D603 Machine Learning, catalog number DTAN 6215, is the three CU graduate course in the WGU Master of Science, Data Analytics covering the building, training, running and testing of supervised and unsupervised models that predict, classify or cluster. It is the graduate counterpart to the identically titled undergraduate course D499, and the difference between them is not the algorithms. It is the standard of proof expected before you are allowed to claim a model works.
The graduate standard of proof
Undergraduate machine learning work asks you to produce a model and report a metric. Graduate work asks you to show that the metric means what you say it means. That difference drives almost every return in DTAN 6215.
Concretely, the standard involves four things a rubric will look for even when it words them differently. A defensible split, so the number you report was computed on data the model never touched. A baseline, so the reader knows whether the model beats guessing. A tuning account, so the reader knows whether the hyperparameters were chosen or stumbled into. And an error analysis, so the reader knows where the model fails rather than only how often.
Error analysis is the one most often missing. A confusion matrix printed without commentary is not error analysis. Naming which class the model confuses, what those cases have in common, and whether the confusion is tolerable given the decision the model supports, is. That single section usually decides whether a submission reads as graduate work.
Turning scored aspects into a section plan
Scoring detail lives in your Course of Study rather than the public catalog, so open the rubric first and count the aspects. Each is judged separately against a three point scale, and a 2 in every one is what passes the task. Nothing averages, so an excellent model with an unexplained preprocessing step still comes back.
Use the aspect list as headings, worded close to the rubric's own nouns. Machine learning students resist this more than most, because a technical report has its own conventional shape. The conventional shape is not the scored shape, and an evaluator should never have to decide which of your paragraphs was meant to answer which aspect.
The word budget, worked. Suppose seven scored aspects and roughly 2,100 words of narrative outside the code appendix. Hold 150 for an opening that names the prediction target and the decision it supports, and 130 for a close. That leaves about 1,820, near 260 words per aspect. Then move 70 words out of each of the three descriptive aspects and give 200 of that to error analysis and the rest to method justification. Error analysis at 90 words is a paragraph that restates a table. At 350 it becomes the section that shows you understand your own model.
A structure that fits a graduate machine learning report
Directions win where they specify a shape. Where they do not, this ordering matches how machine learning aspects are usually written.
| Section | What belongs in it | How it gets read |
|---|---|---|
| Problem and target | What is being predicted, classified or clustered, and what decision changes as a result | Unclear targets produce unclear metrics; this is checked first |
| Data and preparation | Source, size, class balance, preparation steps each with a reason | Class balance is the fact most often omitted and most often needed |
| Method selection | The algorithm, the alternatives rejected, why this one fits this data and this decision | Rejected alternatives are the clearest evidence of graduate reasoning |
| Training and tuning | Split strategy, cross validation, hyperparameter search and what it changed | Scored on whether choices were made or defaulted |
| Evaluation | Metrics with splits named, compared against a stated baseline | A metric with no baseline cannot be judged good or bad |
| Error analysis | Where the model fails, what those cases share, whether that failure is acceptable | The section that separates graduate from undergraduate work |
| Deployment and ethics considerations | What would have to be true to use this, and who could be harmed if it is wrong | Often light, always noticed when missing entirely |
Keep the code appendix reproducible and keep the narrative independent of it. If the report only makes sense with the notebook open, the report is not finished.
Evidence craft when the evidence is a trained model
The hard part of citing machine learning work is that your strongest evidence is something you produced, which means the reader has to be able to check it.
- Attach a split label to every number. Training, validation and holdout figures reported without labels are treated as unverified.
- State the baseline explicitly, including the naive one. Predicting the majority class is a legitimate benchmark and it embarrasses more models than students expect.
- Report the search space for any tuned hyperparameter, not just the winning value.
- Cite the algorithm to a methodological source rather than to a library page. The library implements the idea; it is not the authority for it.
- Fix and report seeds, or state clearly that results vary and by roughly how much.
