Skip to content

Which experiment fits my question?

In short

  • Start from your question, not from a technique. Find it in the table below.
  • Most business questions are Classification / regression: predict an outcome you already know for past cases.
  • No outcome column? Use Anomaly detection or Clustering.
  • Products and customers? Pick between Recommendation, Ranking and Association rules by the unit you score.
  • Every kind starts from Experiments → New experiment (Experiments).

Find your question

You want to… Choose Your data needs
Predict yes or no for each row: will this customer leave, is this claim fraud? Classification / regression (binary) One row per case, and a column with the known outcome
Predict one of several categories: which segment, which reason code? Classification / regression (multiclass) The same, with more than two values in the outcome column
Predict a number for each row: price, amount, days to settle? Classification / regression (regression) A numeric outcome column
Predict how a quantity moves over time: next month's demand? Forecast A time column and the quantity; a series ID if you have many series
Find unusual rows without labels: odd transactions, odd tickets? Anomaly detection Rows and their attributes; no outcome needed
Find natural groups without labels: customer segments? Clustering Rows and their attributes; no outcome needed
Choose which products to offer each customer: next best product? Recommendation Purchases or holdings per customer, with the products
Order a list for each decision: which items to show, in what order? Ranking Decisions, the items offered and the one chosen
Learn which items go together across the catalogue: "A and B go with C"? Association rules Baskets: customers or orders and the items in them

Binary, multiclass and regression are one kind in the app. ParleML tells which one you have from the column you predict (the target), and you can override it.

Three checks before you pick

Do you know the answer for past cases? If every row has a known outcome, the question is supervised: classification, regression or forecasting. If not, it is unsupervised: anomaly detection or clustering. Those describe the data; they don't tell you what to do with a group.

Is time part of the question? To predict a quantity's future from its own history, use Forecast. If each row is an independent case that happens to have a date, use classification or regression, and choose a chronological split so the test rows come after the training rows (Experiments).

What do you score: a row, a customer, or a customer–product pair? Classification scores rows. Recommendation, ranking and association rules score the link between customers and items, so the shape of your table matters. Read the kind's guide (below) before you prepare the data.

Close calls

Details: classification or anomaly detection?

If you have reliable labels, even a few, classification uses them. Anomaly detection ranks rows by how unusual they are, which is not the same as how likely they are to be fraud. Use it when labels are missing or arrive too late to learn from. Don't read its scores as probabilities.

Details: clustering or multiclass classification?

If the segments already exist and you want to predict them for new rows, that is multiclass classification. If you want ParleML to propose the segments, that is clustering.

Details: recommendation or ranking?

Both order products for a customer.

  • A recommender trains one model per product and gives each customer a score for every product. Start here for a product-propensity question.
  • Ranking trains a single model over decision occasions, with the candidate product as a feature. Use it when the question is how to order a list at each decision.
Details: recommendation or association rules?

Recommendation scores each customer. Association rules describe the catalogue as a whole, in statements a person can read and audit. The same data can support both.

Details: regression or forecast?

Regression predicts a number for a row from its attributes. A forecast projects a series forward from its own history, optionally with outside drivers.

Before you launch

  1. Write down the question and the intended use. The classification / regression wizard has a Question / hypothesis field for it; the workspace's Business goal holds it for every kind.
  2. Check the data first: column roles, leakage and the signal check. The signal check tells you in seconds whether the data can support the question at all.
  3. Start with a small baseline, then change one thing at a time (Compare).

Learn more about each kind

Kind Guide Worked example
Classification, regression Supervised Binary, multiclass, regression
Forecast Forecasting Forecasting (sales)
Anomaly detection, clustering Unsupervised Anomaly, clustering
Recommendation Recommendation Next-best-product
Ranking Learning to rank Learning to rank
Association rules Association rules Market basket

Next