How to Approach the BBIM708 Machine Learning Portfolio
Quick answer: The BBIM708 Machine Learning portfolio rewards clear reasoning more than high scores. Pick simple, open datasets, show every script and result as a labelled screenshot, split data properly before tuning, and interpret each metric and fold in your own words. Meeting each item’s word count and keeping the test set untouched earns most marks.
If you are staring at the BBIM708 Assessment 2 portfolio question, the volume of tasks is the first shock: two models in Part A, then Grid Search, Random Search and cross-validation in Part B. A worked layout is in our BBIM708 Machine Learning and AI sample, for reference only. This guide explains how to plan your BBIM708 Machine Learning portfolio so nothing is missed.
Stuck on the code or the write-up? You can get a second pair of eyes through our do my assignment service for coding-heavy coursework.
What the BBIM708 Machine Learning portfolio really asks
This is an individual portfolio worth 60% of the module and 100 marks, assessed against two learning outcomes: evaluating machine learning algorithms and applying fine-tuning methods. You choose your own AI application and an open dataset from sources such as scikit-learn, Kaggle or the UCI repository. The submission is a Word or PDF document with screenshots of scripts and results, plus a zipped folder of scripts and data. APA 7th is required, and quoted or paraphrased material must stay within 15%.
To pass you must attempt every part and reach at least 50%. That matters for planning: a missing section costs more than an imperfect one.
Where the marks are
| Part | Task | Marks |
|---|---|---|
| A (a–c) | Principles, suitable problems, strengths and limitations for one supervised and one unsupervised algorithm | 12 |
| A (d) | Training code screenshots with explanation | 4 |
| A (e) | One diagnostic visual per model plus about 75 words of interpretation | 10 |
| A (f) | Two metrics per model in tables or charts | 10 |
| A (g) | Interpretation, about 150 words per model | 8 |
| B (a) | Grid Search on a classifier | 18 |
| B (b) | Random Search on a regressor | 18 |
| B (c) | Five-fold cross-validation on both models | 20 |
Part B carries 56 of the 100 marks, so do not spend your best energy polishing Part A at its expense. Cross-validation alone is worth 20.
How to approach each task
Choosing algorithms and datasets
Keep it simple. A Random Forest or Logistic Regression for the supervised model and K-means for the unsupervised one are easy to explain and have mature scikit-learn support. The breast cancer dataset and Iris load in one line, which leaves your time for analysis. Avoid huge Kaggle files that take an hour per search; a slow notebook is a risk, not a flex.
Check the dataset suits both parts. For Part B you need a classification target for Grid Search and a numeric target for Random Search, so you may use two datasets, as long as each is described properly.
Part A: principles, code and diagnostics
Write the 100-word answers from your own understanding, not from the library documentation. Markers can spot paraphrased textbook sentences. For strengths and limitations, give each one a concrete example tied to your data, such as K-means struggling when clusters are elongated.
For the diagnostic visual, use a confusion matrix or ROC curve for the classifier and a cluster plot for K-means. Label axes, add a caption and then spend the 75 words saying what the picture shows, not what a confusion matrix is.
Grid Search versus Random Search
Grid Search tries every combination you list, so with two hyperparameters and two values each you get four candidates. It is easy to read and report in a table. Random Search samples from ranges or distributions for a set number of iterations, which suits broader ranges where trying every value is wasteful.
The brief expects you to show both and compare. Explain why you gave the classifier a grid and the regressor a random search, and say honestly if the gain over default settings was small. A modest improvement, explained well, scores better than an exaggerated one.
Cross-validation and stability
Report every fold, not only the average: accuracy and F1 for the classifier, RMSE and R² for the regressor. Then use the roughly 200 words to discuss spread. If one fold is far from the others, say what might cause it, such as class imbalance or a few extreme values. Stability is the point of the exercise.
Want the report side checked as well? Our report writing help for technical portfolios covers structure, captions and APA formatting.
A step-by-step plan
- Read the brief and build a checklist of every item with its word count and marks.
- Choose datasets and algorithms, and confirm they load and train in a few seconds.
- Do the train-test split first and set a random seed, then leave the test set alone.
- Complete Part A code, capture plots and write the interpretations while the results are fresh.
- Run Grid Search, then Random Search, saving the tables as you go.
- Run cross-validation and build the per-fold table.
- Write the explanations, add APA citations and trim any heavily paraphrased text.
- Take final screenshots, zip the scripts and data, and proofread the document.
With a two-week window, give the first three steps two days, Part A three days, Part B five days, and keep the last days for writing and checks.
Mark-maximising checklist
- Every screenshot has a figure number, a caption and a sentence in the text that refers to it.
- Code screenshots are readable at normal zoom and show the output beneath the code.
- Each word count is respected; markers notice both padding and thin answers.
- Hyperparameter tables list every combination tried and mark the best one.
- Test-set results are reported once, after tuning is finished.
- Your zipped scripts run top to bottom without errors.
- Quoted and paraphrased text stays under the 15% limit.
Common mistakes to avoid
| Mistake | Why it costs marks | Fix |
|---|---|---|
| Data leakage | Scaling or selecting features on the full dataset before splitting inflates results | Fit preprocessing on training data only, or use a pipeline |
| Tuning on the test set | The final score is no longer an honest estimate | Tune with cross-validation on training data; test once |
| Unlabelled screenshots | Markers cannot tell which model or step they show | Caption every image and cite it in the text |
| Reporting only accuracy | Hides class imbalance problems | Add F1 or a confusion matrix |
| AI-written explanations | Generic text is easy to flag and says nothing about your data | Write from your own results and name your actual values |
| Ignoring negative scoring | RMSE appears as a negative number in scikit-learn searches | Flip the sign before reporting |
Time pressure is the usual reason these slip in. Students run code late, paste output in a hurry and skip the interpretation. Doing a short write-up after each experiment avoids that.
Already working on the business intelligence side of your programme? Our BBIM601 Assessment 2 business intelligence project guide shows a similar planning approach.
Short on time before the deadline? Get structured support from university assignment help for postgraduate coursework.
FAQs
Can I use the same dataset for Part A and Part B?
Yes, if it suits both a classifier and a regressor, but most students use one classification dataset and one regression dataset. Describe each clearly, since Part B asks for a dataset description of around 150 words.
How many hyperparameters should Grid Search use?
Two, each with two values, giving four combinations that you evaluate in full. Random Search uses two hyperparameters over broader ranges, so you sample a set number of combinations instead.
Do I need to report the test set score?
Yes, report the best configuration’s metrics, but only after tuning. Use cross-validation scores to choose settings, then evaluate on the held-out test data once.
What counts as a good interpretation of fold results?
One that discusses spread, not only the mean. Comment on the best and worst folds, possible causes and what the variation means for how far you can trust the model.
Are screenshots enough, or must I submit code too?
Both. Screenshots of scripts and results go in the document, and the scripts and data go in a separate zipped folder, as the brief requires.
Is it safe to use AI tools to write my explanations?
It is risky. Generic text can be flagged and rarely refers to your actual results. Use tools, if at all, for debugging, and write the interpretation yourself.
If you want expert eyes on your scripts or write-up before submission, you can upload your files through our assignment submission page, or look at Auckland assignment help if you are studying in the city.
About the expert
Harriet Blackwell is listed as a homework writing specialist with a Master’s in Business Analytics from the University of Auckland. Her profile covers tutorial questions, case exercises and problem sets for business, commerce and accounting papers.