- Use APA for all external sources and keep direct quotation short, since definitions are the easiest thing to lift and the easiest to detect.
Name the population limit of your data before an evaluator does. A model trained on one region, one season or one customer segment cannot be claimed as general, and saying so plainly converts a weakness into evidence of judgment.
What separates Competent from a submission sent back
Independent aspect scoring means returns are narrow. Most reports that come back are strong reports with one hollow section.
- Every reported metric names its split and its baseline.
- Every preprocessing step happens inside the split rather than before it, and the report shows the order.
- Error analysis names specific failure cases rather than describing the matrix in words.
- The rejected alternatives are real alternatives, with the reason for rejection tied to this dataset.
- The conclusion answers the decision named in the first section, using the same terms.
Performance assessment work can be revised and resubmitted without a grade penalty, so a return costs time rather than standing. Terms run six months at a flat rate, which makes evaluator queue time the expensive part of any rework.
Six mistakes that cost time in D603
- Leakage through preprocessing. Fitting an encoder, imputer or scaler on all the data before splitting inflates every number that follows.
- Accuracy on imbalanced classes. Ninety four percent accuracy on a target that occurs six percent of the time is a model that predicts nothing.
- Tuning without reporting. A hyperparameter value with no search account looks like a default, and defaults are not choices.
- Treating clustering results as findings. Clusters need interpretation and a stability check before they are allowed to mean anything.
- Silent feature engineering. A derived variable that appears in the code but never in the narrative is invisible to scoring and looks like an accident.
- Ending at evaluation. A model report with no discussion of what deploying it would require is unfinished at graduate level.
Choosing a baseline you can defend
A surprising share of D603 rework comes from a single missing paragraph: the one establishing what a good result would even look like before the model was built. Without it, an evaluator has no way to judge whether 0.82 is impressive or embarrassing, and neither do you.
Build the baseline in layers. Start with the trivial one, which for classification is predicting the majority class every time and for regression is predicting the mean. That number is free to compute and it sets the floor no model is allowed to fall below. Report it even when it is unflattering to the exercise, because a target that occurs in three percent of records gives a do nothing model ninety seven percent accuracy, and that fact reframes every metric that follows.
Add a second layer: the simplest real model that could work. A logistic regression, a shallow decision tree, a rule someone in the business already uses. This is the baseline that matters most in graduate work, because the interesting question is not whether your model beats guessing but whether it beats the obvious approach by enough to justify its extra complexity, its extra maintenance and its reduced interpretability.
Then state the decision threshold in business terms. A model supporting an intervention that costs money per contact has an economically optimal operating point, and it is rarely the point that maximises a symmetric metric. Naming the relative cost of a false positive against a false negative, even approximately, turns a model report into an analysis of a decision.
Written up, this is three short paragraphs near the start of the evaluation section. It is also the single cheapest way to move a machine learning submission from adequate to clearly graduate level.
How support works on this course
Send the rubric from your Course of Study and the task directions. You get an aspect mapped draft: preparation written with reasons, splits and baselines stated, tuning documented, and an error analysis section that actually analyses errors, plus a walkthrough so the reasoning is yours to defend.
D603 is where the MSDA specialization work starts making demands on writing as much as on modelling. Clearing it cleanly the first time is worth real calendar in a six month flat rate term.
Questions students ask about D603
Is D603 the same course as DTAN 6215?
How is D603 different from D499?
Can someone build and submit the model on my behalf?
Where D603 sits in WGU's programs
The July 2026 catalog places this code in 1 current WGU program. Open a program page for the complete standard path and term positions. The live Degree Plan remains authoritative after transfer credit, substitutions, and mentor planning.
The assessments, one by one
The public catalog does not publish this course's PA/OA identity or task count. WGU Tutors publishes at most one PA manual per course and only from a WGU-controlled public rubric. Until that source exists, PA help begins from the student's real Course of Study and OA support remains preparation only.