1
$\begingroup$

I am working on a time series classification problem to identify what items customers would buy in their next order (customers orders different products every week).

Let's say we have a customer who orders every week but different items. Overall, this customer ordered 1k items in the past, and my task would be to identify approximately 80-120 unique items (number of items customer order each week was between 80-120) from these 1k items.

Right now, I am using the xgboost classifier, and my precision is close to 70%, and recall is 65%. I understand xgboost cannot handle seasonality by default. So I created additional features like a week of the month, week of the year, season (Spring/summer/winter/fall), etc.

I wanted to know if anyone in this community worked on a similar kind of problem. I have seen lot of peoples posting same kind of problem but not sure whether they ended up being helped/ moving it to production.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

Try association rules, they will help you in finding the right set of rules like what is chance of customer buying X if he purchased A, B, C in the past.

Also from my personal experience, these kinds of tasks are more business need driven. Like sales capture from the item. So even if your model is 60% accurate but captures 90% sales, you should be good to go. Try to give it a thought in this way.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks, @shivam shah I already have association apriori algorithm integrated to my ML model and coupling items customers bought together more times. 60% of accuracy might be good to go in most of the cases like in the recommendation system, but I am trying to automate the whole ordering system with better precision. $\endgroup$ Commented Oct 4, 2019 at 2:59

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.