Improving Predictions and Outlier Detection in BI With Machine Learning



In addition to a solid background in traditional computer science, RelayiQ engineer Nikola Novakovic has been working heavily in artificial intelligence and machine learning for the last several years. He has been able to rapidly develop his skills with these hot new technologies through a mixture of college classes, coursework on line, and lots of hands-on work in the field.


For Novakovic, offering new machine learning models to RelayiQ customers is really exciting! He is especially proud to be able to offer notifications based on patterns and anomaly detection in data. This is a real game changer and a feature that is unique to RelayiQ. According to Novakovic, “There is a whole library of new machine learning models that we have developed. It helps in prediction, forecasts, and anomaly detection.”

Most machine learning models require massive volumes of training data in order to be accurate. I asked Novakovic if customers need to have the massive quantities of training data on hand for RelayiQ’s machine learning models to work effectively. Novakovic explained that with transfer learning (transferring knowledge gained from one data set to another) and grid search (an algorithm that helps your algorithm pick out the right parameters), a lot of parameters can be tuned so that you can get by with less data. He believes that the data RelayiQ customers have in their dashboards will be enough, but as more data is added, the model will only improve. Novakovic would eventually like to have enough data to have five or six models competing against each other to produce the best possible results.


I asked Novakovic what makes machine learning better than just setting a threshold for data-driven notifications using human judgment. He explained that we as humans are inherently biased and bad at predicting the future, and that is where machine learning models come in. “With RelayiQ, we use a few time-forecasting statistical models, and what they are really good at is recognizing patterns in the past and, based on that, making predictions about the future with reasonable accuracy. And based on that, we can set smart thresholds.” Nikola gives the example of a salesperson trying to make predictions in their forecast. “They may not be great at identifying important thresholds because they can’t keep the context of 10 years’ of sales data in their head. It’s really hard for a human being to perform these analyses without a lot of labor and skill to make those kinds of judgments intelligently. They can guess and pick an artificial threshold, but that is not necessarily an outlier. That is where machine learning comes in. With ten years’ of sales data, machine learning can make reasonable projections and set thresholds based on standard deviations to detect outliers and let you know when you are substantially ahead or behind your historical performance.”


I asked Novakovic if end users will need special skills to take advantage of the machine learning and intelligent notifications in RelayiQ. He explained that the basic notifications can run on their own, but he is developing a beta that will enable users to upload .csv data to train their own models specific to their data.


Knowing that AI and ML are cutting-edge technologies, I wondered how novel the work Novakovic was doing for RelayiQ really was. He did not want to “toot his own horn,” but his enthusiasm was palpable when he responded: “This is raw IP, one of those things that differentiates you from everyone else and nobody can say they are doing this. This is definitely one of those things we can show investors to show that we have exciting state-of-the-art IP here at RelayiQ.” Understandably, Novakovic was careful not to share too many details out of fear of giving away too much of the secret sauce.


Finally, I asked Novakovic what his vision is for RelayiQ customers in the long run. Today, people have to dig through dozens -- if not hundreds -- of dashboards to try and find insights. I wanted to know how machine learning and RelayiQ can change that. Novakovic shared his thoughts: “I think it comes down to humans being really bad at making predictions. We shoot ourselves in the foot a lot when we predict things on our own. Machine learning is just math and statistics. But math and statistics is really good at making predictions using historical data. So you can rely on our machine learning algorithms to help you guide your decisions intelligently without any effort. And with that guidance, you will probably be a lot better than with whatever your homemade threshold would have been.”


The new outlier detection ML features are key IP in the new RelayiQ Detect module in their RPA for BI solution. To learn more or set up a demo or free trial of RelayiQ, we’d love to work with you and hear about your use cases.

330 views

Recent Posts

See All