Big data, data science and machine learning

Using suitable methods from the field of machine learning and adding historical data from the data warehouse, measurable improvements in forecasting quality are achieved compared to the previous forecasting approach. By using machine learning methods (e.g. artificial neural networks, random forest), shipment volumes were forecast for selected logistics locations (HUBs) with low to moderate errors. It has been proven that methods based solely on historical data lead to improved forecasts even without the addition of expert knowledge.

The ParcelShop analysis looks for influencing factors that have an impact on ParcelShop shipment volumes. Using suitable methods from the field of data science (correlation analyses, significance tests, etc.), correlations between ParcelShop characteristics and shipment volumes were established, which will be taken into account in the future when planning and selecting ParcelShop locations.

PTA advises and supports the expansion of operational data warehouse reporting to include modern forms of visualization based on GEO information. In Hermes ParcelRadar, anonymized shipment data is linked with GEO coordinates and polygons and processed visually.

The recipient analysis determines how many consignments a recipient receives from HERMES on average per year. There is no unique recipient identification in the current delivery process. The challenge is to identify recipients based on anonymized, incomplete and non-standardized address data. Such big data problems are known in the literature as duplicate recognition/object identification (record linkage). The performance of the Teradata Aster analysis platform was tested for address matching.