Given the “agnostic” nature of our chosen dataset, I felt a certain level of freedom to explore what the process would be to generate a physicalized network graph versus pursuing a “normal” data analysis with a visual approach.  Here I started the process with a firm framework of the project’s materials (string, nails/screws/ mounting board), mode of output (network graph), and other sensing modalities (tactility and aurality). The data would be secondary to the process of making a physical artifact. Did I just intentionally open the “bias door” with my predetermined framework? More on that later…

The resulting physical artifact in this case would be a data physicalization “whose geometry or material properties encode data” (Jansen et al., 2015). What is it not? It would not be a scale model, architecture model or a figurative sculpture (Dragicevic et al., 2021). Data physicalization does align and sometimes intersect with other areas of research within the human computer interaction realm, like visualizations (of course) – which utilize computers to process “abstract” data into visual representations for easier relatability and understanding. Another area, tangible interaction, focuses on the interactive systems that hinge on physical artifacts as the vehicle to both represent and control “computational media” (2021). It is interesting to note that the visual facet of these other data representation modalities never fully divorce from their definition or explanation. Notably, “Design Principles for Visual Communications” by Agrawala et al., 2011 is referenced multiple times to frame the process of transforming abstract data to a human consumable format via “the three-stage methodology [that] involves first identifying physical characteristics to be incorporated into the design, followed by instantiation of those characteristics, then evaluation of the physical characteristics”(Drogemuller et al., 2019). The sense of sight is front and center in the world of data science and arguably in Digital Humanities as well.

But we have other senses and for some, a limitation on what others assume as the default sense.  Data physicalization focuses on “data-driven tasks, such as data exploration and data communication”(Dragicevic et al., 2021), which opens the door to increased pathways for abstract data constructs to connect to human understanding. Jansen et al. posit “all senses can participate in information gathering and they each have unique characteristics that can be leveraged by physicalizations”(Jansen et al., 2015). 

As the field has developed in this area, the term “data visceralization” entered into the lexicon thanks to Kelly Dobson in the early 2000s, and “later popularized by Luke Stark”(Wernimont, 2022). Defined as the “representations of information [that] rely on multiple senses including touch, smell, and even taste, [that] work together to stimulate our feelings as well as our thoughts”(Stark, 2014),  it has become more prevalent in the discussion and research. There is an interesting nuance that Stark additionally points out in that “visceral data has the potential to level out our reactions the opposite way [from a solely visual consumption of information]: as well as appreciating a problem or issue rationally, users prompted to engage viscerally will have a well-rounded sense of their own intellectual, emotional and physical stance on the matter at hand”(2104).

To arrive at the point of “visceralization” and/or “physicalization” in my own project, lengthy and laborious efforts were required to prepare the data. First I identified the key columns of data that would best suit an output of a network graph. I initially chose: “Unique Identifier”, Episode_ref ,track_name, date, castaway_ref , std_artist, and std_name from the “disk” dataset.   Partly by design, I chose those same columns during the labor intensive encoding error and “missing-ness” identification process Ava, Carol, Lubov, and I performed for the Map & Key. From that point I continued to work with the data in OpenRefine to combine like values through text facets, algorithmic clustering, to remove special characters, and perform visual checks in preparation for the physical output. Essentially I made decisions about what data would stay, be “fixed”, or disregarded.

Another decision point arose with the representation of the data as nodes or edges. I applied a Python script to create the nodes pair and the connecting edges with weight (= occurrences). Here is when the process deteriorated (or perhaps more kindly, iterated through on repeat). Multiple trial attempts of identifying the nodes and edge relations for the final output transpired. Initially I considered the entire “disk” dataset, but that was not a realistic choice as it was too large to transfer with the materials ( or time) I had available. I processed a version where I broke down the episodes by decades (compare 60s vs 00s, eg). From there I iterated through a couple more choices of how to splice the data: song choice as node with the artist as edge, then castaway as node and artist as edge, then the artist as nodes castaways (guests) as edges – their co-occurrence choice. I even attempted to pull a subset of the data with a randomizer (1 choice per year to process using the artists as the nodes and the episode date as the edge, decades). The final attempt, Trial 6, resulted in the selection of the top 8 artist choices (nodes) across all the years two of those top choices co-occurred(edges).

The resulting nodes and edges files were imported into Gephi, an open graph visualization software. I used Gephi to visualize the nodes and edges table to iron out preliminary layout and  run some statistics on degree and modularity. I played around with the layout (color, size, ranking) in preparation for the physicalization build.

Physicalization of the Absence

The resulting data subset I arrived at included only 4478 artist choices for the top 8, leaving 21,782 choices not “used”.  That’s quite a bit of data not used!  I was reminded about an article Cat Hicks published on Data Science by Design where they talk about what is left behind when processing data for an intended outcome, “Sometimes I like to call this ‘when we miss missingness,’ a quip to remind myself that we need to understand data not just in terms of datasets but in terms of the context around it. We need to ask what data was collected and, perhaps more importantly, what data was left out”(Hicks, 2022). Which is why I chose to bring them into the fold and highlight their presence through another tactile mode.

With pen and paper, I created a series of dots that each represented 2 choices each, total of 10,891. I would later affix these textural dot papers to the underneath of the physical network to symbolize that the data doesn’t cease to exist when choices are made to exclude, parse, and foreground visualized data.

Sonification of Labels /Materialization of Connections

The physical network graph assembly involved screwing in screws into the foam core board which served as the nodes that the edges would connect to via crochet thread in various colors representing the top 8 artist choices. The Python nodes and edge table was my instruction list for the wrapping which colored thread (edges) between 2 screws (nodes).  The weight (occurrences) was the amount of thread loops connecting the screws (one thread thickness = 2 occurrences). The nodes’ label was supplied by a miniature recording device with the recorded name of the top artist playing when a button was pressed.