The Journal of Grey System ›› 2026, Vol. 38 ›› Issue (5): 119-133.

Previous Articles    

Cross-Sequence Grey Relational Operators for XGBoost-Based Supplier Selection

  

  1. 1. College of Information Engineering, Ningde Normal University, Ningde, Fujian, 352100, P.R. China 
    2. College of Physics and Information Engineering, Minnan Normal University, Zhangzhou, Fujian, 363000, P.R. China
  • Online:2026-11-16 Published:2026-11-16

Abstract: Supplier selection in manufacturing supply chains is methodologically demanding, and more so under the small-sample conditions typical of emerging economies. This study proposes a hybrid CS-GRA-XGBoost framework that brings the cross-sequence relational structure of grey systems theory together with eXtreme Gradient Boosting (XGBoost). Two structural operators are introduced: Cross Degree (CD), the normalised frequency at which two supplier trajectories cross, and Cross Intensity (CI), the mean displacement magnitude at those crossings. Both are proved independent of the Deng relational grade and complementary to integral cross-sequence models. Because the operators are pairwise, we define an aggregated form of the cross-relational feature vector, which allows reference construction, relational grades, feature computation and hyperparameter search all to be nested inside the training partition of a cross-validation split. The framework is evaluated in a controlled simulation with three scenarios, differing in whether the tier label depends on the level of the performance trajectories, on their stability, or on both; the first is a negative control in which the operators must not help. Where the label depends on stability, the cross-relational representation classifies far better than the relational grade (Decision Accuracy, the proportion of objects assigned to the correct tier, 0.678 against 0.328 at 240 objects). Where it depends on level alone, it classifies no better. The operators extend grey relational analysis in a targeted way; they do not improve it across the board. Feature importance follows the same pattern, the relational grades carrying 0.594 of total gain where the label depends on level and 0.058 where it depends on stability, and the results are insensitive to the distinguishing coefficient. One further finding bounds what the method can be used for: at fifteen objects the learned representations cannot be told apart in any scenario, so populations of sixty or more are needed before the framework’s advantage can be shown at all.

Key words: Cross-sequence Grey Relational Analysis, Cross Degree, Cross Intensity, XGBoost, Supplier Selection, Grey Systems Theory, Small-sample Learning