Machine Learning Insights from C-BARQ Data to Study Canine Personalities

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Feb 7 2024

In a paper published in the journal Scientific Reports, researchers applied machine learning (ML) techniques to classify canine personality types using data from the canine behavioral assessment and research questionnaire (C-BARQ) project. Initially, an unsupervised learning approach was employed to cluster and label the dataset, revealing five distinct personality categories.

Researchers labeled these as "excitable/hyperattached," "anxious/fearful," "aloof/predatory," "reactive/assertive," and "calm/agreeable" personalities. Subsequently, researchers tested four ML models to predict personality traits, and the decision tree (DT) model achieved the highest accuracy. This artificial intelligence (AI)-based methodology showed promise for improving dog selection and training for various roles.

Related Work

Past work in psychology has focused on understanding personality, which encompasses consistent patterns of thinking, feeling, and behaving shaped by genetics and environment. Canine personality, crucial for human-dog relationships, influences surrender rates and welfare. With millions of dog bites annually in the United States (US), personality is also a public health concern. In specialized roles, such as working dogs, personality determines suitability. Researchers have developed tools like the C-BARQ questionnaire for canine personality assessment. Although researchers have utilized AI and ML for human personality prediction, they have yet to apply such methods to dogs.

Canine Personality Analysis Methodology

Researchers sourced the data for this research from the C-BARQ database at the University of Pennsylvania School of Veterinary Medicine, which provides standardized evaluations of canine temperament and behavior. Before analysis, data cleaning removed samples with missing values, resulting in 7807 complete samples. Researchers performed data encoding to convert string-type values to numerical values for ML model training. Among the 157 attributes, they identified 133 as numerical and 24 as non-numerical, requiring encoding. They focused primarily on the 100 scored behavioral items from the C-BARQ dataset for clustering to identify dog personality traits.

Researchers conducted clustering using the k-means algorithm, an unsupervised learning approach to group similar data points (dogs) into clusters. The algorithm iteratively optimizes centroid positions to minimize an objective function, with the number of clusters (K) being crucial. The Elbow method was employed to determine the optimal K value visually, indicating that 5 clusters were most suitable for this dataset. This process ensured effective clustering to identify distinct personality types in dogs.

Four ML models—support vector machine (SVM), k-nearest neighbor (KNN), Naïve Bayes, and DT—were implemented to predict dog personality traits. Researchers evaluated hyperparameter tuning optimized model performance and the models using a five-fold cross-validation method. The DT model emerged as the most efficient classifier for predicting personality types. Additionally, researchers conducted a feature importance analysis to identify influential behavioral attributes that define each cluster, helping in the descriptive labeling of personality types based on the top 20 most important attributes. Mean values of these attributes in each cluster were compared to others, informing behavioral descriptors for each personality grouping.

Classifier Performance Evaluation

The results of the clustering model utilizing the k-means algorithm to identify dog personality types illustrate a t-distributed stochastic neighbor embedding (TSNE) plot showcasing the distribution of samples across different clusters. These clusters, numbered from zero to four, were derived from the behavioral data obtained from the C-BARQ database. Each sample was assigned a cluster label, facilitating further analysis.

Researchers conducted a feature importance analysis for each cluster. These analyses aimed to identify the behavioral attributes defining each personality cluster. The top 20 most important attributes were used for descriptive labeling of personality types, contributing to a comprehensive understanding of each cluster's behavioral patterns.

The clustering results identified five distinct personality groupings, each characterized by specific behavioral traits. These personality types, labeled as "excitable/hyperattached," "anxious/fearful," "aloof/predatory," "reactive/assertive," and "calm/agreeable," were delineated based on the predominant behavioral features observed within each cluster. Researchers generated descriptions of each personality type by comparing mean attribute values within clusters and across the dataset, offering nuanced insights into canine temperament and behavior.

Furthermore, the performance of four ML classifiers—DT, Naïve Bayes, KNN, and SVM—was evaluated for predicting dog personality types. The DT model emerged as the most efficient classifier, demonstrating superior performance compared to the other models. Precision, recall, F1 score, and accuracy metrics were computed for each classifier, providing a comprehensive assessment of their predictive capabilities.

The study also addressed the limitations and considerations inherent in the analysis, such as the need to specify the number of clusters a priori in the K-Means algorithm and the potential impact of outliers on clustering results. Additionally, researchers discussed the performance of the Naïve Bayes model, highlighting its limitations in handling correlated features and nonlinear relationships compared to other classifiers. Finally, they proposed future research directions, emphasizing the need to validate these findings against objective canine behavior and personality measures in real-world contexts.

Conclusion

To sum up, AI and ML techniques, commonly used in predicting human personality types, have been adapted to analyze and predict canine personality traits using behavioral data from the C-BARQ. Researchers identified five distinct personality clusters and labeled them accordingly by applying unsupervised learning and utilizing the K-Means algorithm.

The DT model exhibited the highest accuracy at 99% in predicting personality types. These methods offer promising applications in the selection and training of dogs, but further research is required to validate their effectiveness across different canine populations.