The SOMFlow workflow supports domain experts (e.g., linguistic researchers studying speech intonation) in interactively exploring large collections of time series data through iterative SOM-based cluster analysis, guided by quality and interestingness measures, with full analytic provenance.
The process begins with input and preprocessing, where a collection of time series data items (e.g., pitch contours of recorded utterances) together with associated categorical metadata attributes are normalised and interpolated to produce equal-length feature vectors suitable for SOM training. A SOM training and quality computation phase then produces a set of cells (neurons), each containing a prototype vector and assigned data items, along with quality measures (quantization error, density, u-matrix distances) and interestingness scores (Simpson's diversity index for metadata attributes).
The trained SOM cells are arranged in their 2D grid topology and rendered as a rich composite visualisation comprising cell prototype line charts, switchable overlays (density, u-matrix, quantization error, meta-colouring), and a ranked sidebar of interestingness thumbnails with partitioning recommendation cues.
The expert enters an iterative SOM refinement loop, inspecting the visualisation to identify quality issues such as high quantization error, topology distortions, or confounding effects from inappropriate preprocessing (e.g., unnormalized pitch height causing gender-based separation). If the SOM is inadequate, the expert adapts the configuration—changing normalization strategy, distance metric, temporal weighting, or SOM parameters—and the SOM is retrained. This cycle repeats until quality is satisfactory.
Once an adequate SOM is achieved, the expert proceeds to iterative partitioning and exploration. Guided by recommendations, the analyst selects a partitioning strategy: (a) retrain imprecise cells as a detailed child SOM, (b) split data by an interesting metadata attribute, or (c) split by meta-clustering labels. A child SOM is trained on the extracted subset, visualised, and added to the analytic provenance flow graph. This cycle builds a progressively deeper flow graph until sufficient understanding is achieved.
Finally, in a reflection and knowledge synthesis step, the analyst uses the accumulated flow graph with cross-graph exploration techniques (2D colourmap linking, brushing, global meta-colouring, annotations) to compare cluster structures across branches and levels, synthesising domain knowledge: confirmed or rejected hypotheses, identified data quality issues, and discovered relations between time series patterns and metadata properties.