Pair-wise selective classification with dynamic sampling for shipment importer prediction

Research output: Chapter in Book/Report/Conference proceedingConference Article in proceedingAcademicpeer-review

Abstract

Whenever a shipped package crosses the border, logistic companies have to declare importer information for the clearance process. This information is not always provided by the customer, causing delays and additional expenses. Fortunately, importer information can often be inferred from historical shipments. The current technical standard, even in big companies, is to use a feature-weighted nearest neighbor approach based on domain knowledge. Nearest neighbors assume that each sample point is represented independently and is fixed in some high dimensional space. This makes it difficult to integrate higher-order pair-wise relationships such as transaction frequency from shipper to receiver transaction, because the features are now changing dependent on the pairs. This would require a complex ad-hoc feature engineering and metric learning to capture pairwise relationships properly with nearest neighbors. In this paper, we propose a framework for importer prediction based on a pair-wise classification approach that allows us to capture higher order pair-wise relationships. We also incorporate an auxiliary neural network that can reliably reject shipments that our model could not predict well such as shipments with new importers that are not in the historical data. This allows us to pass the difficult cases to a human agent instead of naively making an incorrect prediction. Our proposed pair-wise solution outperforms the industry standard by a significant margin of precision across a wide range of recall values.

Original languageEnglish
Title of host publicationICMLC 2023
Subtitle of host publicationProceedings of the 2023 15th International Conference on Machine Learning and Computing
PublisherAssociation for Computing Machinery
Pages152-157
Number of pages6
ISBN (Print)9781450398411
DOIs
Publication statusPublished - 7 Sept 2023
Event15th International Conference on Machine Learning and Computing, ICMLC 2023 - Hybrid, Zhuhai, China
Duration: 17 Feb 202320 Feb 2023
http://www.icmlc.org/cft.html

Conference

Conference15th International Conference on Machine Learning and Computing, ICMLC 2023
Country/TerritoryChina
CityZhuhai
Period17/02/2320/02/23
Internet address

Keywords

  • Dynamic sampling
  • Ranking
  • Selective classification

Fingerprint

Dive into the research topics of 'Pair-wise selective classification with dynamic sampling for shipment importer prediction'. Together they form a unique fingerprint.

Cite this