GameGlobs is a two-dimensional visualization of various clusterings of the games in our natural language processing model. A user selects how many clusters (groups of related games) she would like to see the 11,829 games partitioned into and is presented with such a clustering. Each cluster is drawn as a circle that can be clicked to display the games it contains, which are stylized as hyperlinks to their entries in GameNet. The clusterings themselves were derived by applying the classic k-means algorithm to the games’ LSA vectors. GameGlobs includes clusterings using several values for k (number of clusters) spanning between 2 and 2500 and utilizes two key visual cues: clusters with more games appear larger, and clusters are positioned semantically, such that clusters whose games are more similar are nearer one another. To achieve the latter effect, we used a technique called multidimensional scaling (MDS), which is a way of building low-dimensional visualizations of high-dimensional data. This technique is represented by a suite of algorithms; we submitted the LSA vectors of our cluster centroids to a variant called locally linear embedding (LLE) to derive their 2D coordinates. This is a collaboration with James Ryan at the Expressive Intelligence Studio.

GameGlobs is live here. For more information on the model and GameGlobs, please check out publications.