D499 Machine Learning carries the banner number DTSC 3221 and is worth 3 competency units. It walks the end to end process of investigating data through a machine learning lens, across both supervised algorithms that learn from labeled examples and unsupervised algorithms that find structure without them. D499 and DTSC 3221 are one requirement. Note that WGU also offers a graduate course with the identical title, D499 being the undergraduate one, so check the code on your own degree plan before you start hunting for materials.
The split that organizes everything: labeled or not
Machine learning becomes navigable once you sort every method by what it is given. Supervised methods receive examples with known answers and learn to reproduce them on new cases. Unsupervised methods receive no answers and look for structure that was already there. Every algorithm in an introductory course sits on one side of that line, and knowing which side a problem is on determines the method family, the evaluation approach and what counts as success.
Inside supervised learning, the second sort is by what is being predicted. A number is a regression problem, a category is a classification problem, and the metrics differ completely. Reporting the wrong metric for the problem type is a common and expensive error, because it suggests the writer did not understand what the model is doing.
Evaluation is the part that separates people who understand machine learning from people who have run it. A model scored on the same data it learned from will look excellent and mean nothing. Holding out data, or cross validating, exists because the only question worth asking is how the model performs on cases it has never seen. Any result reported without saying how it was validated is unusable, and an evaluator will treat it that way.
Overfitting is the concept that ties the course together. A model that memorizes its training data fits it beautifully and generalizes badly, and every technique for controlling complexity is a response to that problem. Being able to describe overfitting in terms of your own results, rather than as a definition, is one of the clearest markers of understanding in student work.
Class imbalance deserves early attention because it silently ruins otherwise careful projects. When one outcome is rare, a model that always predicts the common outcome can post an impressive accuracy while being useless. Recognizing that situation and choosing metrics that survive it is a mark of competence. WGU judges this Competent or Not Competent instead of assigning a letter grade, keeps no ordinary grade point average, and sets these 3 competency units within a flat rate six month term.
Turning scored aspects into a modeling report plan
If your version of D499 is assessed by a performance assessment, the aspects your evaluator scores define the report and its weighting. WGU requires a score of 2 in each aspect for a task to pass and scores each aspect independently, so strong modeling will not carry an aspect on justification or limitations that you left thin.
Budget before drafting. Assume a rubric carrying five scored aspects and a target near 2,250 words. Reserve 160 words for the problem and the dataset, and 130 for the close, leaving 1,960. Three aspects that carry technical argument, typically algorithm selection, the modeling process itself and evaluation, take 440 words each for 1,320, because each needs the choice, the reason, the procedure and the outcome. The two remaining aspects, usually covering data preparation and the interpretation or recommendation, take 320 each for 640. Adding 1,320 to 640 gives 1,960 exactly.
Use a fixed order in every technical section: what the data allowed, what you chose, how you validated it, what the numbers were, and what those numbers mean for the original problem. Consistency here is worth real points because each aspect becomes individually findable, which is exactly what aspect based scoring rewards.
Protect budget for limitations. In modeling reports the limitations section is where genuine understanding shows, and it is invariably the section written at midnight in three sentences. Naming the specific weakness of your specific model beats any amount of general caution.
Shape for a machine learning project report
D499 deliverables usually build and evaluate a model against a described problem. These proportions carry that report.
| Section | What belongs there | Share |
|---|---|---|
| Problem and target | What is being predicted or discovered, and what a useful result would let someone do. | 10 percent |
| Data and features | Variables available, what each measures, and which ones entered the model. | 14 percent |
| Preparation | Encoding, scaling, missing values, imbalance handling, and the split into training and held out data. | 16 percent |
| Algorithm selection | The method chosen, the family it belongs to, and the alternative you rejected. | 17 percent |
| Training and tuning | What was fitted, what was adjusted, and how you avoided tuning on the held out data. | 15 percent |
| Evaluation | The metrics appropriate to the problem type, on unseen data, reported honestly. | 18 percent |
| Limitations and close | Where this model fails, and what would improve it next. | 10 percent |
Reporting results so a reader can trust them
Trust in a modeling report is built by disclosure. Say how the data was split and in what proportion. Say whether any tuning touched the held out set, because if it did the reported performance is optimistic and a careful reader will know. Say what the baseline is, since a model has to beat something to be worth reporting, and the simplest sensible baseline is the honest comparison.
Report the metrics the problem type demands rather than the one that flatters. For a rare outcome, overall accuracy is close to meaningless, and reporting it alone reads as either naive or evasive. Showing the breakdown of correct and incorrect predictions by class costs a few lines and answers most questions a reader would raise.
Cite methods and libraries. Algorithms come from the literature and your course materials, and the tools you used come with documentation and versions. In a field that moves quickly, naming the version is not pedantry, since default behavior changes between releases and a result that cannot be reproduced is a result with an asterisk.
Handle the ethical dimension where it exists. Models trained on historical data reproduce whatever pattern that history contained, and a model that affects people needs a sentence about who could be disadvantaged and how you would check. Use the citation style your task specifies, cite at the point of the claim, and keep figures labeled and discussed rather than dropped in.
What earns Competent, and what comes back
Competent reports validate honestly and interpret conservatively. The split is described, the baseline is stated, the metrics fit the problem type, the model is compared to something, and the limitations are specific. They also connect the model back to the original decision it is supposed to support.
Returns follow five patterns. Performance is reported on the training data, which invalidates it. Accuracy alone is reported for an imbalanced problem. The algorithm choice is asserted with no reason and no rejected alternative. Preparation steps that materially change the data are omitted. Or the report ends with a metric rather than an answer, leaving the reader to work out whether the model is useful for the described purpose.
A check before you submit: write one sentence saying what a person could now do that they could not do before, based on your model and its measured performance. If that sentence is hard to write, the interpretation aspect needs work regardless of how good the numbers look.
Because a performance assessment at WGU can be revised without a grade penalty, submit once every aspect carries a genuine answer. If your section also uses an objective assessment, that exam is proctored and our boundary does not move: preparation only, meaning concept drills, metric selection practice and an honest read of your preassessment. Sitting an assessment is not part of what we offer, we take no role while one runs, and portal credentials are never requested or held.
Model built, report not convincing?
Send the D499 rubric and your project. We rebuild it with an honest validation story, the right metrics for the problem type and limitations that are specific.
Six mistakes that cost time in D499
- Scoring on training data. Performance means performance on cases the model has not seen. Anything else is a description of memorization.
- Accuracy on an imbalanced problem. When one outcome is rare, accuracy rewards a model that never predicts it. Report the breakdown instead.
- No baseline. A model needs something to beat. Without a comparison, a number on its own means nothing.
- Tuning against the held out set. Once it has influenced your choices, it is no longer unseen and your reported result is optimistic.
- Algorithm chosen without a reason. Name the family, the reason it fits this data and target, and the option you rejected.
- Ending on a metric. Say what the model lets someone decide, and where it should not be trusted to decide it.
Three questions students ask about D499
Is D499 the undergraduate or the graduate machine learning course?
How much programming does this course expect?
Is D499 the same course as DTSC 3221?
Where D499 sits in WGU's programs
The July 2026 catalog places this code in 2 current WGU programs. 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.