Data Analysis: Spotfire vs Tableau

A few weeks ago I compared the abilities of Spotfire and Tableau for generating beginner-level data visualisations. Overall I found Tableau easier to handle and was happier with the amount of control I had over the aesthetics of my visualisations. However, writing that review led me to wonder how these two pieces of software compare when it comes to a more detailed, quantitative analysis of a dataset. I was spurred on to extend my initial research and present my findings below.

So, how do Spotfire and Tableau compare in terms of their abilities as data analysis tools..?

Time-series: Moving Averages and Forecasting

For this review I'm using the same test dataset as my visualisation article: Temperature observations from cities around the world since 1743 (open source data provided by Berkeley Earth). The dashboards above show the variation in average temperatures recorded in selected countries through time. For the upper visualisations I generated 10 year moving averages for each country which are represented by the semi-transparent lines through each time series. The moving averages were fairly easy to generate in both software packages, although the function wasn't immediately obvious to find in either instance.

The lower visualisations show 20 year forecasts generated by each software suite. Clearly different algorithms are at play here! Tableau could not detect any cyclic/seasonal element to the input data so reverted to a linear regression line surrounded by a 95 % confidence interval. A report describing the forecasts and analysing their quality was also generated, featuring several metrics (such as RMSE) as well as qualitative QC descriptors ("good", "ok" or "poor").

Whilst Tableau automatically estimated all forecasting parameters, Spotfire needed a bit more supervision. Clearly some of my forecasts need a bit of a tweak! Some users might prefer Tableau's simple (yet somewhat "black box") approach but those with a bit more forecasting background might favour Spotfire's more hands-on analysis. Note that although Spotfire hasn't used a linear regression approach to its forecasting, it does have a whole module devoted to detailed linear regression.

Box-plots

I love both of these visualisations! The data you're looking at is the difference between maximum and minimum temperatures recorded in every city over the period from 1860 to 2013. Spotfire's visualisation allows you to display the distribution of actual data alongside the semi-transparent box-plots - a feature which I think is fantastic. The table underneath can be populated by as many or as few statistics as you choose, making the overall visualisation highly informative. Don't look too closely though or you'll notice that I haven't succeeded in pivoting the data quite as I intended!

In Tableau you can overlay a box-plot on virtually anything, even if it's not meaningful. I'm a particular fan of the density plot shown above (correctly pivoted!), which gives a similar impression to the distributions generated by Spotfire. In both visualisations you can interpret that for all decades, most cities experienced a low magnitude range of temperatures (maybe up to 10 C), but there are also quite a few cities which have seen temperature swings of 25 C or greater. For those that are interested, it seems to be high latitude (often Russian) cities which experience the largest temperature variations.

Cluster Analysis: Spotfire

Spotfire can carry out two types of cluster analysis: K-means and Hierarchical. I opted for K-means clustering which requires the user to input a line chart. I don't know why Spotfire enforces this limitation, but to satisfy the criteria I chose the time-series data shown earlier. Once the cluster analysis is complete, Spotfire partitions the input line chart so you can see which lines were assigned to which clusters (upper image). This was very useful for reviewing what the clusters actually represent. The output clusters can then be displayed on other visualisations to allow further interrogation (lower images). A lot of my data could not be assigned (I suspect largely due to user inexperience), but the clusters I was able to generate looked sensible on these further displays.

Cluster Analysis: Tableau

Tableau only performs K-means cluster analysis but has no specifications for how the data needs to be input. For this reason my Tableau cluster analysis shows something a bit more interesting than the Spotfire example above. I created the upper plot showing latitude of cities vs their observed temperature range (i.e. maximum - minimum) and then started the cluster analysis. The inputs were: Range of temperature, latitude and a Boolean variable for north/south hemisphere, with cities as the level of detail. The clusters appeared immediately on my visualisation and the input variables could be interactively tweaked until I was happy with the result. A summary report was also generated which included details on each cluster's attributes and analysis of variance. Plotting the clusters on a world map revealed a very sensible looking distribution which reassured me that the clustering had made use of all input variables.

The Verdict

When creating these examples it was clear that Spotfire demands greater knowledge from the user, both in terms of how to control the software as well as having a deeper understanding of how various data analysis techniques work. Whilst Tableau can "do" data analysis, Spotfire tends to offer several different techniques as well as a host of parameters to control. That said, Tableau's analysis processes automatically generated descriptive/QC reports which would be highly valuable metadata for any real-life project.

So which software wins? Well it depends...

I have a strong background in data analysis but I'm no expert - geophysics is my main discipline. I felt most comfortable controlling Tableau, however, there were elements which were quite "black box" so a data analysis specialist might get more out of Spotfire. But they would need to spend more time learning how to use the software!