Formalizing prediction tasks that lead to value creation

When applying Machine Learning in your company, the 1st challenge is to formalize a Prediction Task that connects to a Value Proposition. I’ve created a new tool to help do this.

In the same way that the Value Proposition Canvas helps formulate a Value Proposition, which is one aspect of the Business Model Canvas, the Prediction Task Canvas helps detail one aspect of the Machine Learning Canvas, which is the Prediction Task.

There are 4 parts to it:

• ENTITY: Object on which a prediction is to be made.

• WAIT: Duration (or event) until an outcome can be observed.

• OUTCOMES: Entity aspect (or events) we can only observe after waiting.

• HEURISTIC: What’s a simple way to predict the outcome, when we first observe the entity?

There’s also a time axis that helps see that the outcome we want to predict at time T can only be observed at time T+W.

The heuristic is useful to provide a baseline that we need our predictive model to beat. It’s also a good way to start thinking about what aspects of the entity need to be taken into account when making a prediction.

​

Here are some simple examples, just to give you a better idea:

• Gmail's Priority Inbox

  • ENTITY: email

  • WAIT for end of next user session

  • Possible OUTCOMES to observe / predict: email discarded without opening, or email opened and archived / replied to / left in inbox

  • HEURISTIC: if sender is in address book, email will be opened

• Real-estate price prediction

  • ENTITY: property

  • WAIT for sale

  • OUTCOME: transaction price

  • HEURISTIC: linear function of surface

• Fake review detection

  • ENTITY: review

  • WAIT for hand-labelling, or visitor to report fake review

  • OUTCOMES: ‘fake’ or ‘real’

• Credit risk

  • ENTITY: credit application + bank customer

  • WAIT for 2 years

  • OUTCOMES: no repayment delay, or all delays < 90 days, or 1 delay of 90+ days

• Email marketing

  • ENTITY: subscriber

  • WAIT for them to open email and consider purchase

  • OUTCOMES: sale, unsubscribe, or none

• Churn

  • ENTITY: customer

  • WAIT for subscription to end in 15 days

  • OUTCOMES: renewal or cancellation

• Fraud Detection

  • ENTITY: transaction

  • WAIT for 45 days, or hand-labelling

  • OUTCOMES: chargeback or none

​

The type of possible outcomes determines the nature of your prediction task, e.g. regression, binary classification, N-class classification... Looking back at the bigger picture, in the Machine Learning Canvas there’s a Decisions block in between Prediction Task and Value Proposition: it’s where you would describe the process for turning the output of the prediction task into value for the end-user.

I hope this will new tool will be as useful to you as it's been for me — let me know in the comments if so!