Challenges in Streaming Data Analysis for Building an Adaptive Model for Handling Concept Drifts

Incremental learning from non-stationary environments with evolving streaming data is one of the main challenges in streaming data analysis (SDA). The advances in data acquisition technology have ignited the importance of data analysis on streaming data. Analysis on streaming data has to be performed with constrained memory and single pass over the data. The main challenges are in window adjustments, various drift detections, feature selection, dynamic model selection, adaptability of models and it’s real-time processing. The change in characteristics of data stream often make our model obsolete. Such changes are called concept drift. This concept drift mandates model adaptations in SDA. Adaptations are required at the feature selection, parametric adjustments and model selection. Making models with the capability of detecting abrupt, gradual and virtual drift is one of the challenges addressed by the researchers in this field. Upon detecting concept drifts, it is required to make necessary changes to the models for its acceptance. Statistical constraints have to be overcome for developing such adaptive algorithms. The model adaptability can be achieved much faster with good feature selection strategies which in turn reduces resource conception without affecting performance of model. Here the objective is to survey on building an adaptive model which is capable of handling concept drifts occurring in non-stationary data with limited resources without degradation in performance.