DSpace community: 管理學院

資料前處理與分類器建構之集成學習技術於類別不平衡資料之研究;A Study on Ensemble Learning Techniques for Data Preprocessing and Classifier Construction in Imbalanced Data

Fri, 06 Mar 2026 10:55:11 GMT

title: 資料前處理與分類器建構之集成學習技術於類別不平衡資料之研究;A Study on Ensemble Learning Techniques for Data Preprocessing and Classifier Construction in Imbalanced Data abstract: 在真實世界中，類別不平衡（Imbalanced Data）問題廣泛存在，如設備故障預測與醫療診斷。由於傳統機器學習模型通常偏向預測多數類別，因此如何提升分類器對少數類別的辨識能力成為重要課題。目前針對類別不平衡問題的解決策略主要可分為資料層級、演算法層級以及混合層級三大類，但在資料層級方面，現有文獻尚缺乏對於如何將集成學習（Ensemble Learning）概念應用於資料前處理的深入探討。此外，也鮮有研究將集成學習運用於多重分類器的篩選中。因此本研究針對此缺口，探討集成學習在資料前處理與分類器建構上對分類表現的影響，使用來自KEEL資料庫的42個類別不平衡資料集，並設計兩組實驗：（1）選用三種重採樣演算法（SMOTE、Cluster Centroids和SMOTEENN）與四種案例選取演算法（ENN、DROP3、IPF和CVCF），並設計12種不同的資料前處理流程，比較不同資料前處理方法（單一與集成）對分類表現的影響，以找出最佳的資料前處理方法；（2）結合六種動態選取演算法（OLA、MLA、MCB、DES-KNN、KNORA-U和DES-P）進行多重分類器建構，評估資料層級與分類器層級集成的協同效果。實驗結果顯示，採用重採樣的多重交集方法能提升訓練資料的多樣性與品質並增強分類效能，而所有分類器中以Random Forest表現最優異。而在整合策略方面，將SMOTE後搭配ENN，並結合SVM、CART、KNN三個分類器與KNORA-U動態選取技術，可在AUC指標上取得最優表現（0.863）；若重視少數類別的預測能力則建議採用IPF後進行重採樣的聯集，並搭配SVM、KNN、Random Forest（或XGBoost）三個分類器與KNORA-U，在F1-Measure指標上表現最佳（0.739），最終整合策略可依據實際應用情境與預測重點來做選擇。;Imbalanced data is common in real-world applications such as equipment failure prediction and medical diagnosis. Traditional machine learning models often favor the majority class. Therefore, improving a classifier’s ability to recognize the minority class has become a key challenge. However, current literature lacks exploration of how ensemble learning can be incorporated into data preprocessing at the data level. Additionally, few studies have applied ensemble learning to the selection of multiple classifiers. To address these gaps, this study investigates the impact of ensemble learning on classification performance in both data preprocessing and classifier construction. A total of 42 imbalanced datasets from the KEEL repository were used, and two sets of experiments were designed: (1) Twelve distinct data preprocessing workflows were designed by three resampling algorithms (SMOTE, Cluster Centroids, and SMOTEENN) and four instance selection algorithms (ENN, DROP3, IPF, and CVCF). These workflows, incorporating both single and ensemble learning based data preprocessing approaches, were identify to determine the most effective preprocessing strategy for handling imbalanced data; (2) Integrate six dynamic selection algorithms (OLA, MLA, MCB, DES-KNN, KNORA-U, and DES-P) for multiple classifier construction to evaluate the synergistic effects of combining data-level and classifier-level ensembles. Experimental results show that employing a multi-intersection resampling approach can enhance the diversity and quality of training data, thereby improving classification performance. Random Forest demonstrated the best overall performance. Regarding integration strategies, applying SMOTE followed by ENN, and integrating SVM, CART, and KNN with the dynamic selection technique KNORA-U, achieved the highest AUC (0.863). For tasks prioritizing minority class prediction, the recommended strategy is to apply IPF followed by a union of resampling approach, combined with SVM, KNN, and Random Forest (or XGBoost), along with KNORA-U. This approach achieved the best F1-Measure (0.739). The final integration strategy can be selected according to specific application scenarios and predictive objectives.

