D801 Machine Learning for Computer Scientists, catalog number ICSC 6206, is the four-CU graduate course in the WGU School of Technology that covers the foundational algorithm families and what they are good for. The part students underestimate is that a machine learning submission is an experimental report, not a program. A model that scores 0.94 on a metric proves nothing until the reader knows what the baseline was, how the data were split, what leaked and what the number means for the decision the model is supposed to support. The code is the easy half.
What ICSC 6206 is actually testing
The first competency is problem framing. Before any algorithm is chosen, a graduate submission has to state what is being predicted, from what, for whom, and what a good prediction is worth compared with a bad one. Fraud detection where a missed case costs a thousand times more than a false alarm is not the same problem as recommending a film, even when both are binary classification with the same accuracy. Aspects that ask you to justify an algorithm choice are really asking whether you understood the problem before you reached for a library.
The second is honest evaluation. Most of the ways a machine learning report goes wrong are evaluation failures rather than modelling failures: scaling fitted on the whole dataset before splitting, a target-derived feature quietly included, hyperparameters tuned against the test set, class imbalance hidden behind accuracy, time series shuffled so tomorrow trains the model that predicts yesterday. Evaluators in a graduate course look for the split described before the results are shown, because the description is what makes the results readable.
The third is interpretation. A confusion matrix is not a conclusion. What the numbers imply for use, where the model fails, which errors are tolerable and which are not, and what would have to be true before anybody deployed it: that is the reasoning the rubric aspects reward, and it is usually the section written last and shortest.
Turning scored aspects into a report plan
Scoring detail for D801 lives in your Course of Study, not in the public catalog. Open it, count the scored aspects and treat the list as a specification. Each aspect is scored on its own three-point scale and each needs a 2 to pass, so a beautiful modelling section cannot compensate for an evaluation section that skipped the split.
Where the course is assessed by a performance assessment, give every aspect a heading in the rubric's language. Where a submission includes code or a notebook, keep the written document self-contained: an evaluator should be able to score every aspect without executing anything.
The word budget, worked. Assume seven scored aspects and directions asking for roughly 2,500 words. Reserve 150 for a framing opening and 120 for a closing recommendation, leaving about 2,230 for scored content, or 320 words an aspect. Then weight it. Evaluation methodology and results interpretation are the two aspects where returns concentrate, so take 60 words from each of three descriptive aspects and add 90 to each of those two. Data preparation deserves protection as well; it is the aspect students compress into a sentence and the one that most often hides the flaw that invalidates everything downstream.
Tables and figures buy you words back. A results table with metrics for three models across two splits replaces four paragraphs of prose, but each table still needs a caption and a sentence saying what the reader should notice in it.
A structure that fits a machine learning report
Follow the task directions where they prescribe an arrangement. Where they do not, this order matches how machine learning aspects are usually written and keeps the evaluator from hunting.
| Section | What belongs in it | What earns the aspect |
|---|---|---|
| Problem and value | The prediction target, the decision it supports and the cost of each error type | Scored for specificity; a target defined in one sentence beats a page of context |
| Data | Source, size, features, class balance, missingness and any licensing or privacy limits | Scored for candour about the data you have rather than the data you wanted |
| Preparation | Cleaning, encoding, scaling and feature work, with the point of the split stated | Scored for order of operations; fitting transforms before splitting is a defect |
| Baseline | The simplest sensible predictor and its score | Scored where named, and it makes every later number mean something |
| Models and rationale | Algorithms tried and why each suits this data and this cost structure | Scored for reasoning, not for the number of models attempted |
| Evaluation protocol | Split or cross-validation scheme, tuning procedure, metrics and why those metrics | Scored for validity; the section that decides whether results can be believed |
| Results | Metrics per model in a labelled table, with the errors characterised | Scored for completeness; one headline number is not a result |
| Limitations and use | Where the model fails, who it could harm, what deployment would require | Scored for judgment; hedging with no specifics reads as unmet |
State your random seed and your library versions once, in a line. Reproducibility is cheap to provide and expensive to be missing when an evaluator cannot reconcile two numbers in your own report.
Evidence craft when the evidence is a metric
Machine learning writing has a peculiar hazard: numbers look like evidence even when nothing supports them. These habits keep a graduate report defensible.
- Report the baseline first. On a dataset with ninety-four percent negatives, a classifier at ninety-four percent accuracy has learned to say no.
- Choose metrics from the cost structure and say why. Precision, recall, the balance between them, area under a curve and calibration answer different questions, and quoting all of them without a reason is not thoroughness.
- Describe the split before the scores. Which rows were held out, how many, and whether the split respects time, grouping or subject identity.
- Name every source of leakage you checked for. A short paragraph confirming that scaling and imputation were fitted inside the training fold buys more credibility than another model.
- Characterise the errors, not just their count. Which class, which region of the feature space, which subgroup, and whether the failures share anything.
- Cite the dataset and any algorithm reference in APA. A public dataset has a provider, a date and terms of use, and using it without attribution is a citation failure whatever the code does.
One sentence separates the strong reports: a plain statement of what this model must not be used for. Graduate evaluators read that as maturity, and it is the natural home for the fairness and privacy considerations that a rubric may score separately.
What separates Competent from a submission sent back
Because each aspect stands alone, most D801 returns are narrow and mechanical to fix.
- A baseline exists, and every reported score is read against it.
- The evaluation protocol is described before the results and is valid for the data type.
- Preprocessing is stated in order, with fitting confined to training data.
- Metric choice is justified by what an error costs in this problem.
- Results appear in labelled tables, and the prose says what to notice in each.
- Limitations name specific failure conditions rather than gesturing at more data.
Performance assessment work at WGU can be revised and resubmitted with no grade penalty, which matters more in machine learning than in most subjects, because a return often exposes a methodology flaw worth catching. The cost is calendar time inside a six-month flat-rate term, and a four-CU course reopened twice is a course that did not close.
If any part of the course is assessed by a proctored objective assessment, that exam is yours to sit. We prepare only: concept drilling, worked problems on algorithms and metrics, and an honest readiness call. We never sit assessments and never ask for portal credentials.
Six mistakes that cost time in D801
- Tuning against the test set. Once hyperparameters are chosen by test performance, the test score is a training score wearing a disguise.
- Reporting accuracy on imbalanced data. It hides the only errors anybody cares about and invites a return on the evaluation aspect.
- Skipping the baseline. Without it, no reader can tell whether the model added anything at all.
- Shuffling time-ordered data. Random splits on a temporal problem leak the future into the past and produce scores that cannot survive contact with reality.
- Letting the notebook be the report. Cell output pasted into a document is not analysis, and an evaluator scoring an interpretation aspect needs sentences.
- Writing limitations as a formality. Two lines about needing more data score nothing. Name the subgroup where the model fails and the condition under which it should be retired.
How support works on this course
Send the task directions plus what your Course of Study says about scoring, and if you have started, the dataset description and what you have run so far. What comes back is an aspect-mapped report plan, a data preparation order that survives scrutiny, an evaluation protocol chosen for your problem, and a results section that reads as analysis instead of output.
Most students on this course do not need more modelling. They need the write-up to say clearly what they already did, in the order an evaluator scores it, which is usually one focused session rather than another week of experiments.
Questions students ask about D801
Is D801 the same course as ICSC 6206?
How much mathematics does D801 assume?
Can you build and submit the model for me?
Model built and the report will not come together?
Send the task directions and your Course of Study rubric. You get an aspect-mapped plan, a defensible evaluation protocol and a results section written as analysis.
Where D801 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.