Predicting categories from data having no targets

I have training data that just contains transaction history of a store which includes user id of the customer, the product purchased and the cost of the purchase. There are repeated transactions from same customer at different entries, only one product is bought at once. What I have to predict from this are the top 3 categories of products that the user might purchase from in the future. The target is not in ‘top-3 in future’ format. The so called target is just one product that a user buys in future for each user.

What I have thought is I’ll train a neural network with the training data and make generalized top-3 predictions for a general user based on the budget. Then I take this data and train it with the instances for each user separately extracted from the training data itself so that the general network gets biased for each user based on their average purchase budget in training data and type of products bought by them. Such biased networks will be produced for each user and be used to make the top 3 predictions.

I don’t know if this is effective, moreover how do I use the training targets, I have no idea about the approach and implement optimization for this, how do I go about this problem ?