An analytics or machine learning workflow is not a linear process. As we saw in some of the examples we provided, discovering certain things about our data or getting disappointing results from our models can force us to go back to square one and think about how we can improve our data. This can mean getting more data, getting better data, or very probably both. Getting more data doesn't always help, since we saw that even not-so-large samples can do a good job of representing the overall population. However, getting more data can be especially helpful if we need to model very unusual behaviors (sometimes called rare events) using a lot of features. Getting better data can refer to two things:
- Using new features in the hope that they will reduce noise in our models. A new feature here can be a new source of data, or a feature that was derived directly from the existing feature set. As we saw in the last exercise, this approach can considerably reduce noise. However, obtaining or engineering new features is not always easy, as it may require a lot of domain knowledge about the business and data sources.
- Doing a better job of cleaning and transforming our existing feature set before we use them to build a model. In this case, what we mean by cleaning and transforming depends somewhat on the model we want to build. Therefore, using this approach requires a certain amount of machine learning expertise.
With its parallel data processing and analytics capabilities,
RevoScaleR can help a data scientist to reduce the time spent on iterating and improving models, and once ready, models can be deployed to a production environment with relative ease. This allows the data scientist to keep their focus on the analytics and modeling and spend less time being side-tracked into the hurdles faced in distributed computing environments.