The workflow supports financial analysts in performing credit card expenditure segmentation — grouping transactions based on customer demographics and financial metrics to identify spending behaviour profiles. The distinguishing characteristic is that all computational methods (PCA, clustering) operate progressively: they produce approximate results within one second and iteratively refine them as more data is consumed, enabling uninterrupted human-computer dialogue. Analysts can renounce an ongoing computation at any time and redirect the analysis with a new feature selection without waiting.
The input consists of a high-dimensional dataset of credit card transactions (300,000+ transactions from 5,000+ customers) with dozens of attributes: location, amount, demographic information (job type, marital status, education, income, age), and financial metrics (mean transfer, deposits, risk score, response score, number of credit cards, entropy of transfers). The analyst's goal is to locate transaction segments — customer groups with similar spending behaviours — suitable for targeted marketing or risk management.
The analyst selects a subset of features from the difference view (e.g., EFT features, mean transfer, acceptance rate) to initiate progressive computations. Two computational modules run simultaneously: online PCA (incremental SVD updating) produces progressively improving 2D projections, and online clustering (mini-batch k-means) produces progressively refining cluster assignments. Both guarantee a first response within one second and improve with each subsequent batch via adaptive sampling. Visualisations animate between successive approximations at 10 Hz, and a progress bar indicates data consumption percentage.
As computations progress, the analyst inspects the evolving visualisations. From the clustering view, they identify automatically formed subsegments (AFS — automatic feature-based subsegments) by selecting a cluster's small multiple. From the PCA view, they identify user-defined subsegments (UDS) by lasso-selecting visually separable groups. At each moment, the analyst assesses whether the progressive results have stabilised sufficiently to act upon, or whether the current feature set is unproductive and should be abandoned (renounce). During the case study, analysts renounced 13 out of 14 feature combinations before full computation, typically based on visual stabilisation of the views rather than progress percentage alone.
The analyst composes a consolidated segment by combining subsegments (from both AFS and UDS) using boolean operations (union, intersection). They further fine-tune the composite segment by brushing on auxiliary views — for example, selecting transactions near the airport on the map view, or excluding small transactions via histogram brushing. The difference view continuously characterises the current segment against the rest of the data, showing which attributes distinguish the segment (using Cohen's D). This composition-and-refinement cycle may involve multiple feature selections and progressive computations until a coherent, well-characterised segment is identified.
The analyst synthesises findings: identified customer segments with their defining characteristics (e.g., customers with low EFT entropy, low credit card limits, high campaign acceptance rates, and low mobile banking usage — good candidates for further credit card offers), tested and confirmed or rejected hypotheses about segment compositions, and documented the segmentation rationale for decision-making.