In this chapter, you learned and applied a set of techniques that enable us to effectively build and finesse datasets for machine learning, starting from very little initial data. These powerful techniques enable a data scientist to turn seemingly shallow datasets into opportunities. We demonstrated this power using a set of customer service tweets to create a travel disruption predictor.

In order to take that solution into production, though, we'd need to add some functionality. Removing some locations in the penultimate step was a questionable decision; if this solution is intended to identify journey disruption risk, then removing locations seems like a non-starter! This is particularly true given that we do not have year-round data and so cannot identify the effect of seasonal or longitudinal trends (like extended maintenance works or a scheduled station closure). We were a little hasty in removing these elements and a better solution would be to retain them for a longer period.

Following on from these concerns, we should recognize the need to start building some dynamism into our solution. When spring rolls around and our dataset starts to contain new climate conditions, it is entirely likely that our model will fail to adapt as effectively. In the next chapter, we will be looking at building more sophisticated model ensembles and discuss methods of building robustness into your model solutions.