Impact of data quality on the prediction success of auto-classification tools

Nowadays, compagnies with fast-paced logistics processes and involved in international trade must have streamlined processes to efficiently comply with international trade regulations and customs requirements. This is particularly true when companies must adapt, in a timely manner, to external regulation changes or internal events, such as the launching of a new product line. Product classification, also known as the Harmonized System (HS) of goods at World Customs Organization (WCO) level or as the Combined Nomenclature at EU level, is one of the domains impacted by the above-mentioned changes. This thesis aims at better understanding one specific aspect of the products classification: the impact of data quality on the auto-classification tools prediction scores.

After a familiarization with the theoretical concepts of product classification, two products data sets are analyzed. The first step of the analysis consists in confirming that the auto-classification tool chosen beforehand (based on criteria such as availability and ease of use) is appropriate for the classification of goods with the data sets at our disposal. This is achieved by submitting the unaltered (except formatting) data set to the auto-classification tool and compare the actual classification with the predicted classification. The tool is considered appropriate if its accuracy is above 90% (a subjective threshold) when used with both data sets.