利用三元組網絡處理不完整資料集之應用

Abstract

一種根據特徵子集合缺失的資料集應用場景被重新定義及命名為子集合缺失資料集。這種重新制定及定義有助於之後的場景應用以及簡化許多領域都會遇到的因資料集部分缺失的模型問題。此研究利用偽三元組神經元網絡之調用正反配對例子的特性來應對子集合缺失資料集的問題。此研究也涉及了一套資料前處理流程用於處理資料集以便用於模型之輸入。此研究對於多種資料集設定了多種不同比例的子集合缺失用於鑒定所設計模型的分類效能

The data scenario in which the absence of its value depends on the feature subset of the value is defined and formulated as Subset-incomplete data scenario. The formulation of the data scenario contributed to the future application of multiple field problems. Pseudo-triplet networks with positive and negative pairing and their corresponding data flow are proposed to tackle the subset-incomplete data scenario. Data preprocessing pipeline is designed to process subset-incomplete data into model-accepted data with positive and negative pairing. Experiments with various data feature settings for evaluation are observed to support the efficacy of the pseudo-triplet network.