Improving Human Mobility Forecasts: A Study on Outlier Correction with Multi-Agent Techniques

Abstract This research is part of designing and implementing an AutoML platform for time series forecasting. Having discussed missing value imputation and outlier detection in other research papers, this research is focused on outlier correction to forecast better in time series. This research has used human mobility data as a sample that was collected in Hiroshima, Japan, over a year. Master-Slave multi-agent design pattern was used with an outlier co-ordinator agent that direct other subagents to identify the best regression technique and correct the outlier data points. Extensible regression agents are used to determine the best regression technique using seasonal data and evaluated with the harmonic mean of regression evaluation parameters. Outlier correction was done iteratively to find the optimum outlier corrections, as fixing all the outlier data points is impossible. This research achieved different levels of outlier corrections for the ten selected locations in Hiroshima, Japan. These decisions are integrated with classification techniques as an emerging knowledge capability in the multi-agent.