Paper Type
short
Description
In dynamic business environments, the underlying true data pattern changes rapidly. Machine learning models built upon historical data may not be responsive to the changes. A simple solution is to re-train a machine learning model using the re-collected current data. However, current data are often scarce. Therefore, it would be optimal to adapt the machine learning model built on historical data to the current period. In this study, we propose a two-step transfer learning method for enhancing machine learning in dynamic data environments. Our insight is that, by comparing current data and historical data, we gain information on the change of data environments, which guides the training of machine learning using historical and current data sets simultaneously. In this research-in-progress, we evaluate our method and an existing state-of-art algorithm in the earnings prediction tasks. Preliminary results show the effectiveness of transfer learning in dynamic business environments.
Recommended Citation
Peng, Jiaxu, "Transfer Learning in Dynamic Business Environments: An Application in Earnings Forecast for Public Firms" (2019). ICIS 2019 Proceedings. 6.
https://aisel.aisnet.org/icis2019/data_science/data_science/6
Transfer Learning in Dynamic Business Environments: An Application in Earnings Forecast for Public Firms
In dynamic business environments, the underlying true data pattern changes rapidly. Machine learning models built upon historical data may not be responsive to the changes. A simple solution is to re-train a machine learning model using the re-collected current data. However, current data are often scarce. Therefore, it would be optimal to adapt the machine learning model built on historical data to the current period. In this study, we propose a two-step transfer learning method for enhancing machine learning in dynamic data environments. Our insight is that, by comparing current data and historical data, we gain information on the change of data environments, which guides the training of machine learning using historical and current data sets simultaneously. In this research-in-progress, we evaluate our method and an existing state-of-art algorithm in the earnings prediction tasks. Preliminary results show the effectiveness of transfer learning in dynamic business environments.