2026-10-14T17:32:02Z https://www.scidb.cn/oai

10.57760/sciencedb.12458 2024-08-28T17:11:22Z

2024-08-28 Bidirectional Weighted Self-Organizing Incremental Neural Network for Multi-Label Learning doi:10.57760/sciencedb.12458 en Data Generation Program:- Construction of Multi-Label Datasets: The dataset comprises multiple features and their corresponding sets of labels. For instance, an image may encompass various semantic concepts such as "sky," "lake," and "building."- Feature Selection Techniques: These include filtering methods, wrapper methods, and embedding approaches, utilized to reduce data dimensionality and enhance the performance of multi-label learning models.- Self-Organizing Incremental Neural Network (SOINN): An SOINN algorithm with adjusted parameters is proposed to strengthen supervised learning and rapid classification.- Correntropy Induced Metric (CIM): A similarity measurement method based on the Gaussian kernel is introduced.Data Processing Methods and Steps:- Partitioning of Label Space: An improved Fibonacci affinity propagation (AP) clustering method is employed for dividing the label space based on label correlations.- Multi-Label Rough Membership Function: The concept of multi-label classification margin is introduced, with definitions for adaptive neighborhood radius and neighborhood class.- Multi-Label Fuzzy Neighborhood Rough Set Model: Fuzzy neighborhood entropy and multi-label fuzzy neighborhood entropy are defined, and a bidirectional multi-label feature selection (BMFS) algorithm based on the weighted information granulation of fuzzy neighborhood entropy is proposed.- C-SOINN Algorithm: The integration of SOINN and CIM is achieved, controlling the insertion of nodes through parameter adjustment and similarity thresholds based on inter-class insertion.- Experimental Evaluation: The C-SOINN algorithm is experimentally evaluated using multiple real-world multi-label datasets and compared with existing multi-label learning methods.Equipment and Tools Used:- Computational Environment: Python 3.8 environment with an AMD Ryzen 7 4800H processor, 40GB RAM, and an NVIDIA GeForce RTX 2060 graphics card.- Datasets: Standardized multi-label datasets from Mulan and Extreme Classification.- Evaluation Metrics: Average Precision (AP), Hamming Loss (HL), Ranking Loss (RL), and Coverage (CV).- Classification Framework: Multi-Label K-Nearest Neighbors algorithm (MLKNN).Experimental Design:- Parameter Tuning: For machine learning algorithms, parameter tuning is crucial for performance. Grid search was used to adjust the parameters for each algorithm in the experiments.- Performance Comparison: The effectiveness of the proposed algorithm is validated by comparing its performance with other multi-label classification methods. Multi-label learning; Feature selection; Fuzzy neighborhood entropy; neural network Weiguo Yi PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank