基于视图可靠性与冲突协调的多视图序贯三支决策模型

A Multi-View Sequential Three-Way Decision Model Based on View Reliability and Conflict Coordination

摘要: 随着数据来源和信息获取渠道的多样化，同一决策对象通常可以由多个具有不同特征结构和信息质量的视图共同描述。现有多视图序贯三支决策研究为多视图、多层次信息的渐进式处理提供了有效框架，但大多默认视图质量同质且判断一致鲜有将视图可靠性评估、序贯信息融合与视图冲突协调纳入统一框架的系统性研究。若忽略视图质量差异或直接融合存在冲突的视图信息，易导致不可靠信息过度影响综合决策，增加对象被过早接受或拒绝的风险。针对上述问题，本文提出一种基于视图可靠性与冲突协调的多视图序贯三支决策方法。首先，根据各视图对决策类别的概率判别表现构造视图可靠性指标，并据此建立由高可靠视图向低可靠视图逐步扩展的视图序列。其次，在各阶段融合已引入视图的条件概率信息，并利用视图间概率判断的加权离散程度构造对象级视图冲突度。进一步，根据“冲突越强，决策确定性应越低”的原则设计冲突修正机制，将存在较大视图分歧的综合条件概率向不确定状态进行收缩。最后，将修正后的条件概率引入三支决策规则，使可确定对象在当前阶段完成接受或拒绝，仅将边界域对象传递至下一阶段继续引入新的视图信息。该模型有机整合了视图质量评价、跨视图冲突协调与序贯渐进决策三个核心环节，实现了三者的统一数学描述与机制协同。基于多组公开多视图数据集的对比实验结果表明，通过与传统多视图三支决策、经典序贯多视图三支决策及消融基线模型对比，所提方法在决策准确率、延迟率、平均使用视图数及综合决策损失等指标上均取得最优性能，有效兼顾了决策精度与信息利用效率。

Abstract: With the diversification of data sources and information acquisition channels, the same decision object can often be described by multiple views characterized by heterogeneous feature structures and varying information quality. Existing studies on multi-view sequential three-way decision-making provide an effective framework for the progressive processing of multi-view and multi-level information. However, most of them assume homogeneous view quality and consistent judgments, and few have systematically integrated view reliability evaluation, sequential information fusion and inter-view conflict coordination into a unified framework. Ignoring differences in view quality or directly fusing conflicting information from multiple views may allow unreliable information to exert excessive influence on the aggregated decision, thereby increasing the risk of prematurely accepting or rejecting objects. To address these issues, this paper proposes a multi-view sequential three-way decision-making method based on view reliability and conflict coordination. First, a view reliability measure is constructed according to the probabilistic classification performance of each view with respect to the decision categories. On this basis, a view sequence is established in which views are progressively introduced from higher to lower reliability. Second, at each stage, the conditional probability information from the views introduced thus far is aggregated, and an object-level view conflict measure is constructed using the weighted dispersion of the probabilistic judgments across views. Furthermore, following the principle that stronger conflict should correspond to lower decision certainty, a conflict correction mechanism is designed to shrink the aggregated conditional probability toward the uncertain state when substantial disagreement exists among the views. Finally, the corrected conditional probabilities are incorporated into the three-way decision rules. Objects that can be determined with sufficient certainty are accepted or rejected at the current stage, whereas only objects in the boundary region are transferred to the next stage, where additional view information is introduced. While preserving the deferment mechanism of three-way decision-making, the proposed model provides a unified representation of view reliability, inter-view conflict, and sequential information fusion. The proposed model organically integrates three core components: view quality evaluation, cross-view conflict coordination and sequential progressive decision-making, and realizes a unified mathematical formulation and synergistic mechanism among them. Comparative experiments on multiple public multi-view datasets demonstrate that, compared with traditional multi-view three-way decision-making, classical sequential multi-view three-way decision-making and ablation baseline models, the proposed method achieves optimal performance in terms of decision accuracy, deferment rate, average number of views used and overall decision loss, and effectively balances decision accuracy and information utilization efficiency.