Overview
Domain 2 moves from vocabulary to mechanics. You need to explain how a machine learning model is built: collect data, choose features, train the model so it fits patterns in that data, then use the trained model for inference on new inputs. The section then splits ML into its three learning styles. Supervised learning uses labelled examples and produces either a number (regression) or a category (classification).
Unsupervised learning has no labels and finds structure on its own, most often by clustering similar records. Reinforcement learning trains an agent through trial and error, using rewards from an environment to improve a policy. Oracle's course uses simple examples such as house prices, spam filters, customer segments and game-playing agents, and the exam questions follow that level.
How it's tested
Most items present a short business problem and ask which learning type, algorithm family or output fits. Read for three clues: whether historical labels exist, whether the target is a number or a category, and whether the system learns by acting and receiving feedback.
-
Labels or no labels
A dataset is described and you decide between supervised and unsupervised learning based on whether known outcomes are present.
-
Number or category
For a supervised task, you choose regression for a continuous value or classification for a discrete class.
-
Learning by reward
A robot, game or pricing agent improves from feedback, which signals reinforcement learning rather than supervised training.
A bank has five years of loan records, each marked as repaid or defaulted, and wants to score new applicants. Because every historical record carries a known outcome, this is supervised learning. The outcome is one of two categories, so the task is binary classification, not regression. A logistic regression model, despite its name, fits because it outputs a probability that is turned into a class.
Objectives 4
Official objectives from the Oracle outline, in outline order.
2.1 Explain machine learning basicsFeatures, labels, training and inference.
- What it means
- Machine learning builds a model that maps input features to an output by learning from examples. Training adjusts the model's parameters to reduce error on known data, and inference applies the trained model to inputs it has never seen.
- What to understand
- Know the vocabulary: features are the input attributes, a label is the known answer, and the model is the learned function. Data is usually split into training and test sets so performance is measured on unseen examples. Overfitting means the model memorised the training data and performs poorly on new data; underfitting means it is too simple to capture the pattern.
- How it shows up
- Questions ask which phase uses labelled data, or what overfitting looks like in training versus test results.
- Common confusion
- Confusing training, which changes the model, with inference, which only uses it.
2.2 Discuss supervised learning fundamentals (regression and classification)Learning from labelled data to predict values or classes.
- What it means
- Supervised learning trains on records where the correct output is known. Regression predicts a continuous number such as a price or temperature, while classification predicts a discrete class such as spam or not spam.
- What to understand
- Linear regression fits a line through data points by minimising the gap between predicted and actual values, often measured as mean squared error. Logistic regression applies a sigmoid function to output a probability between 0 and 1, then uses a threshold to assign a class. Classification can be binary or multi-class. Evaluation for classifiers uses accuracy, precision and recall, while regression uses error measures.
- How it shows up
- Expect a target variable to classify as continuous or categorical, or a question on what logistic regression outputs.
- Common confusion
- Assuming logistic regression is a regression task because of its name; it is used for classification.
2.3 Discuss unsupervised learning fundamentalsFinding structure in unlabelled data.
- What it means
- Unsupervised learning works on data without labels and looks for natural groupings, patterns or unusual points. Clustering is the main example: records that are similar to each other end up in the same group.
- What to understand
- K-means clustering picks a number of clusters, assigns each point to the nearest centre, then moves the centres and repeats until the groups stabilise. Similarity is measured by distance between feature values. Common uses are customer segmentation, grouping documents, and anomaly detection, where points far from any cluster are flagged. Dimensionality reduction is another unsupervised technique that compresses many features into fewer.
- How it shows up
- A scenario with no historical outcomes, such as grouping customers by behaviour, points to clustering.
- Common confusion
- Expecting clusters to arrive with meaningful names; people interpret and label groups afterwards.
2.4 Discuss reinforcement learning fundamentalsAgents learning from rewards in an environment.
- What it means
- Reinforcement learning trains an agent that takes actions in an environment. After each action the environment returns a new state and a reward, and the agent learns a policy that maximises total reward over time.
- What to understand
- There is no labelled dataset; feedback arrives as rewards or penalties, often delayed. The agent must balance exploration, trying new actions, against exploitation, repeating actions known to pay off. Typical examples are game playing, robotics, autonomous driving and dynamic pricing. Reinforcement learning from human feedback is also used to align language models, which links this topic to Domain 4.
- How it shows up
- You match agent, environment, state, action, reward and policy to parts of a described scenario.
- Common confusion
- Calling the reward a label; it scores an action rather than giving the correct answer.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Feature
An input attribute used by the model.
- Remember
- Features go in, predictions come out.
- Tested as
- Vocabulary items.
- Trap
- Calling the target column a feature.
-
Label
The known correct output for a training record.
- Remember
- Labels define supervised learning.
- Tested as
- Deciding supervised versus unsupervised.
- Trap
- Assuming every dataset already has labels.
-
Training vs inference
Training fits parameters; inference applies the fitted model.
- Remember
- Only training changes the model.
- Tested as
- Lifecycle questions.
- Trap
- Thinking inference keeps improving the model automatically.
-
Overfitting
Strong training results but weak results on new data.
- Remember
- Fix with more data, simpler models or regularisation.
- Tested as
- Interpreting train versus test accuracy.
- Trap
- Reading high training accuracy as proof of a good model.
-
Regression
Supervised prediction of a continuous value.
- Remember
- Output is a number on a scale.
- Tested as
- Target-type questions.
- Trap
- Using it for yes or no outcomes.
-
Classification
Supervised prediction of a discrete class.
- Remember
- Binary or multi-class.
- Tested as
- Most common scenario type.
