The February 2019 release of Microsoft Power BI includes an interesting preview visual. It’s the Key Influencers Visual, which incorporates artificial intelligence. Among other things, this allows Power BI to determine which customer characteristics influence purchasing behavior. All you need is a dataset and the free-to-download Power BI Desktop. Does this mean we can finally say goodbye to the extensive data mining functionality in Microsoft SQL Server Analysis Services? That functionality was marked as “deprecated” in SQL Server 2017 and will no longer be available in future versions. Additionally, the question arises as to whether Azure Machine Learning Studio is already obsolete as well. Let’s compare the tools.

Key Influencers Visual in Power BI

In the examples, a dataset from Microsoft was used for a targeted mailing campaign.

To use the Key Influencers Visual, you must check the box next to this visual under the "Preview Features" heading in the Power BI Desktop Options. This will make the icon shown below appear in the Visuals section. This feature is available starting with the February 2019 update of Power BI Desktop.

The dataset for the Targeted Mailing Campaign contains customer attributes and an indicator showing whether or not a bicycle was purchased in the past (BikeBuyer). The BikeBuyer field must be in the “Analyze” segment of the Key Influencers Visual, and the customer attributes must be in “Explain by.”.

Power BI will use this to determine the influence of customer characteristics on whether or not a bicycle is purchased. The result is immediately visible, with a “Bachelor’s” degree appearing to have the greatest influence on the purchase of a bicycle (“What influences BikeBuyer to be 1”).

To effectively select customers for a targeted mailing campaign, it is important to consider multiple customer attributes. In the Key Influencers Visual, you can therefore view segments (combinations of field values across multiple customer attributes).

The percentage and population size are displayed, after which you can zoom in for more details. So it doesn’t just show the result, but also provides information on how it was calculated. If necessary, you can determine whether it would be useful to further subdivide a segment by adding other fields.

SQL Server Analysis Services Data Mining

All in all, the Key Influencers Visual does a good job of explaining how the result was derived. However, the built-in method cannot be further customized. This is quite different in SQL Server Analysis Services Data Mining. There, various mining models can be applied, such as Decision Tree, Naive Bayes, and Clustering. Each model offers a wide range of configuration options, and data subsets can be used to create branches. Analysis Services Data Mining can also determine which model appears most suitable for the problem at hand. For the dataset used, the so-called Lift Chart points to the Decision Tree model.

You can navigate through the models, or narrow or widen the field of view, as shown below for the Decision Tree. The color intensity indicates the degree of influence, and the number of levels and histograms help you navigate through the Decision Tree.

The Equation

You can save the results of each data mining run to, for example, a database for future use or as a reference. In Data Mining, you can further test and train the models to increase their accuracy even more. Speaking of accuracy, let’s compare the results of the Key Influencers Visual with those of the Data Mining exercise to see how they relate to each other.

I had the Decision Tree results from Data Mining exported to a database and added them to Power BI as a second query. This dataset excludes non-bicycle purchases, which makes sense for use in a marketing campaign. This data was then filtered based on the prominent attributes in the Key Influencers Visual. On the one hand, for the best Buyer Segment, and on the other hand, for the best Non-Buyer segment. The results show that no Data Mining records are returned when filtering for Non-Buyers, which is consistent with the underlying concept. Conversely, a substantial subset of records is returned with an average probability of 71%. This demonstrates that there are no extreme discrepancies between the methods of the two tools.

Alternatives

The Key Influencers Visual isn’t the only tool Microsoft offers. Starting with version 2017, Machine Learning Services (Python & R) have been added to SQL Server, and these are also being gradually integrated into Azure SQL Database. The standalone Microsoft R Server has been renamed Machine Learning Server following its expansion to include Python. In addition to the preview visual, Power BI can utilize Azure Cognitive Services and custom models created in Azure Machine Learning. Azure Machine Learning itself goes much further, offering features such as image recognition and sentiment analysis of text fields, in addition to standard statistical operations. As shown below, Azure Machine Learning Studio provides a user-friendly interface that allows you to use components similar to those found in Data Mining in Analysis Services.

Conclusion

At first glance, the Key Influencers Visual in Power BI seems to perform somewhat similarly. However, the model within the visual cannot be further customized and is therefore unsuitable for in-depth predictive needs. For end users who simply want a first impression of influential values, the Key Influencers Visual works just fine. However, a “dash of artificial intelligence” in an end-user tool is no substitute for the use of statistical models. The recommendation remains to consult with statisticians or data scientists after forming an initial impression, after which one can seriously set to work on making sound predictions. There are countless tools available to support that process.

Azure Cognitive Services already includes some models that can be used in Power BI. Azure Machine Learning is capable of much more, but Azure must fit within the cloud strategy—if there even is one. The results from Azure Machine Learning cannot simply be written to a database, and for now, it is not possible to connect to an Analysis Services cube. However, it is possible to achieve seamless integration with business processes.

If a company wants to continue using an alternative to Analysis Services Data Mining on-premises, SQL Server Machine Learning Services offers advantages for Python and R developers, but it lacks a user interface and is therefore quite technically intensive. My conclusion is that data mining as a concept is more alive than ever, including at Microsoft. The gap left by the Analysis Services Data Mining tool has been almost completely filled in the cloud—and in some cases even surpassed—but on-premises, certain capabilities will be missing starting with SQL Server 2019. The question, of course, is how big of a problem this will be once everyone is in the cloud anyway…