DRL-DC: An Adaptive Deep Reinforcement Learning Framework for Criminal Subject Data Filtering

• Lei Guan,
• Donghong Zhang,
• Jinduan Yan,
• Dongzhi Cao,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.009, Download Paper (Free):

• Data Filtering
• Deep Reinforcement Learning
• Markov Decision Process
• Data Preprocessing

Abstract

With the rapid advancement of information technology, multi-source criminal data has become increasingly complex, noisy, and unreliable. Conventional cleaning techniques based on static models struggle to handle such heterogeneity. To overcome these limitations, this paper proposes DRL-DC, an adaptive data filtering method based on deep reinforcement learning (DRL). The proposed framework formulates the data cleaning process as a Markov Decision Process (MDP) and utilizes a Deep Q-Network (DQN) to learn optimal filtering strategies through interaction with the environment. Experimental evaluations on real-world Facebook data and public benchmark datasets demonstrate that DRL-DC achieves an 8.2% improvement in identity recognition accuracy compared with Learn2Clean, a 15.4% reduction in regression error relative to AutoML, and a clustering accuracy of 0.823 on Google Play datasets. Furthermore, DRL-DC consistently maintains a precision close to 0.9, confirming its robustness and adaptability across diverse data sources.

Statistics

Cite this article