As organizations increasingly turn to machine learning applications to support decisioning, marketing managers must reassess what kind of decision-supporting data a dashboard should display.
A good starting point is the usual primary starting point for creating a dashboard – to know the business objectives of the people who will use it. In many instances, the audience for the dashboard might include an executive or an executive team that has deemed the data project critical to the company. Generically, dashboards should show trends – and deviations – in metrics relevant to the objectives of interest. Does the dashboard quickly identify potential opportunities to make decisions or trigger essential operations? In many cases the dashboard will present metrics associated with one particular medium – a website or an app.
For a machine learning project, however, what might be needed is a set of dashboards covering a number of activities. Machine learning involves building the right training dataset, deciding which models to apply to the set for testing, then reviewing the results to select an optimal model. Effective dashboard options should reveal a flow of information that ultimately reveals the data relationships underlying these activities.
For example, marketers may want to train a machine learning model on sales data from different channels. An algorithm will seek to establish a relationships within the data. A dashboard should display the meta-information on the models that convey the relationship – the advanced statistics from that relationship. It should display a visualization, with capacity to update in real-time, and to display the statistics alongside the visuals so that users can drill down into the details conveniently. Users can then draw conclusions on those results to select algorithms to be used regularly.
A good set of dashboards will also enhance data exploration. It’s critical to machine learning to monitor issues which arise within the training dataset – which data types are in the data, for example; the volume of non-available fields in the dataset; and other format concerns. Fortunately, data visualizations are becoming options on the databases that organizations often use. MongoDB, for example, introduced a dashboard feature called Charts, which allows users to observe visual data patterns from the collections in a database.
A good set of dashboards supports collaboration across a team Many operational practices in data science, let alone machine learning, are linked to dynamic business operations. This means a need for subject matter experts to interpret test and train results against those dynamics. The right dashboards can keep the members of such a team in sync with each other.
If you or your team is unsure of what processes a dashboard set should cover, consider examining the data management process against CRISP-DM (Cross-Industry Standard Process for Data Mining). It’s an industry-proven methodology for data mining efforts. The process, shown below, occurs as a cycle, so comparing an internal data management process against it can reveal ideas about what to monitor, and what collaboration is needed to support a dashboard set for machine learning.
For overall success in managing machine learning initiatives, marketers shouldn’t be seeking one dashboard “to rule it all” and instead develop a series of interfaces that highlight the logical relationships between the data and business objectives. The best dashboard set offers a clear view of statistical patterns that highlight how a model can innovate for the benefit of products, services, and customer experience.