- Trap
- Treating ordered ratings as regression without thinking about the classes.
-
Logistic regression
Classifier that outputs a probability via a sigmoid curve.
- Remember
- A threshold converts probability to a class.
- Tested as
- Name-versus-purpose trick.
- Trap
- Classing it as a regression algorithm.
-
Clustering
Grouping unlabelled records by similarity.
- Remember
- K-means is the standard example.
- Tested as
- Customer segmentation scenarios.
- Trap
- Choosing classification when no labels exist.
-
Reinforcement learning
An agent learns a policy from rewards.
- Remember
- Agent, environment, state, action, reward, policy.
- Tested as
- Robotics and game scenarios.
- Trap
- Expecting a labelled dataset.
-
Exploration vs exploitation
Trying new actions versus repeating known good ones.
- Remember
- Both are needed to find the best policy.
- Tested as
- Conceptual RL items.
- Trap
- Thinking pure exploitation is always optimal.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Three learning styles
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Supervised | Learns from labelled examples. | Needs known outcomes. | Picking it when the data has no answers. |
| Unsupervised | Finds structure without labels. | Output is groups or patterns, not predictions of a known target. | Expecting it to predict a specific named outcome. |
| Reinforcement | Learns from rewards while acting. | Feedback comes from an environment over time. | Confusing rewards with labels. |
Regression vs classification
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Linear regression | Fits a line to predict a number. | Continuous output. | Applying it to category targets. |
| Logistic regression | Predicts class probability. | Discrete output after a threshold. | Being misled by the word regression. |
| Multi-class classification | Chooses among three or more classes. | More than two possible labels. | Assuming classifiers only handle yes or no. |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Logistic is classification
- What candidates pick
- Regression
- Why it looks right
- The algorithm name contains the word regression.
- Why it's wrong
- Logistic regression outputs a probability that is thresholded into a class, so the task it solves is classification.
Clue to watch for Yes or no, churn or stay, fraud or legitimate.
-
No labels present
- What candidates pick
- Classification
- Why it looks right
- Grouping sounds like assigning classes.
- Why it's wrong
- Without known outcomes there is nothing to train a classifier on; clustering discovers the groups instead.
Clue to watch for Words like segment, group or discover.
-
Reward is not a label
- What candidates pick
- Supervised learning for a game agent
- Why it looks right
- The agent seems to learn right and wrong moves.
- Why it's wrong
- The agent receives scores for its actions from an environment rather than correct answers for fixed examples.
Clue to watch for An agent acting over many steps.
-
Training accuracy only
- What candidates pick
- The model with the highest training score
- Why it looks right
- Bigger numbers look better.
- Why it's wrong
- A very high training score with a low test score indicates overfitting and poor generalisation.
Clue to watch for Two scores reported side by side.
-
Numbers coded as classes
- What candidates pick
- Regression for product category IDs
- Why it looks right
- The target column holds numbers.
- Why it's wrong
- Category codes are labels, not quantities, so averaging them is meaningless and classification applies.
Clue to watch for Codes, IDs or ratings with fixed options.
Answer framework
Work through these checks whenever a Domain 2 item asks which learning approach or output fits a scenario.
- Ask whether the system learns by acting in an environment and receiving rewards; if so, the answer is reinforcement learning.
- Otherwise, check whether the training records include a known outcome for each example, which makes the problem supervised.
- If there are no outcomes and the goal is to find groups or unusual records, choose unsupervised learning such as clustering.
- For supervised problems, decide whether the target is a measurable quantity for regression or a fixed set of classes for classification.
- Check any reported results for signs of overfitting, such as a large gap between training performance and test performance, and remember that more data or a simpler model usually narrows that gap.
Practice question
Want more like this? The 1Z0-1122-26 practice test has questions for every domain.
Checklist
0 of 10
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Plan five to seven hours. Watch the machine learning lessons in Oracle's free AI Foundations course first, paying attention to the worked examples on house prices, email spam and customer segmentation, because exam scenarios use the same style.
Build a decision table with three columns: learning style, what the data looks like, and what comes out. Fill it with at least three examples per row. Then add a second table for supervised learning only, splitting regression from classification and listing the evaluation measures that go with each.
Practise by writing twenty short scenarios, or using a question bank, and labelling each one within thirty seconds. When you miss one, write down which clue you ignored: labels, target type or reward signal. Spend a little extra time on the logistic regression naming trap and on overfitting, since both show up regularly in foundation-level ML exams.
Finally, connect this section forward: note which learning styles OCI Data Science and Oracle Machine Learning support, so Domain 5 questions feel familiar.
FAQ
How many questions come from Intro to ML Foundations?
Oracle does not publish weights for 1Z0-1122-26. Secondary estimates suggest about six of the 40 questions come from this machine learning section.
Do I need to calculate anything?
No. Questions are conceptual. You should understand what mean squared error or a sigmoid does, but you will not compute them.
Which algorithms should I know by name?
Linear regression, logistic regression and k-means clustering cover nearly every example used in Oracle's course.
Is reinforcement learning tested deeply?
Only at the level of its components and typical uses. Know the agent and reward loop and the exploration versus exploitation trade-off.
What is the most common mistake in this section?
Choosing the wrong learning style because the presence or absence of labels was missed in the scenario.
How does this section connect to OCI services?
Domain 5 asks which OCI tools build and deploy custom ML models, so understanding the learning styles here makes those choices easier.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 2 knowledge
You have covered ML basics, regression, classification, clustering and reinforcement learning. Try timed scenario questions on choosing the right approach.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the 1Z0-1122-26 AI tutor Get any concept explained your way