This AI Predicts Student Grades — Now It Shows Its Work

Imagine a teaching assistant that watches every click a student makes during online math practice and predicts whether they'll get the next question right. Schools increasingly use these systems — called "knowledge tracing" — to spot kids who need help. The problem: the AI never says why. A teacher told "this student will probably fail the next problem" has no idea whether to reteach fractions, slow down, or check if the child is just tired.

This paper builds a three-part test for those explanations. First, is the AI actually accurate? Second, does it give the same reasoning when you slightly change the data? Third, do the factors it blames really matter — if you remove them, does accuracy actually drop? The team used a simple, well-understood model (XGBoost) paired with a popular explanation tool (TreeSHAP) and compared it against four fancy deep-learning rivals on real student practice data. Interesting result: when everyone got the same information, they performed almost identically. The simpler model's advantage came from better data, not better math.

Along the way they caught a real problem. A famous public dataset, ASSISTments 2009, duplicates multi-skill questions into one row per skill — and because those copies all share the same right-or-wrong answer, the answer leaks into the model's "past performance" clues. Rebuilding the dataset lowered accuracy scores and reshuffled which factors looked most important. In other words, some published results may have been too optimistic.

The team is careful about limits: their student-by-student examples are illustrations, not proven teaching advice. But the bigger message is reassuring. AI in classrooms can be forced to show its reasoning, and that reasoning can be checked — not just trusted.