Correcting Instrumental Variation and Time-Varying Drift Using Parallel and Serial Multitask Learning

Abstract

When instruments and sensor systems are used to measure signals, the posterior distribution of test samples often drifts from that of the training ones, which invalidates the initially trained classification or regression models. This may be caused by instrumental variation, sensor aging, and environmental change. We introduce transfer-sample-based multitask learning (TMTL) to address this problem, with a special focus on applications in machine olfaction. Data collected with each device or in each time period define a domain. Transfer samples are the same group of samples measured in every domain. They are used by our method to share knowledge across domains. Two paradigms, parallel and serial transfer, are designed to deal with different types of drift. A dynamic model strategy is proposed to predict samples with known acquisition time. Experiments on three real-world data sets confirm the efficacy of the proposed methods. They achieve good accuracy compared with traditional feature-level drift correction algorithms and typical labeled-sample-based MTL methods, with few transfer samples needed. TMTL is a practical algorithm framework which can greatly enhance the robustness of sensor systems with complex drift.