BUSI2001 Chap.7 Predictive Analytics and Machine Learning
Predictive Analytics and Machine Learning
Predictive analytics uses historical data and machine learning to forecast outcomes such as which customers will leave. This chapter distinguishes supervised, unsupervised and semi-supervised learning by the labels each needs, and compares the three models the churn exercise uses, from a random forest to Naive Bayes and logistic regression.
It explains why models are trained on one file and scored on another, and how a confusion matrix separates caught churners from missed ones and false alarms. The worked material shows that the most accurate model is not always the best business choice once the cost of each error is counted, and the practice closes with checks before a monthly flagging system goes live.
Throughout, the emphasis stays on interpreting results for a manager rather than on the mathematics inside each algorithm.
What this chapter covers
- 01
Predictive analytics uses
- 02
Supervised learning
- 03
Unsupervised learning
- 04
Semi-supervised learning
- 05
Logistic regression
- 06
Naive Bayes
- 07
Random forest
- 08
Confusion matrix
- 09
Precision and recall
Worked example · free
Placing three business tasks in the right family of machine learning
- +1(a) is supervised, because every past transaction carries a label to learn from.
- +1(b) is unsupervised, because the goal is to discover clusters without labels.
- +1(c) is semi-supervised, because a small labelled set can be combined with many unlabelled records.
Key terms
- Unsupervised learning
- Machine learning that finds structure such as clusters or anomalies in data without labels.
- Random forest
- A model that averages the predictions of many decision trees built on random subsets of the data.
- Precision
- The share of cases a model flags as positive that really are positive.
- Confusion matrix
- A table that crosses actual outcomes with predicted outcomes to count each kind of correct and incorrect prediction.
- Overfitting
- A model learning noise in its training data so closely that it performs poorly on new data.
- Binary target
- An outcome variable with exactly two possible values, such as churned or stayed.
- Testing file
- Held-out data used only to judge a trained model, never to fit it.
Predictive Analytics and Machine Learning FAQ
Why train on one file and test on another?
Scoring a model on the same data it learned from flatters every algorithm, because it may simply have memorised those cases. The churn exercise therefore fits models on a training file and judges them on a separate testing file, which gives an honest estimate of how each will perform on customers it has never seen.
Is accuracy enough to compare churn models?
Rarely. When most customers stay, a model that flags nobody can still score high accuracy while catching no churners. Report precision and recall from the confusion matrix as well, and choose between models using the business cost of a missed churner compared with a wasted retention offer.
What should the Kaggle classification task include?
Choose a dataset with a binary target, convert ordinal and nominal variables with label or one-hot encoding, apply every algorithm covered in the topic, and compare their performance in a table. Save the work as a presentation deck, upload it to Discussions, and be ready to explain what the differences mean.
What is a flagging system in the churn exercise?
It is the final stage of the exercise, in which a trained model scores new customers and flags those likely to churn so the business can act. Building it requires the new data to be prepared exactly like the training data, a threshold chosen from the business cost of errors, and regular checks on performance.
Assessment move
Learn this chapter by computing, not memorising. Draw a confusion matrix for an invented test set and calculate accuracy, precision and recall by hand until you can do it in under two minutes. Then change one cell, for example by catching twenty more churners at the cost of forty more false alarms, and recompute to see which measures move.
Practise the cost comparison from the worked example with different cost ratios, so you understand when the least accurate model becomes the best choice. For the workshop task, keep a single table that lists each model with the same measures on the same test data, and add one sentence on what the differences mean for the business.
Before treating any model as finished, write down how its inputs were encoded, which variables could act as unfair proxies, and when it should be re-evaluated. Explain the difference between precision and recall to someone outside the course in two sentences without using either word. If you can do that clearly, you understand the trade-off well enough to justify a model choice in writing.
Working through Predictive Analytics and Machine Learning in BUSI2001? Sia is AskSia’s AI Business Analytics tutor — ask any BUSI2001 Predictive Analytics and Machine Learning question and get a clear, step-by-step explanation grounded in how BUSI2001 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.