A Framework for Automated Feature Extraction with Continuous Feedback | Original Article

ABSTRACT:

The emphasis of machine learning exploration has mainly been on the learning various algorithms. This may be because of confined amount of data available. Over the period, the technology became more advanced and has created the opportunity to considerably more data. With the increase of stream data, it has become clean that the representation of such data, which is the input for any learning algorithm, can have a sizable impact at the performance of algorithms. The crucial thing these days is that data is not static rather it is continuously changing. However, this affects the already trained deployed models as new data trends will interfere with what models has already learned. This occurrence can lead to abrupt decrease in performance of a model. Although, this can be solved by retraining model every time new data is generated but this process is computationally expensive and also challenge to deploy new model in the same environment by maintaining accuracy as well. To limit this issue and keep results accurate feedback loop is used to ensure that model performance is maintained and improves when new data is added. The technique used is to feed model with fresh test data and by considering data which model has already predicted which ensures it is learning from new data and performing better in the future.