Innovation Series: Demystifying Machine Learning

Dr. Tom Mitchell, former chair of the Machine Learning department of Carnegie Mellon University, offers an elegant definition of Machine Learning in his book. He suggests that, “the field of Machine Learning is concerned with the question of how to construct computer programs that automatically improve with experience.”

Many different types of Machine Learning exist today, but the one that is most widely used for business applications is Supervised Learning. Supervised Learning uses algorithms such as linear and logistic regressions, and multi-class classification, to analyze a series of input variables (X) to produce an output (Y) through a mapping function, think y=f(x).

The parallel of Supervised Learning is one of a teacher and student, where the student is trained on a subject by the teacher. Supervised Learning requires that the algorithm’s possible results be known and that the data used to train the algorithm is labeled with the correct answers.

The majority of Supervised Machine Learning applications usually involve the following steps:

1. Gather the data set to be evaluated
2. Extract the set of parameters and attributes to support predictions
3. Choose the Machine Learning algorithm
4. Train the model
5. Make predictions using the deployed model
6. Adjust parameters to refine the model

Imagine that you work in the FP&A group of a mobile apps developer and are trying to predict the future sales of various mobile apps. There are many variables such as the supported platform, price, global availability, online critic scores, and user reviews which could influence a mobile app’s profitability, and this is where Machine Learning could be used to predict success.

It all starts with the available data that may be used in a model. The more data that is available, the greater the opportunity for the Machine Learning algorithm to come up with the correlation between a set of attributes to improve future predictions.

During the Model Training phase, a Machine Learning algorithm is selected and used to evaluate the gathered data. In our example, we would feed the model data containing selected attributes from a group of mobile apps that were hits and compare it to the predicted outcomes from the model.

The experiments may be repeated using the same data set but with different Machine Learning algorithms to determine which algorithm is most effective at predicting the results. An Evaluation Model is created after the initial Model Training is complete, and when we are satisfied with the efficacy of the chosen algorithm.

We can now deploy the Evaluation Model to make predictions. The real magic or “learning” aspect is when we compare the predicted values from our Evaluation Model to actuals, as they occur over time. Through incremental adjustments, we can refine the model parameters, increase or decrease the resolution of the data set, and the Evaluation Model can rerun the predictions to determine whether the adjustments to the parameters (and/or data) improved the prediction accuracy.

Overall, Machine Learning applications will greatly supplement and enhance FP&A capabilities. These applications will not only drive more effective analysis and higher precision predictive models but it will also allow confidence to be attached to forecasts.

To learn more click here.