Analysing Visual Clusterings
Critique
Following Arnold and Tilton’s argumentation, distant viewing is an attempt to formalize the act of interpreting visual material, and counteract interpretative bias. They outline how the “process of coding [individual] images in this way is both destructive and interpretive” [@arnoldDistantViewingAnalyzing2019, p. 5], since the description and labeling of an image is intimately tied to the researcher. Distant viewing should be able to circumvent this, by describing and labeling through a formalized system. It could be argued, that this merely shifts the problematics to the structural level. The important aspect is that their observation is based on working with metadata resulting from object classification and face recognition. My research on the other hand works with the clustering of visual material based on image similarity, which results in a new type of image. The newly created visualisation is a meta-image [@blaschkeMetabilderDigitaleBildkulturen2021] or operational image [@parikkaOperationalImagesVisual2023].
Volker Pantenburg outlines operational images as “visualizations of data” and describes them as a “working image” or an image that “performs work” [@pantenburgWorkingImagesHarun2016]. It is then an image with functional purposes rather than being made for aesthetic appreciation. These images are created to assist in tasks or operations, frequently involving machine automation, and are not meant to be representations of the real world for human viewing. The term was introduced by German filmmaker Harun Farocki in 2000, in his audiovisual installation Eye/Machine, which showcased such images used by militaries, including footage from laser-guided missiles during the Gulf War.
Despite being of machine origin, the operational image produced through the clustering of visual material reintroduced the problem of interpretation. Having went through a formalized process of description and labeling, the results are yet another image that needs to be interpreted. Distant viewing an image clustering is a specific case of the problematic that Johanna Drucker outlined in regards to producing knowledge through data visualisations. The insights, the knowledge and the data are not inherent in the visualisation, but are actively produced through the researcher [@druckerHumanitiesApproachesGraphical]. She argues to think of data rather in terms of something that we capture, as capta. Image clustering visualisations add to the problematic, by being made in a seemingly intelligent manner. The researcher sets up the process and the results can look pleasing, ordered and conciously designed. The impression can arise, that these visualisation must hold some knowledge to be uncovered.
In fact and reality, the researcher must now interpret again. After choosing from a set of variables for the process, the results can vary. The final visualisation should not be to tight and not be to loose, but strike a balance towards being interpretable. After this initial step, the researcher looks at the visualisation, makes notes, reconfigures what is on screen and takes further step to capture insights. I found a lack of actual descriptions on how this interpretative work should be done in the literature. More often then not, scientific paper applying image clusterings display sentences such as “By closely observing the groupings, it was possible to establish an initial hypothesis, indicating that there seems to be” [@teixeiraMedalsLikesMethodology2024, p. 7], and completely omitting this crucial part of the research process in their descriptions.
To fill into this gap, I continue to outline my own process of practical steps that helped me in analysing the image clusterings. The process involved multiple sessions of intense engagement with the resulting visualisation, with the manual and automatic disovered hotspots, different configurations of the visual and metadata, perspectives, and hiding or focusing on specific aspects. Parts of this process can be based on a research protocoll, but others need to be delegated to trusting in serendipity through experimenting and prolonged engagement with the visualisation.
Practical Guidance
- Check global form for general clusters, transitional lines, holes, islands, peninsulas and bridges in the layout
- Investigate oddities in global and local structures for aspects that afford visual complexity
- Manually build and analyse clusters through the visual material’s formal similarities
- Explore different settings for automatic cluster-detection and analyse what those bring forth in terms of global and local structures, as well on what basis the cluster was built
- Label and describe the clusters in a structured way
- Reconfigure the layout with the help of metadata and iterate
- Check in how first findings correlate with research questions at hand and loop back to step one. Repeat until finding saturation occurs.