Partial Granger causality-based feature selection algorithm for workload prediction in cloud systems.

The development of the cloud technology led to the rising interest in multi/hybrid cloud and emergence of artificial intelligence for IT operations(AIOps). The core elements of AIOps involve predicting how each metric of the data center will change in the future. Compared to general multivariate time-series prediction problems, causal relationships between each variable have a significant impact on cloud metric prediction. This paper focuses on the causality between variables, which partially arises when a specific event occurs in a cloud data center to achieve good predictive performance. The proposed model detects partial causality and sums it up again to extract key variables that explain the target variable well. Through this, variables with higher predictive performance than existing methods were found. We also propose a structure that improves the performance of the prediction model and minimizes inference time through a variable selection technique based on partial causality. By applying this to the actual operating cloud environment, it was proved that it can be effectively applied to the real world.