BUSINFO700 Chap.4 Machine Learning Across the Analytics Stages
Machine Learning Across the Analytics Stages
This chapter connects the maturity stages from DELTA-Plus to the analysis a team can realistically run, following the course reading by Paas (2023).
Analytically impaired firms start with univariate and bivariate statistics; localized firms add if-then segmentation on recency, frequency and monetary value, together with dashboards; firms with analytical aspirations adopt unsupervised learning, such as clustering and association rules, and supervised learning that predicts a behaviour; analytical companies combine models into next best action tables; analytical competitors automate.
You will learn why supervised models measure their inputs in an observation period and the outcome in a later period, why roughly 80% of the data trains a model and 20% tests it, and why simple cross-tables remain the check on complex output. The chapter comes from a set reading rather than a lecture week, and it supports the analytical-techniques factor of DELTA-Plus.
A worked example judges whether a prepaid phone provider is ready for a churn model, and a drill sorts five business requests by stage and method.
What this chapter covers
- 01
Five stages mapped to analytical methods
- 02
Inductive applied analytics versus deductive research
- 03
Rule-based segments and dashboards
- 04
Unsupervised learning for customers and products
- 05
Supervised learning, training and test sets
- 06
Judging readiness for a predictive model
Worked example · free
Choose the stage 3 branch for a churn question after rule-based segments
- 1Rule-based segments and dashboards place the provider at stage 2, localized analytics, so the next methods are at stage 3.
- 1There is a behaviour to predict, leaving or staying, so the supervised branch applies rather than unsupervised clustering.
- 1Measure the explanatory variables, such as plan, tenure and fault calls, at the end of this year (observation period) and record churn during next year (outcome period).
- 1Train the model on about 80% of customers, 8,000, and judge accuracy on the remaining 20%, 2,000, which the model never saw.
- 1Check any surprising effect with a simple cross-table before managers act, and monitor accuracy after launch.
Key terms
- Unsupervised learning
- Machine learning that groups observations without a dependent variable, for example cluster analysis.
- Supervised learning
- Machine learning that uses explanatory variables to predict a recorded outcome, such as churn or response.
- RFM segmentation
- If-then grouping of customers by recency, frequency and monetary value of their purchases.
- Association rules
- Patterns identifying products that are often bought or owned together by the same customer.
- Test set
- The share of data, often about one fifth, held back to measure a model's accuracy on unseen cases.
Machine Learning Across the Analytics Stages FAQ
What is the difference between supervised and unsupervised learning?
Unsupervised learning has no outcome variable; it finds structure, such as customer clusters or products bought together. Supervised learning predicts a recorded behaviour, such as responding to an offer or churning, from explanatory variables. Ask whether there is a target behaviour, and the branch follows.
Should every company aim for deep learning?
No. Paas argues that small bricks-and-mortar firms may reasonably stop at localized analytics with dashboards and rule-based segments, because later stages can cost more than they return. Online start-ups, by contrast, may begin at a more advanced stage. Match the method to the firm's data, people and goals.
Why split data into training and test sets?
A model always fits the cases it learnt from better than new ones. Holding back about 20% of the data as a test set gives an honest estimate of accuracy on customers the model has not seen, and helps reveal overfitting caused by too many variables.
Why do simple statistics still matter at advanced stages?
Cross-tables and histograms let analysts check whether a complex model's effects make sense. A churn model that flags very old customers may only be recording customers who have died, and targeting them would be unethical. Paas stresses that human interpretation stays relevant at every stage.
How can I tell whether a firm is ready for a supervised model?
Check four things drawn from the reading. Is there a recorded outcome, such as a response or a defined churn event? Can inputs be measured in an observation period before a separate outcome period? Have simple statistics already shown the data are clean? Is there a plan to test the model on held-out cases and monitor it after launch?
Assessment move
Draw the stage-to-method map from memory: the five stages down one side and, beside each, the methods Paas assigns to it. Check it against the chapter, then redraw it a day later. For every method on the map, write one business question it answers and one reason a firm at an earlier stage could not yet use it.
Practise the supervised set-up on invented cases by naming the outcome, the observation period, the outcome period and the training and test split. Then take two or three invented firms and judge their readiness for a proposed predictive model, listing what is in place and what must come first.
Finally, connect this chapter to DELTA-Plus: for each stage, describe what the analytical-techniques factor would look like, so that you can justify that score in a case answer with a concrete method rather than a label.
Working through Machine Learning Across the Analytics Stages in BUSINFO700? Sia is AskSia’s AI Statistics tutor — ask any BUSINFO700 Machine Learning Across the Analytics Stages question and get a clear, step-by-step explanation grounded in how BUSINFO700 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
