Alex Zolot's weblog: From Data to Knowledge

Objective

In this work we have deal with the problem of visualization for Client service events to optimize work of client service.

The raw data have form

Case	DateTime	Event	Type
Case1	T1	E1	Type1
Case1	T2	E3	Type1
Case1	T3	E7	Type1
Case2	T4	E3	Type1
……

Table 1. Raw data

Service events could be case; Creation – Received – First_Contact_SW – First_Contact_HW – Request – Pending – Closed.

The same type of visualization can be done for cross-sell and up-sell analysis, where events could be purchases of specific products by a customer, e.g. we could have deal with opening a bank sequence of accounts; then instead of Case we have CustomerID, and Event can have values of Open_Checking_Acct, Open_Saving_Acct, Open_Loan, Close_Checing_Acct and so on;

The Type could be BranchID or Group of Clients and can be used for the Classification of cases.

To analyze the table we sort it by Case,DateTime and create variables Previous Event (PrEv) and Time between Events T:

Case	DateTime	Event	Type	Time	PrEv
Case1	T1	E1	Type1	0	0
Case1	T2	E3	Type1	T2-T1	E1
Case1	T3	E7	Type1	T3-T2	E3
Case2	T4	E3	Type1	0	0
……

Table 2.

To analyze the quality of service we aggregate the data in Table 2 calculating count and average through Case and obtain two tables: Frequency (or Count) and Time that is average of Time in Table 1 :

Event	PrEv	Freq	Time	Type
E1	E1	Fr11	T11	Type1
E1	E2	Fr12	T12	Type1
…				Type1
E2	E1	Fr21	T21	Type1
…

This table we can transform to two “wide” tables:

Freq

PrEv \ Ev	E1	E2	E3	…
0	Fr01	Fr02	Fr03
E1	Fr11	Fr12	Fr13
E2
…

and

Time

PrEv \ Ev	E1	E2	E3	…
0	T01	T02	T03
E1	T11	T12	T13
E2
…

The simplest way of visualizing these two tables is to put in the cells of the table rectangulars (bars) with width proportional T and length proportional Fr:

PrEv \ Ev	E1	E2	E3	…
0	▌	████	██
E1	███	█	█
E2	█████	███	▌
…

In this table the rows show frequency and average time of transactions following events PrEv and the columns show transactions that led to events Ev.

One disadvantage of this method is that each event is presented in the table twice: in raw header as “From” and in a column header as “To”.

To visualize this table without doubling the events, we present events as circles or other figures (e.g. “houses”) with area proportional frequency of the events and represent frequency F₁₂ and Time T₁₂ as arrow (or bar or petal) from Ev₁ to Ev₂ with width proportional to F₁₂ and length proportional T₁₂, color of the arrow is the same as the color of circle Ev₂:

Fig.1. Three variants of visual representation: arrows, bars and petals

We can choose positions of the circles arbitrarily; the simplest case is to put it on a big circle where all event circles “can see” each other:

Fig.2. “Flower bed” chart

During data aggregation from Table 1 we could use the same type of chart but length of rectangular could be proportional mean(1/T) or Scale(T) =exp(mean(ln(T))). The latter makes sense because the distribution of time between events could be Weibull rather than normal.

We prefer to plot length of bars (or arrows) proportional mean (T) because sometimes lost for servicing company is proportional to time of service multiplied number of cases; in such situation areas of rectangulars (bars or arrows) are proportional to $$$ amount of loss related to these transactions, so just a short glance at the chart shows which process creates the majority of issues for the company.

We named the chart “flower bed” chart. Another alternative could be to use standard techniques for weighted multidigraph visualization [1], but we think our “flower bed” chart is easier for interpretation and visual perception.

We have to create the “flower bed” chart (Fig. 2) for each Type of case to compare quality of service between different Types.

References

1. www.graphviz.org/Gallery.php