Where Decision Trees Actually Appear
Decision trees are among the most widely deployed AI structures in the world, largely because they combine two properties that most other models lack: they are effective enough to be useful, and transparent enough to be explained. That combination makes them especially attractive in regulated industries where decisions must be justifiable.
Credit Scoring and Loan Approval
The loan example in Reading 2 is not hypothetical. FICO scores — the credit scores used by most American lenders — were originally derived from decision tree-based analysis of repayment patterns. Many banks still use decision trees or tree-based models as the first filter in loan evaluation, because regulators require that lending decisions be explainable. A bank that denies a loan must be able to say why, in terms a customer can understand. "The model said so" is not legally sufficient. "Your debt-to-income ratio exceeds our threshold" — the kind of explanation a decision tree naturally provides — is.
Medical Diagnosis Support
Decision trees are widely used in clinical decision support systems: tools that help physicians make diagnoses or treatment decisions by prompting them through a structured set of criteria. The classic flowchart on a hospital wall — "If the patient presents with X and Y, consider Z" — is a hand-crafted decision tree. Modern clinical decision support software uses learned trees built from patient outcome data. The transparency requirement here is even more acute than in lending: a physician needs to understand the reasoning well enough to override it when clinical judgment demands.
Student Early Warning Systems
This brings us back to the educational context. Many school districts and universities use early warning systems — tools that flag students at risk of failing, dropping out, or needing additional support. These systems frequently use decision trees or closely related models, built from historical student data: attendance, grades, course completion, demographic factors.
The appeal is clear: a well-built early warning system can identify at-risk students earlier than any individual teacher could, allowing intervention before patterns become crises. The risks are equally clear, and they connect directly to the Week 5 SEC discussion of graduation risk scores: if the historical data reflects past inequities, the tree learns those inequities as patterns. A tree trained on data from a district that systematically underserved certain student populations will reproduce that underservice, automatically and at scale.
Spam Filtering and Content Moderation
Our running spam filter example fits naturally here. Early spam filters were essentially hand-crafted decision trees — explicit rules about keywords, sender patterns, and message structure. Modern spam filters use more sophisticated approaches, but the decision tree structure underlies many of the interpretable components: "if the sender is not in the contact list AND the subject contains certain patterns AND the body contains a link to an unfamiliar domain, route to spam." That is a decision tree path.
Strengths and Limits
Decision trees are not the right tool for every problem. Understanding where they work well and where they break down is part of understanding AI systems generally.
What Decision Trees Do Well
- Transparency. Every classification comes with a traceable path. This is the decision tree's defining strength and the reason it remains widely used in high-stakes domains decades after more powerful models were developed.
- Handling mixed attribute types. Decision trees can split on numerical attributes (income above $50,000?), categorical attributes (employment status is stable?), and binary attributes (flagged by teacher?) without requiring special treatment of each type.
- No assumptions about data distribution. Many statistical models require that data follow a particular pattern (often a normal distribution). Decision trees make no such assumption — they learn directly from the examples they are given.
- Teachability. A decision tree can be drawn on a whiteboard, traced with a pencil, and understood by students at almost any level. No other AI model has this property.
What Decision Trees Do Poorly
- Complex relationships between attributes. Decision trees make one decision at a time, considering one attribute per node. They struggle with problems where the right answer depends on a subtle combination of many attributes simultaneously — the kind of problem that neural networks handle well.
- Continuous outputs. Decision trees naturally produce categories (approve/deny, refer/monitor/no action). Predicting a continuous value — "what will this student's final exam score be?" — requires adaptation and is not a natural fit.
- Instability. Small changes in training data can produce very different trees. If you retrain a decision tree on slightly different data, the root node might change entirely. This makes individual trees less reliable than ensemble methods that combine many trees.
- Overfitting. As we discussed in Reading 2, deep trees tend to memorize training data rather than learn from it. Pruning helps, but the fundamental tension between accuracy on training data and generalization to new data is a persistent challenge.
Decision Trees as a Teaching Tool
Beyond their role in real AI systems, decision trees have a property that makes them uniquely valuable for K-12 instruction: they are the most hands-on, pencil-and-paper-friendly AI structure that exists.
At the elementary level, decision trees can be introduced as sorting games. Give students a collection of animals and ask them to build a tree that correctly sorts them: "Does it have four legs? Can it fly? Does it live in water?" The tree emerges from the students' own questions, and they can test it against new animals to see if it works.
At the middle school level, students can build simple decision trees from small datasets — a spreadsheet of thirty students with a few attributes and a known outcome — by hand. Which attribute creates the purest split? Students can argue about this without any formulas, using only the intuition from Reading 2.
At the high school level, the full treatment — training data, learning algorithms, overfitting, pruning, random forests — is appropriate and connects directly to AP Computer Science Principles' data and algorithm content.
In all cases, the key pedagogical value is the same: decision trees make the reasoning explicit. Students can see every question the system asks, every branch it follows, and every conclusion it reaches. That transparency is rare in AI, and it makes decision trees the ideal entry point for teaching students to think critically about automated decisions.
The Bridge to Supervised Learning
Decision trees are the clearest example of a pattern that will run through all of Week 6: supervised learning.
Supervised learning is the category of machine learning where an algorithm is trained on labeled examples — data where the correct answer is already known — and learns to predict the correct answer for new, unlabeled examples. Every loan application in the training data was labeled "Repaid" or "Defaulted." The algorithm used those labels as a teacher, adjusting the tree structure until it could predict the label from the attributes reliably.
The word "supervised" refers to the role of those labels: they supervise the learning process by providing feedback on whether the system is right or wrong. Without labels, there is no signal to learn from. With good labels on enough examples, a surprising amount of structure can be extracted automatically.
In Topic 6c, we mapped the full landscape of learning types — supervised, unsupervised, and reinforcement — and showed how decision trees fit into the bigger picture. In Topic 6e, we will look at artificial neural networks: a supervised learning approach that is far more powerful than decision trees for many problems, but far less transparent. The tradeoff between power and explainability is one of the central tensions in modern AI, and it is a question every educator, citizen, and future computer scientist should be able to reason about clearly.