跳转至

基于血缘引导的演化概念定义下的增量学习

文章背景与核心概要

传统的概念漂移方法通常依赖统计异常或预测误差来检测输入数据的变化,往往会忽略对目标定义所做的显式策略或规则更新。本文提出了一种血缘引导的增量学习框架(provenance-guided incremental learning framework),旨在解决规则引发的概念漂移(rule-induced concept shift)问题——在这种情况下,目标定义的概念被直接修订,从而在底层数据未变的情况下改变了历史标签。

为了评估该方法,作者推出了 RuleShift-Bench,这是一个涵盖金融、人口统计、网络安全和图结构数据集的综合性基准测试。所提出的方法能够高效追踪变更、验证历史记录的有效性并最小化计算开销,与完全重新标注和重新训练相比,它在实现高准确率的同时大幅降低了更新延迟。


📋 执行摘要 (Executive Summary)

传统的概念漂移方法通常依赖统计异常或预测误差来检测输入数据的变化,往往会忽略对目标定义所做的显式策略或规则更新。本文提出了一种血缘引导的增量学习框架,旨在解决规则引发的概念漂移——即目标定义概念被直接修订,在底层数据保持不变的情况下改变了历史标签。

Traditional concept-drift methods typically rely on statistical anomalies or prediction errors to detect changes in incoming data, often missing explicit policy or rule updates made to target definitions. This paper introduces a provenance-guided incremental learning framework designed to address rule-induced concept shift—where target-defining concepts are revised directly, altering historical labels without changes to the underlying data.

为了评估该方法,作者推出了 RuleShift-Bench,这是一个涵盖金融、人口统计、网络安全和图结构数据集的综合性基准测试。所提出的方法能够高效追踪变更、验证有效历史记录并最小化计算开销,与完全重新标注和重新训练相比,它在保持高准确率的同时大幅降低了更新延迟。

To evaluate this approach, the author presents RuleShift-Bench, a comprehensive benchmark spanning financial, demographic, cybersecurity, and graph-structured datasets. The proposed methodology efficiently tracks changes, certifies valid historical records, and minimizes computational overhead, achieving high accuracy and substantially reduced update latency compared to complete relabeling and retraining.


🔍 核心概念与方法论 (Key Concepts & Methodology)

  • 规则引发的概念漂移: 专注于目标概念直接发生变化(通过修改规则、策略或查询)的场景,这导致存储的实例在原始数据未变的情况下获得新的语义标签。
    • Rule-Induced Concept Shift: Focuses on scenarios where the target concept changes directly (via modified rules, policies, or queries), causing stored instances to acquire new semantic labels despite unchanged raw data.
  • 结构化规则增量: 将连续的概念定义编译为增量结构,以精确指出具体发生了哪些变化。
    • Structured Rule Deltas: Compiles consecutive concept definitions into delta structures to pinpoint exactly what changed.
  • 历史血缘追踪: 穿透历史数据追踪发生变化的组件,以认证其先前标签依然严格有效的记录。
    • Historical Provenance Tracking: Traces changed components through historical data to certify records whose previous labels remain strictly valid.
  • 局部候选区域: 将重新评估限制在相关区域内,对修订部分执行自动重新标注,同时通过选择性监督来路由含糊不清的案例。
    • Localized Candidate Regions: Restricts reevaluation to relevant areas, executing automated relabeling for revisions while routing ambiguous cases through selective supervision.
  • 版本化概念记忆: 维护历史上下文,以无缝支持循环出现的概念定义。
    • Versioned Concept Memory: Maintains historical context to seamlessly support recurring concept definitions.

📊 基准测试与实验结果 (RuleShift-Bench)

RuleShift-Bench 利用阈值、谓词、逻辑、关系、循环和混合概念修订,对不同领域(金融、人口统计、网络安全和图结构数据)的系统进行了严格测试。

RuleShift-Bench rigorously tests systems across diverse domains (financial, demographic, cybersecurity, and graph-structured data) using threshold, predicate, logical, relational, recurring, and mixed concept revisions.

  • 准确率 (Accuracy): 92.3%
    • Accuracy: 92.3%
  • 宏平均 F1 分数 (Macro-F1 Score): 90.2%
    • Macro-F1 Score: 90.2%
  • 数据效率 (Data Efficiency): 仅重新处理历史集合的 14.7%,同时保留了 94.6% 的未受影响记录。
    • Data Efficiency: Reprocesses only 14.7% of the historical collection while retaining 94.6% of unaffected records.
  • 更新延迟 (Update Latency): 平均更新延迟达到 179秒,相比完全重新标注和重新训练所需的 993秒 实现了巨大改进。
    • Update Latency: Achieves an average update latency of 179 seconds, a massive improvement over the 993 seconds required for complete relabeling and retraining.