolean label column tells the model so, and training becomes supervised.

Clustering is the other common unsupervised workload. The Snowflake Auto ML skill lists it as a task type for segmenting data into groups, for example customer segmentation or grouping anomalies. Unlike classification, clustering is given no classes in advance: it discovers the groups itself, so there is no label column to name.

Because no label column exists, the result cannot be scored against correct answers the way a classifier's withheld rows are. Snowflake's Feature Store guide fits a KMeans model and lists techniques for judging clusters: Within-Cluster Sum of Squares, Silhouette Score and Gap Statistics. The guide itself only plots the clusters and reviews them visually. The same guide scales the input variables with min-max scaling before fitting, so that columns on large numeric ranges do not dominate the grouping.

Clustering is not the only unsupervised technique. Snowflake's documentation on GPU-accelerated libraries also describes dimensionality reduction, which shrinks high-dimensional data (such as text embeddings) to fewer dimensions. It uses that step together with clustering to extract topics from product reviews, and it lists UMAP and HDBSCAN among the libraries. None of this needs a label column either.

Sources4256

5.Reinforcement learning: not covered by this documentation

The documentation does show one contrast. Every model in this lesson learns from a fixed dataset you hand it when you create the model. Once trained, it does not keep learning as it is used. "Models are immutable and cannot be updated in place." To include new outcomes, you drop the model and train a new one, often with CREATE OR REPLACE. So if a question describes a model that keeps changing its behaviour based on feedback from new results, that is not what the ML Functions shown here do.

Sources1

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.Anomaly detection is always unsupervised, because outliers are never labelled in advance.Why is that wrong?

   SNOWFLAKE.ML.ANOMALY_DETECTION can train without labels (LABEL_COLNAME => '') or with a Boolean column that marks known anomalies. With that column, training is supervised.

   Covered in Unsupervised learning: finding outliers and groups without labels

2. 2.Any function with a TARGET_COLNAME argument is doing supervised learning.Why is that wrong?

   In anomaly detection, the target column is just the metric being tracked over time. Whether training is supervised depends on the separate label column.

   Covered in Unsupervised learning: finding outliers and groups without labels

3. 3.Classification evaluation metrics are measured on the rows the model was trained on.Why is that wrong?

   Evaluation trains an additional model with some rows withheld, then compares its predictions on those withheld rows with their actual classes.

   Covered in Labels make evaluation possible, and parameters are learned

4. 4.Any column can be the label, even one with a single value or a unique value on every row.Why is that wrong?

   The label column must have more than one distinct value and fewer distinct values than there are rows. It may also hold at most 255 classes.

   Covered in Supervised learning in practice: training on labelled rows

5. 5.Predicting a number, such as price or demand, is classification because it is supervised.Why is that wrong?

   Both classification and regression are supervised, but classification predicts a category and regression predicts a number.

   Covered in Supervised learning in practice: training on labelled rows

6. 6.Clustering is just classification under another name, so it needs a label column.Why is that wrong?

   Classification sorts rows into classes using a labeled target column. Clustering segments data into groups without labels, so its groups are judged by measures such as silhouette score instead of correct answers.

   Covered in Unsupervised learning: finding outliers and groups without labels