For business process artificial intelligence (AI) work, data scientists often build teams of algorithms. This is called an ensemble model – taking different algorithms and combining to an end result.
Building a model with just one algorithm doing all the work of analyzing and predicting may work well in some business situations. When the conditions and data are conducive, a single algorithm may be the best solution. In reality, most business processes have so many influencing factors that a single model or algorithm is not likely to capture all of them or estimate their influences well in all situations.
This is also a math issue. Certain algorithms respond differently to changes in values and potential relationships between influencing factors. It is not possible to say one type of algorithm is always better than another in every circumstance.
Data scientists often use different algorithms over the same data set to derive independent predictive models. A few years ago, this was time consuming to do since computing power was more precious and expensive than it is today. Now, it is a common and useful procedure to use the same data to create multiple models in parallel using different machine learning algorithms.
When it comes to making predictions, the results (predictions) of the individual model are combined to come up with a single number. This is usually referred to as creating an ensemble model – the output is a single number that, behind the scenes, is really a combination of other predictions.
There are a few big reasons ensemble models often perform better than a single algorithm model.
As mentioned above, different algorithms have different strengths. This is not just between different data sets but often within a single data set. One algorithm type may be better at predicting values in a certain range but another algorithm type performs better outside that range.
Certainly, a data scientist could spend the time examining the performance of different models over all ranges of all variables. However, that takes a lot of time. Often the models only differ by a small amount in their predictions. Combining them makes an easy solution to smooth out wild predictions or spot small trends not every model may pickup. This is the genius of the ensemble model.
It is not perfect but every prediction will not be perfect anyways. AI for business processes is about probability. The ensemble is an easy way of making predictions a little more accurate in many, many business process models.