2D vocal expressive space of cats following a short isolation period

Abstract

A sample of cats was recorded following a mild stress experiment. After a brief isolation period, the cats were reunited with their caretaker, and their recordings were analyzed. Principal component analysis (PCA) and t-distributed stochastic neighbour embedding (t-SNE) were used to reduce the high dimensionality of their vocalizations to two and analyzed using a battery of three different clustering algorithms: Affinity Propagation, MeanShift and HDBSCAN. Every clustering algorithm showed similar results; however, to identify overrepresented clusters, the sizes of each cluster set were tested for uniformity with a Kolmogorov-Smirnov test. This test revealed that the cluster sizes analyzed by HDBSCAN were not uniform and some values stood out more than others. A z-score test was conducted to determine which cluster was the most overrepresented and this was followed by a Chi-Square test to find the rest of overrepresented clusters. Visual inspection of the spectrograms shows that a change in frequency follows a predetermined path in the t-SNE plot, highlighting a gradient and a polarisation in their vocalizations. One extreme exhibited visible harmonics, which progressively diminished as it approached the opposite region, with vocalizations that work over broader, noisier, bandwidth with sparser harmonics (purrs). The number of vocalizations and their length were tested with a Kruskal-Wallis test and a one-way ANOVA, respectively, both of which found statistical significance. This analysis was followed by post-hoc tests, which verified significant differences among the tested cats in the number of vocalizations and their duration.