基於超圖神經網路與自注意力機制的服裝搭配相容性預測;基於超圖神經網路與自注意力機制的服裝搭配相容性預測

Fri, 06 Mar 2026 10:54:47 GMT

title: 基於超圖神經網路與自注意力機制的服裝搭配相容性預測;基於超圖神經網路與自注意力機制的服裝搭配相容性預測 abstract: 時尚產業為一持續發展的領域，消費者對穿搭組合的需求日益提升，尤其隨著電子商務平台的快速發展，如何提供兼具準確性與多樣性的服裝推薦成為推薦系統研究中的重要議題。然而，傳統服裝相容性預測方法常面臨多模態訊息不一致、圖文特徵整合困難，以及模型難以捕捉高階依賴與全局上下文語意等挑戰，導致實際應用中無法有效掌握穿搭準確性，特別是在圖像與文字特徵落於不同嵌入空間的情況下，現行方法的表現仍顯不足。本研究旨在提升整體服裝相容性預測效能與搭配準確性，提出一種整合大型語言模型（LLM）、FashionCLIP、超圖神經網路（HGNN）與Transformer的多階段相容性預測架構。首先，研究僅以單品圖像作為輸入，並運用先進的大型語言模型gpt-4o生成語義描述，涵蓋單品的色彩、材質、風格等語意特徵，進而補足視覺訊息所無法直接表達的語義資訊。接著，透過FashionCLIP對圖像與文字進行對齊與特徵融合，將每個單品轉化為語意一致的多模態嵌入表示。為進一步捕捉單品間的高階依賴關係，本研究構建以服裝類別為節點的超圖（Hypergraph），並使用HGNN進行消息傳遞與語意聚合，有效保留搭配中隱含的結構關係與多對多互動特性。此外，研究引入Transformer以強化全局語意建模能力，進一步整合HGNN後的節點嵌入表示，彌補其在長距依賴建模上的不足。本模型最終產生一組整體outfit相容性表示，並於Polyvore與Zalando兩大公開資料集上進行實證分析。實驗結果顯示，所提方法在相容性排序準確度上明顯優於現有多項基準模型，驗證本研究在多模態整合與上下文語意建模上的創新設計具有效益與實用性。;Fashion is a long-standing industry, with consumers exhibiting a persistent need for outfit coordination. As the importance of fashion recommendation systems increases in e-commerce and styling applications, outfit compatibility prediction has emerged as a key research focus. However, previous methods face limitations due to inconsistencies in auxiliary information, challenges in multimodal integration, and insufficient modeling of high-order dependencies and global context. These issues restrict both generalizability and accuracy in practical applications. Specifically, current methods remain suboptimal when visual and textual features lie in different representational spaces or when hypergraph neural networks (HGNNs) suffer from information aggregation loss in fine details. The objective of this study is to improve outfit compatibility prediction and enhance matching accuracy by proposing a multi-stage prediction model that integrates a Large Language Model (LLM), FashionCLIP, Hypergraph Neural Network (HGNN), and Transformer. The model begins by using only the fashion item image—which contains the most visual information—as input. A rapidly evolving LLM, GPT-4o, is then employed to generate detailed textual descriptions of each item, including specific fashion attributes, as well as its category label. FashionCLIP, which is tailored for fashion-related tasks, is used to align the image and textual features, forming consistent multimodal inputs. These aligned features are used to construct a hypergraph, where a message-passing mechanism from a graph convolutional network aggregates neighboring node information and updates node representations. The updated node embeddings are subsequently fed into a Transformer to capture fine-grained details and long-range dependencies within the global context, addressing the limitations of HGNN. The combination of HGNN and Transformer forms the final compatibility prediction module. The proposed model is evaluated on the Polyvore Disjoint and Zalando datasets. Experimental results demonstrate that the proposed method outperforms existing approaches in outfit compatibility prediction, thereby validating its effectiveness.

探討社會支持對軟體開發人員在勞動敏捷性的影響：外控取向作為調節作用;Exploring the Impact of Social Support on Workforce Agility among Software Developers: The Moderating Role of External Locus of Control

