VaBank — Visual Analytics for Banking Transactions

To analyse and detect fraudulent patterns in banking transactions, most fraud analysts use spreadsheets which makes the overall process time-consuming and complex. In this project, we propose a visualization tool that aims to ease the analysis of banking transactions over time and the detection of the transactions’ topology and of suspicious behaviours. Our main contributions are: (i) a user-centred visual tool, developed with the aid of fraud experts; (ii) a method that characterises the transactions topology through a self-organising algorithm; (iii) the visual characterisation of transactions through complex glyphs; and (iv) a user study to assess the tool effectiveness.


The tool


Figure 1

Transaction History view and its components.


VaBank is divided into three views: the transaction history, the transactions topology, and the transactions relations. The first arranges all transactions by time and amount. The last two display the results of a SOM algorithm in a grid and through a force-directed graph, respectively. All views have access to a GUI panel (Fig.1, A).


In the Transaction History view, the main representation, which occupies more canvas space, is the the transaction matrix (Fig.1, B). It divides the space in different ranges of amounts on the y-axis and temporal values on the x-axis. The transactions, represented as glyphs according to their characteristics (Fig.2) are then distributed by the cells of the matrix, according to their date and amount. If more than one transaction with the same characteristics occur within the same cell, they are aggregated and its glyph grows in size.


Figure 2

Glyph elements that characterise each transaction (side A) and timeline bar composition and respective colour ranges (side B)


In the bottom and right sides of the matrix, histograms are drawn to show the total number of transactions per column and row, respectively (Fig.1, B and C). The histogram’s bars are coloured according to the number of transactions: the darker, the higher the number of transactions. By hovering each bar, the total number of transactions is shown. In the bottom right corner of the matrix area, we draw a small matrix of glyphs that represents the result of a SOM algorithm, concerning: amount, transaction type, and fraud (Fig.1, C). With this, we aim to enhance the understanding of typical/atypical transactions.


In the bottom, we placed an interactive timeline, so the user can select and visualise different periods of time in the data (Fig.1, D). The different time periods are defined by a hierarchical aggregation algorithm.


Transactions Topology


Figure 3

Projections of the SOM results for the same bank client through the matrix projection (left) and force-directed graph (right).


We applied a SOM algorithm to enhance the detection of atypical transactions, which can be related to fraudulent behaviours (Fig.3). The results of the SOM are visualised through a matrix or a force-directed projections. The goal of the first (Fig.3, left) is to represent the distribution of different types of transactions present in the data and extrapolate at a higher level the characteristics of the dataset. The goal of the second (Fig.3, right) is to express the relations among clustered data and emphasise the most typical transaction, allowing a more detailed analysis of the dataset.


Evaluation and Results


To evaluate the tool usefulness and effectiveness in the analysis of banking transactions, we performed a user testing with 5 fraud analysts—which were not present during the tool development.


The tests were performed as follows: (i) we introduced the transactions representation, the views of the tool, and its interaction mechanisms; (ii) we asked the analysts to perform 18 tasks concerning: transaction history (G1); interpretability of the glyphs (G2); SOM matrix (G3); and SOM graph (G4); (iii) the analysts analysed two clients in terms of behaviour and fraud; and (iv) the analyst gave feedback on the models concerning aesthetics, interpretability, aid in the analysis, and learning curve. The second and third part of the tests were timed and, in the end of each, the analysts were asked to rate the difficulty of each exercise and certainty of their answer—from 1 to 5.

The Figure bellow summarises the results concerning difficulty, certainty, accuracy, and duration for each group of tasks. Hereafter, we further analyse each group and discuss the results from the third part of the test and the analysts feedback.


Figure 4

Difficulty, Certainty, Accuracy and Time for the four groups of tasks.


Transaction History and Glyphs: The tasks related to the analysis of the Transaction History view (G1) and the glyphs (G2) were the ones which arouse more difficulty. Regarding the Transaction History, the analysts had more difficulties in interpreting the positioning of the glyphs in the grid and the histograms.


SOM Visualisation: The groups of tasks related to the SOM analysis took less time to perform (20 seconds, on average), had 100% of accuracy, and were the ones in which the analysts had more certainty in their answers and less difficulty in completing the tasks. Comparing both projections, the graph (G4) had a lower duration and the difficulty of completion was also considered low.


VaBank analysis: The second part of the test was concerned with the free exploration and analysis of the transactions of two clients. All analysts could interact properly with the tool. They stated that after the tasks completion they were more familiarised with the tool, and could use easily all functionalities.


Feedback: In the end of each test, the analysts rated each view in terms of aesthetics, interpretability, aid in the analysis, and learning curve. The Transaction History view, got the higher rate in terms of aesthetics and aid. Additionally, it was defined as the easier to learn and interpret. This enhance the fact that, although in the tasks it was the most difficult, after interaction it got easier to interpret. This view was well received by the analysts which defined it as a good auxiliary for their work. Concerning the graph and matrix views of the SOM, with the matrix view the analysts took more time to complete the tasks, and rate it with higher values of difficulty. However, the matrix grid was seen as a better aid to analyse the transaction patterns and was also defined has easier to learn.


  • C. Maçãs, E. Polisciuc, and P. Machado, “VaBank: Visual Analytics for Banking Transactions,” in 24th International Conference Information Visualisation, IV 2020, Melbourne, Australia, September 7-11, 2020, 2020, pp. 336-343.

  • C. Maçãs, E. Polisciuc, and P. Machado, “Visualization and Self-Organising Maps for the Characterisation of Bank Clients,” in Integrating Artificial Intelligence and Visualization for Visual Knowledge Discovery, B. Kovalerchuk, K. Nazemi, R. Andonie, N. Datia, and E. Banissi, Eds., Cham: Springer International Publishing, 2022, pp. 255-287.