潜在パターン抽出による行動誘導：ユーザ軌跡分析における次元性駆動型イノベーション
Year: 2025
Type: dissertation
Abstract: To address the risk level assessment of multi-view traffic risk records, we propose a Self-weighted Multiview Spectral Clustering on Multiple Manifolds (SwMMC) algorithm. SwMMC leverages the Constrained Laplacian Rank framework to directly partition data by imposing rank constraints on the similarity matrix, eliminating the need for post-processing steps like K-means. To reduce noise, it combines similarity measures from both high-dimensional and projected low-dimensional spaces during matrix construction. By alternatingly optimizing the ideal affinity matrix and view weights, SwMMC adaptively fuses multiple views. Comparative experiments on two real-world datasets demonstrate its superior performance in risk level classification across diverse conditions. For traffic networks, traffic signal control is crucial for improving transportation efficiency. Traditional reinforcement learning methods for traffic control have problems like slow learning speed and poor adaptability to new situations. To solve these issues, we introduce MetaSignal, a meta-reinforcement learning approach that uses Fourier basis linear function approximation. This approach maps high-dimensional traffic states to low-dimensional linear combinations via Fourier basis, reducing memory and computation costs while capturing feature interactions like queue intensity and phase time. Combined with a model-agnostic meta-learning framework, MetaSignal enables fast adaptation to dynamic traffic without extensive retraining. Empirically, the proposed method shows promising and stable performance for traffic signal control through comprehensive comparison experiments in both synthetic and real-world traffic networks. In summary, this research proposes two paradigm-specific frameworks for behavioral trajectory guidance: SwMMC for multi-view traffic risk assessment and MetaSignal for traffic signal control. By integrating constrained spectral ranking and Fourier basis approximation, the study advances dimensionality-adaptive methodologies, demonstrating enhanced performance in risk classification and traffic control across real-world scenarios.
DOI: 10.15118/0002000404
Authors: HUANG, SHUNING, ホアン, シュニン
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7110826699
Cites:
Cited by:
Related to: