Graph-based dynamic ensemble pruning for facial expression recognition

Ensemble learning is an effective method to enhance the recognition accuracy of facial expressions. The performance of ensemble learning can be affected by many factors, such as the accuracy of the classifier pool’s component members and the diversity of classifier pool. Therefore, choosing the component members of ensemble learning reasonably can be helpful to maintain or enhance the recognition rate of facial expressions. In this paper, we propose a novel dynamic ensemble pruning method called graph-based dynamic ensemble pruning (GDEP) and apply it to the field of facial expression recognition. The GDEP’s main intension is to solve the problem that in the dynamic ensemble pruning methods, the classifier selection process is heavily sensitive to the membership in test sample’s neighborhood. Like all other dynamic ensemble pruning methods, GDEP can be divided into three steps: 1) Construct the neighborhood; 2) Evaluate the classifiers’ performance; 3) Form the selected classifier subset according to the classifiers’ capacity for recognizing a specific test image. And in order to achieve the GDEP’s intension, in the first step, this paper chooses neighborhood members more carefully by taking use of the statistics of classifiers’ behavior to characterize the intensity and similarity of emotions in data samples, and using the geodesic distance to calculate the data samples’ similarity. In the second step, GDEP builds the must-link and cannot-link graphs in the neighborhood to measure the classifiers’ performance and reduce the impact of inappropriate samples in the neighborhood. The experiments on the Fer2013, JAFFE and CK+ databases show the effectiveness of GDEP and demonstrate that it can compete with many state-of-art methods.