Preparing data for a live prediction engine

Question

I need to build a live prediction engine (either using Ensemble methods/ RNN/ Keras Classifier) that could learn from historical data what kinds of users are likely to transact and, in real-time, generate transaction probabilities for current users on a DAILY basis. I do have their demographics, transactional and app activities data. There is an amount limit that can be transacted by the users in a month. How do I prepare the historical data to predict the next day if he is going to transact or not? Thanks.

Demographics:

CustomerID | CustomerAge | Gender | Nationality | Joining Date | Loyalty Program Opt-in | Address

1222 | 25 | M | British | 21-Apr-2016 | Yes | Old kent

Transactions:

Beneficiary | CustomerID| TransactionDateTime | TransactionAmount

4534534 | 545544 | 19/02/2016 16:02:56 | 1888.85

4353222 | 545544 | 23/02/2016 08:49:49 | 3242.43

2453453 | 535455 | 26/02/2016 10:10:56 | 4776.57

5353532 | 435353 | 11/02/2016 12:43:38 | 4862.16

4532522 | 345353 | 23/02/2016 17:42:22 | 4923.54

Answer 1

You'll have to do some feature engineering on your data. Beneficiary and Customer ID has no meaning for machine learning, so you first have to join all other data, taking care of date types (split hour, day of week, month) and should be ordereed by the ones you will not use for training (beneficiary, customerID). than you can create new features like for each transaction the time spent since the last one for each customer, count of transcactions by day/day of wee/month, min, max, median amout spent and son on. It'll be a binary classification (SVM, random forest, logistic regression, etc.).