ADAPTIVE MACHINE LEARNING IN DELAYED FEEDBACK DOMAINS BY SELECTIVE RELEARNING

摘要

We present a novel hybrid technique for improving the predictive performance of an online machine learning system: Combining advantages from both memory-based and concept-based procedures selective relearning tackles the problem of learning in gradually changing domains with delayed feedback. The idea is based on training and retraining the model only on the subsegment of the historical dataset which has been identified as the one most similar to the current conditions. We exemplify the effectiveness of our approach by evaluation in a well-known artificial dataset and show that selective relearning is rather insensitive to noise. Additionally, we present preliminary experimental results for a complex synthetic dataset resembling an online diagnostic system for the tile manufacturing industry and show that the procedure for selecting the best segment yields favorable training results in terms of the mean-squared error of the predictions.