D789 Applied Machine Learning for Business Solutions is banner number ITSW 6106 and three competency units in the School of Technology, at the graduate level. The catalog describes it as familiarizing students with machine learning through the lens of a software engineer, which is a specific and useful framing. The competency is not deriving algorithms. It is deciding whether a business problem is a machine learning problem at all, choosing an approach, evaluating it with a metric that matches the business cost, and building something that can run in production.
Framing the problem is most of the work
Through the lens of a software engineer means the questions are engineering questions. Is there a decision being made repeatedly that could be automated. Is there labeled history of that decision. What does a wrong answer cost, and does the cost differ by direction. Could a simple rule do it. Answering those four before touching a model is what separates an applied paper from an academic one, and graduate aspects usually ask for the reasoning explicitly.
Problem type follows from the framing. A decision between categories is classification, a quantity is regression, grouping without labels is clustering, and ordering options is ranking. Each carries different evaluation and different failure modes, and misidentifying the type invalidates everything downstream.
The metric is where business and technique meet, and it is the most common weakness in student work. Accuracy is nearly useless on imbalanced data, since a fraud detector that flags nothing is accurate almost all the time and worthless. Precision and recall trade against each other, and which one matters depends entirely on the relative cost of a false positive against a false negative in the business context. A paper that reports accuracy for an imbalanced problem has skipped the reasoning the course is about.
Data leakage is the failure that quietly produces excellent results and cannot survive production. Any feature that would not be available at the moment of prediction, any statistic computed over the whole data set before splitting, and any time-ordered data split randomly will inflate your numbers. Checking for leakage explicitly and saying you checked is a mark of engineering maturity that evaluators notice.
Turning scored aspects into a workflow and a word budget
Each aspect is scored on its own at WGU and needs a 2. Here the aspects usually track the stages of a machine learning workflow, so map each to a stage and decide what evidence that stage produces before you begin.
Budget it. Suppose eleven scored aspects and a report of about 2,200 words with tables and charts. Eleven into 2,200 is 200 each. Sort them. Two frame the business problem and justify the approach, at 300 each, or 600. Two cover data understanding and preparation, at 250 each, or 500. One covers model selection with alternatives considered, 300. Two cover evaluation and metric choice, at 280 each, or 560. One covers deployment or integration considerations, 200. Three are shorter, an introduction, a chart reading and a conclusion, at 70 each, or 210. That totals 600 plus 500 plus 300 plus 560 plus 200 plus 210, which is 2,370, so trim the preparation sections to 200 each and the paper lands near 2,270, with framing and evaluation still carrying half.
Decide the metric and the baseline before you train anything. A baseline can be a simple rule or the current manual process, and a model that does not beat it has not made the case for itself no matter how sophisticated it is.
On the D789 project now?
Send the aspects and the business problem. You get a framing, a metric chosen against the cost of each error type, and an evaluation plan.
Matching the metric to what an error costs
| Business situation | Expensive error | Metric to lead with |
|---|---|---|
| Screening for a rare failure before it happens | Missing a real case | Recall, with precision reported alongside |
| Flagging transactions for human review | Wasting reviewer time on false alarms | Precision at the volume the team can actually review |
| Both error types carry real cost | Neither dominates | A combined measure, with the trade-off curve shown |
| Ranking leads or recommendations | Good options buried below the cutoff | A ranking measure at the position users actually see |
| Forecasting a quantity | Large misses matter more than small ones | An error measure that penalizes large deviations, in business units |
| Imbalanced classes of any kind | The rare class is the one that matters | Anything except plain accuracy |
Write the cost of each error type in the business's own terms, in currency or hours, before choosing the row. That sentence is what turns a metric choice into a justification an aspect can score.
Honest evaluation, data ethics and citation
Evaluation is only credible when the protocol is stated. Report how the data was split, whether the split respected time ordering, whether validation was used for tuning separately from the final test, and how many runs the reported figure represents. A single number with no protocol behind it is the weakest form of evidence in this subject.
Data ethics belongs in a business-facing paper. Say where the data came from and whether its use is permitted. Say what personal information it contains and what you did about it. Examine whether performance differs across groups the business cares about, because a model with strong overall numbers and uneven performance is a business risk rather than a success, and identifying that is the kind of judgment the graduate framing expects.
For sources, established machine learning texts, library documentation for behavior claims and published work on evaluation practice are the strong citations, in APA where your program requires it. Cite the source of any data set. Where you adapt example code, attribute it, and be able to explain every transformation, since an aspect asking why a feature was engineered a particular way is unanswerable about somebody else's pipeline.
What clears, and what returns
Work at WGU is Competent or Not Competent, with no letter grades, and a returned performance assessment can be revised freely. Inside a six month flat rate term the days are the cost, so a first submission with a stated baseline and a defensible metric is the efficient target.
Reports that clear frame the business problem before the technique, justify the metric against error costs, compare the model to a baseline, state the evaluation protocol, and discuss what would have to be true for the model to keep working after deployment. They also say when the answer is not machine learning, which is a legitimate and sometimes correct conclusion.
Returns come from four habits. Accuracy reported on imbalanced data. No baseline, so improvement is unquantified. Leakage unexamined, which usually shows up as results that are too good for the problem. And a model presented without any account of how it would be integrated, monitored or retrained, which leaves the applied half of the course unaddressed.
Six mistakes that cost D789 students time
- Choosing the technique before the problem is framed. The interesting decision is what kind of problem this is and whether a model is warranted at all.
- Reporting accuracy on imbalanced data. It is the metric that makes a useless model look excellent, and evaluators check for it.
- Splitting time-ordered data randomly. The model then learns from the future, and the reported performance cannot be reproduced in production.
- Tuning on the test set. Once the test set has influenced a choice, it is no longer a test set, and the final number is optimistic.
- Skipping the baseline. A simple rule or the existing process is what the model has to beat, and without it there is no case to make.
- Stopping at the model. The applied framing expects integration, monitoring and a plan for when the data shifts under the model.
How we work on this course
D789 support starts at framing. Send the scored aspects and the business problem and you get a problem framing with the error costs written in business terms, a metric chosen and defended against those costs, a baseline to beat, an evaluation protocol that survives a graduate reading, a leakage checklist, and a model report in the register WGU evaluators expect. If an objective assessment sits in your course, readiness coaching is the limit of what we provide, and because those sittings are proctored we take no part in one and never handle your portal login.
Applied machine learning papers go wrong early and expensively, because a framing error invalidates everything after it. Fixing the framing first is what keeps this course inside one six month flat rate term.
Three questions D789 students ask
Is D789 the same course as ITSW 6106?
How much mathematics does this course need?
Can I use any data set I like?
Where D789 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.