Prediction model of pet rescue outcome based on LightGBM and Bayesian optimisation

Abstract

Due to owner abandoning, pets getting lost, and irresponsible breeding, the number of strays is rising steadily as more individuals opt to have pets. Stray animals are not only in a precarious position, but they also wreak havoc on the lives of humans. To reduce the number of stray animals and enhance the pet rescue mechanism, this paper develops a prediction model of pet rescue outcome based on LightGBM (Light Gradient Boosting Machine). The model selects data features using the Pearson correlation coefficient, employs SMOTE (Synthetic Minority Oversampling Technique) to balance the data, and then uses LightGBM modelling to predict the outcome of each pet, such as adoption, transfer, death, etc. Simultaneously, the Bayesian algorithm is used to optimise the model’s hyper-parameters in order to generate the final model. Comparing the prediction results of SVM, Random Forest, and BP neural network models with those of the model based on LightGBM, the experimental results on the real animal centre dataset reveal that the model based on LightGBM has a superior predict performance, and the Bayesian optimisation effect is evident. In addition, the paper assesses the importance ranking of model features to aid animals in animal shelters more effectively.