Classification of Data Streams Applied to Insect Recognition: Initial Results

Applications such as intelligent sensors should be able to collect information about the environment and make decisions based on input data. An example is a low-cost sensor able to detect and classify species of insects using a simple laser and machine learning techniques. This sensor is an important step towards the development of intelligent traps able to attract and selectively capture insect species of interest such as disease vectors or agricultural pests, without affecting non-harmful species. The data gathered by the sensor constitutes a data stream with non-stationary characteristics, since the insects´ metabolisms are influenced by environmental conditions (such as temperature, humidity and atmospheric pressure), circadian rhythm and age. Algorithms that classify data streams often assume that once a prediction is made, the actual labels are provided to assist in updating the classifier. In the case of intelligent sensors, these labels are rarely available. The objective of this paper is to evaluate methods that adapt concept drifts by regularly updating the classification models applied to insect recognition in a data stream. We show in our initial results that the philosophy of inserting and removing examples from the training set are of essential importance. We also show that a simple criterion to insert examples with high classification confidence can significantly improve the accuracy.