Worked through three 8/16/32 clustering attempts and labeled the clusters. To that end I made a little script that takes some random samples from each cluster. Then I manually checked the samples and tried to summarize some formal and semantic aspects. Usually the formal clustering dominated.
Re-released the cleaned dataset under new name, “Video Game History Dataset” (VHS-D). Run into problems with git lfs, and then into problems with releasing on Zenodo. This gave me the opportunity to release a cleaner state of the dataset.
Distant viewing image clusters as generous interfaces.
Researching more literature to backbone the image clustering paper.
2024-08-20
Went again through the 8 and 32 clusters and described and labeled them in a more structural approach. Went once through the 128 cluster and made a quick summary of each, and picked those that are of interest. The overall trajectory is from roughly formal clusterings into clear formal elements and semantic clusterings. I also picked a few interesting clusters from the 1228 set to use in the paper.
isometric 6, 15, 92
horrorcore 12
boxing 23
nsfw 26, 102, 119
I like this approach of generating subsamples and analyze those in different configurations. Maybe that could find it’s place in the paper?
Some clusters are mirroring genre through formal aspects, eg graphic adventures, flight simulators or team sports games. Some others by semanto-formal aspects, eg boxing (23), where perspective can change from 3/4 to POV, but the elements of boxing gloves and ring-ropes stay. Some formal aspects transverse genre, such as street settings (42, 117, 122)