Transfer Kernel Extreme Learning Machine Based on Bidirectional Cross Domain Approximation

Kernel extreme learning machine (KELM) not only inherits the high learning efficiency of the traditional extreme learning machine (ELM), but also mitigates the instability caused by random initialization of input layer parameters, thus establishing its success as a machine learning model. However, KELM will be degraded when there is an insufficient amount of labeled training data available. In this study, we propose a novel approach called transfer kernel extreme learning machine based on bidirectional cross domain approximation (TKELM-BCDA) to address this issue within the context of unsupervised transfer learning (UTL). Firstly, TKELM-BCDA decomposes cross domain mean approximation (CDMA) into two single directional cross domain mean approximations (SDCDMA), aiming to rectify the distribution discrepancy between source and target domains in UTL. Subsequently, these SDCDMAs are integrated with KELM to construct two transfer KELMs respectively. Furthermore, we introduce two output weight approximation terms into these transfer KELMs - referred to as source transfer KELM and target transfer KELM - in order to enhance knowledge transfer across domains. Finally, by combining the predictions from both transfer KELMs, our model significantly boosts its robustness. To validate the effectiveness of our proposed method, we conducted classification experiments on four publicly available datasets. An analysis of the experimental results reveals that TKELM-BCDA achieved an average classification accuracy improvement of 14.14% compared to the KELM baseline method. These comparative results clearly demonstrate that our proposed algorithm offers satisfactory classification accuracy and strong robustness.