Fri, 06 Mar 2026 10:54:36 GMT

title: 探討社會支持對軟體開發人員在勞動敏捷性的影響：外控取向作為調節作用;Exploring the Impact of Social Support on Workforce Agility among Software Developers: The Moderating Role of External Locus of Control abstract: 隨著時代快速變遷，企業唯有持續交付創新且具高價值的產品，方能維持市場競爭力，尤以變動迅速的軟體開發環境為甚。具備勞動敏捷性特質之開發人員因而成為關鍵角色，如何有效提升其勞動敏捷性，亦成為當前極需關注的核心議題。有鑑於此，本研究結合社會支持相關理論與模型，探討不同類型之社會支持是否能協助開發人員因應工作挑戰，並促進其正向行為表現。同時，引入外控取向作為調節變項，進一步探討個體特質在社會支持與勞動敏捷性關係中的影響機制，從而建構一套整合勞動敏捷性、社會支持與外控取向的創新研究模型，以補足過往文獻的不足。本研究透過問卷調查法蒐集樣本資料，共回收312份有效問卷，並採用偏最小平方法結構方程模型 (PLS-SEM) 進行資料分析與假說驗證。研究結果顯示，資訊性支持、情感性支持與評估性支持皆對勞動敏捷性產生正向且顯著的影響。對於未達顯著水準之假說，亦有進行深入討論與合理解釋。整體而言，本研究有助於豐富軟體開發領域中關於社會支持、開發人員個人行為表現與特質之理論探討，並可作為實務上進行資源配置與組織規劃之參考依據。最後，針對研究限制，提出相應之未來研究方向，以供後續實務與學術研究發展。;In an era of rapid change, enterprises must continuously deliver innovative, high-value products to remain competitive—especially in the fast-paced software development environment. Accordingly, developers with workforce agility have become key players, and enhancing their agility has emerged as a critical issue. This study integrates theories and models of social support to explore whether various types of support can help developers cope with work challenges and foster positive behavioral outcomes. The concept of external locus of control is also introduced as a moderating variable to examine how individual traits influence the relationship between social support and workforce agility, thereby addressing gaps in existing literature through the construction of an integrated research model. A total of 312 valid responses were collected via questionnaire, and data were analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM). The findings show that informational, emotional, and appraisal support each have a significant positive impact on workforce agility. In addition, non-significant hypotheses are also thoroughly discussed and reasonably interpreted. Overall, this study contributes to theoretical discourse on social support, personal traits, and behavioral outcomes in software development, while offering practical implications for organizational planning and resource allocation. Future research directions are also proposed as a foundation for continued academic inquiry and practical advancement.

Summarization-Enhanced BERT with Phonetic and Glyph embeddings for Chinses Spelling Check

Fri, 06 Mar 2026 10:54:11 GMT

title: Summarization-Enhanced BERT with Phonetic and Glyph embeddings for Chinses Spelling Check abstract: 因中文字結構複雜、同音字多與字形相似等特性，中文拼寫檢查（Chinese Spelling Check, CSC）面臨諸多挑戰，使得錯字偵測高度依賴語境理解。本研究提出一個新穎的 CSC 框架—Summarization-Enhanced BERT（SE-BERT），結合句子摘要特徵、字音資訊與字形嵌入，以提升模型在錯誤偵測與糾正任務中的語意感知能力。該模型由摘要模組、偵測網路與糾正網路三部分組成，並加入錯誤導向遮罩機制，提供更具針對性的修正指引。在 SIGHAN 標準資料集上進行實驗後顯示，SE-BERT 在準確率與錯誤識別能力方面皆優於現有基準模型，且能有效降低過度修正的情形。注意力視覺化與個案分析亦驗證模型能聚焦於語意關鍵位置。整體而言，本研究證實整合語意、語音與視覺資訊對於提升中文拼寫校正成效的重要性，並提供一個具結構性且可擴展的拼字校正解決方案，適用於語言特性多變的應用場景。;Due to the structural complexity of Chinese characters, the high occurrence of homophones, and visual similarity among glyphs, Chinese spelling check (CSC) presents unique challenges. These factors make typo detection highly context-dependent. This study proposes a novel CSC framework, Summarization-Enhanced BERT (SE-BERT), which integrates phonetic and glyph embeddings with sentence-level summarization features to enhance context awareness in error detection and correction. The model consists of a summarization module, a detection network, and a correction network, augmented with an error-guided mask that guides more precise correction. Experiments conducted on benchmark datasets, including SIGHAN, demonstrate that SE-BERT achieves superior performance compared to existing baselines, particularly in reducing miscorrections and improving accuracy. Attention visualization and case studies further confirm the model′s ability to focus on key contextual cues. These findings highlight the importance of multi-source information integration like semantic, phonetic, and visual, for effective CSC, offering a structured and adaptable approach for spelling correction in linguistically complex environments.