Sheep's coping style can be identified by unsupervised machine learning from unlabeled data


Çakmakçı C.

Behavioural Processes, no.194, pp.1-10, 2022 (SCI-Expanded)

  • Publication Type: Article / Article
  • Publication Date: 2022
  • Doi Number: 10.1016/j.beproc.2021.104559
  • Journal Name: Behavioural Processes
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Social Sciences Citation Index (SSCI), Scopus, Academic Search Premier, PASCAL, Animal Behavior Abstracts, BIOSIS, CAB Abstracts, EMBASE, Geobase, MEDLINE, Psycinfo, Veterinary Science Database
  • Page Numbers: pp.1-10
  • Van Yüzüncü Yıl University Affiliated: Yes

Abstract

The objective of this study was to define coping style of sheep by using unsupervised machine learning approaches. A total of 105 Norduz sheep (age 3–5 years) were subjected to a 5-minute arena test. Agglomerative Hierarchical Clustering (HCA) was performed on scores of selected principal components retained from Principal Components Analysis (PCA) on arena behaviors to identify sheep coping style. Initially, the variables retained for the PCA were determined with Bartlett’s test for sphericity and Kaiser-Meyer-Olkin (KMO) measure of sample adequacy. Seven behavioral variables with KMO values greater than 0.5 were used for final PCA: the average distance to group sheep (DTG), the average distance to stimulus (DTS), the duration of locomotion (LOC), the total number of zone boundaries crossed during the test (CRS), the total number of times that tested sheep sniffed stimulus (NSS), latency to the first sniff the stimulus (LSS), and subjective scores (SCR) scored by an observer on a scale from 1 to 5 (1: extremely calm, 5: extremely restless). The first two components, which were the only ones with an eigenvalue greater than one, accounted for 70.32% of the total variation and were used for clustering analysis. Clustering tendency showed that the scores for the first two components were suitable for clustering (Hopkins' H = 0.852). Several cluster validity indexes were used to obtain aggregated results to determine the most appropriate clustering method and number of clusters. Five different clustering methods: k-means and hierarchical clustering with Ward, average, single and complete linkage were compared. Bootstrap resampling was used to evaluate the stability of a given cluster using the Jaccard coefficient. The clustering method and number of clusters corresponding to the highest rank aggregation score from the bootstrap resampling indicate that the hierarchical clustering method with average linkage and 5 clusters is the most suggested clustering method. However, Ward's algorithm identified the strongest clustering structure for hierarchical clustering, as it had the highest agglomerative coefficient value (0.98). When both Jaccard and aggregation scores are considered together, Ward’s method with 3 clusters was selected as the most appropriate method. Sheep were classified into three coping styles (CS) based on HCA results as reactive (Cluster 1, n = 71), intermediate (Cluster 2, n = 22) or proactive (Cluster 3, n = 12). Coping style had significant effect on behavioral variables, DTG, DTS, LOC, CRS and NSS (P < 0.05). The individuals that have proactive coping style had the highest mean values for the variables DTG, DTS and LOC and SCR (P < 0.0001). This indicates that proactive sheep are more active then reactive sheep. The CRS, LOC and NSS mean values were higher for intermediate sheep compared to reactive sheep (P < 0.05). The NSS values were higher for intermediate sheep compare to proactive sheep (P < 0.0001). The findings of the current study show that distinct coping styles in sheep may be identified based on behaviors recorded in an arena test. The findings also revealed that sheep's coping style can be objectively identified by unsupervised machine learning from unlabeled behavioral data.