Addressing bias in machine learning algorithms: A pilot study on emotion recognition for intelligent systems

Recently, there has been an explosion of cloud-based services that enable developers to include a spectrum of recognition services, such as emotion recognition, in their applications. The recognition of emotions is a challenging problem, and research has been done on building classifiers to recognize emotion in the open world. Often, learned emotion models are trained on data sets that may not sufficiently represent a target population of interest. For example, many of these on-line services have focused on training and testing using a majority representation of adults and thus are tuned to the dynamics of mature faces. For applications designed to serve an older or younger age demographic, using the outputs from these pre-defined models may result in lower performance rates than when using a specialized classifier. Similar challenges with biases in performance arise in other situations where datasets in these large-scale on-line services have a non-representative ratio of the desired class of interest. We consider the challenge of providing application developers with the power to utilize pre-constructed cloud-based services in their applications while still ensuring satisfactory performance for their unique workload of cases. We focus on biases in emotion recognition as a representative scenario to evaluate an approach to improving recognition rates when an on-line pre-trained classifier is used for recognition of a class that may have a minority representation in the training set. We discuss a hierarchical classification approach to address this challenge and show that the average recognition rate associated with the most difficult emotion for the minority class increases by 41.5% and the overall recognition rate for all classes increases by 17.3% when using this approach.