We're using randomForest and R to do some forecasting for labor schedules. Specifically, we are using randomForest to do:

  1. Forecast / predict the number of staff needed (shifts).
  2. Forecast / predict the start time of each shift (start time).
  3. Forecast / predict the end time of each shift (end time).
  4. Finally, forecast / predict the actual staff member.

In some cases, this works quite well. In other cases, it's not working well at all. For example, for one customer who is always closed on Sundays, the software is predicting staff every Sunday.

We've attempted to expand the algorithm to identify staffing needs on say, Sunday only, but find that this creates some other problems in our forecasting.

  • Any good resources for this type of predictions / forecasting?
  • Is there something better than randomForest that we should be using?
  • Is there a better way to do this?

We don't have a strong background in O/R or machine learning ...

asked Sep 27 '10 at 18:35

Anthony's gravatar image

Anthony
16112

You might want to cross-post this on http://www.or-exchange.com/ to see if anyone in OR might have any ideas.

(Sep 28 '10 at 00:45) Phillip Mah

One Answer:

You might want to investigate multi-task prediction and consider adding inputs (or outputs) for whether the customer is open the next day. I recommend Caruana's Multitask Learning paper as a good entry point for thinking about multitask modeling. The advantage of a multitask model is that it can learn the structural constraints between the different predicted outputs, which can be helpful for avoiding logical violations (e.g., the customer is not open Sunday) and improving prediction accuracy.

answered Sep 28 '10 at 13:01

Art%20Munson's gravatar image

Art Munson
64611316

Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.