This can be a critical step of the general endeavor because it determines the trouble needed to integrate the dashboards inside a cluster right into a single entity. The more disparate the dashboards inside a cluster, the more effort and time needed to mix them right into a single unit. We’ll walk through a case study, where we wish to consolidate seven dashboards (shown in Figure. 1) into 2–3 groups.
A series of steps is advisable for the clustering:
1) Understand the aim of every dashboard by talking with current users and developers. This voice of customer is crucial to capture at an early stage to facilitate adoption of the consolidated dashboards. We might also unearth latest information in regards to the dashboards and have the ability to update our initial assumptions and definitions.
2) Assign weights to the various dimensions — as an illustration, we should want to assign a better weightage to metrics over the opposite aspects. In our example above, we give metrics a 2x weightage vs. the others.
3) Convert the data right into a dataframe conducive for applying clustering techniques. Figure 2 shows the dataframe for our case study accounting for the suitable weights across dimensions.
4) Apply a normal clustering approach after removing the names of the dashboards. Figure 3 shows the dendrogram output from hierarchical clustering with Euclidean distance and Average linkage. If we overlay the dashed green line, it produces 3 clusters with the dashboards in our example {A, F}, {G, B, C, D}, {E}.
5) Iterate on variety of clusters to reach at a set of balanced clusters that make business sense.
A caveat here is that a given metric could also be an element of various dashboards across multiple clusters. We will either document this occurrence to tell users or we could remove the metric based on business judgment from K-1 dashboards, where K is the full variety of clusters where the metric appears. Nonetheless, such a judgment-based elimination might be sub-optimal.
One other challenge with a standard clustering approach is that it might not group dashboards which are subsets of other dashboards in the identical cluster. For instance, Dashboard A is a subset of Dashboard E as might be seen in Figure 1 (i.e. metrics, user personas, filers and data sources in Dashboard A are also present in Dashboard E) but they’re grouped in several clusters (Figure 3). The concept behind capturing subsets is to eliminate them since an alternate (superset) dashboard is accessible that also exposes the identical metrics together with others to users. To mitigate this issue, we propose an alternate clustering algorithm to assist group together subsets.