Software Frameworks for Explainable AI in Time Series Classification: A Systematic Review
本文系统地回顾了时间序列分类中可解释AI的软件框架，比较了不同维度的方法，指出了现有框架的局限性，并探讨了未来研究方向。
0 citationsRead paper
Selected work
Representative Papers
Beyond Local Inspection: Global, Guideline-Grounded Evaluation of Post-hoc XAI Methods for ECG Classification
This work addresses the limitations of existing post-hoc explainable AI (XAI) methods in electrocardiogram (ECG) classification, which often produce explanations misaligned with clinically relevant features, obscuring whether models rely on genuine pathological signals. The authors propose the first guideline-based global evaluation framework, leveraging cardiologist-defined critical ECG regions from clinical guidelines as ground truth to systematically assess the reliability of 13 gradient-based XAI methods across four binary classification models on the PTB-XL dataset. By aggregating multi-beat explanations and employing Spearman correlation alongside region-wise attribution ratios, they reveal that most methods disproportionately emphasize high-amplitude QRS complexes while neglecting diagnostically crucial low-amplitude ST segments—for instance, LRP-ε assigns only 4.6% attribution to ischemic ST segments compared to LRP-SIGN’s 63.8%. Alarmingly, nine out of thirteen methods perform below random baseline for at least one pathology, exposing a systemic failure of vision-domain XAI techniques when naively transferred to medical time-series signals.
0 citationsRead paper
Breadth-First Search in Succinct Planar Graphs
This work addresses the challenge of efficiently performing breadth-first search (BFS) and related graph algorithms directly on highly compressed representations of planar graphs while strictly limiting auxiliary space usage. The authors propose the first succinct encoding of planar graphs that can be constructed in linear time, supports native BFS traversal, and requires only o(n) additional bits of space. This representation enables constant-time queries on both the BFS tree and the original graph structure. Furthermore, it extends to several advanced applications, including computing balanced separators of size O(√n), diameter-based tree decompositions, triangulation, and bipartiteness testing. By integrating techniques from succinct data structures, planar graph embedding theory, and dual-graph traversal, the method achieves sublinear space complexity without sacrificing linear-time efficiency.
0 citationsRead paper
Sim4IA-Bench: A User Simulation Benchmark Suite for Next Query and Utterance Prediction
Existing user simulation studies lack reproducible and interpretable evaluation benchmarks, particularly for next-query and next-utterance prediction tasks, due to the absence of authentic conversational context. To address this, we propose Sim4IA-Bench—the first benchmark for user behavior simulation in information retrieval—built upon 160 real search sessions (including a publicly available subset of 70). It establishes the first systematic alignment between authentic interaction trajectories and multi-turn simulated predictions. We introduce a dual-task evaluation framework: Task A (next-query prediction) and Task B (next-utterance prediction), incorporating novel metrics tailored to intent drift and query reformulation. The benchmark leverages high-fidelity simulation environments derived from CORE search engine data. Sim4IA-Bench provides a unified, reproducible platform for evaluating user simulators and significantly advances interaction-aware retrieval evaluation.
0 citationsRead paper
THM@SimpleText 2025 -- Task 1.1: Revisiting Text Simplification based on Complex Terms for Non-Experts
This study addresses the challenge of low comprehensibility of context-poor, ultra-short scientific sentences—such as abstracts and titles—for non-expert readers. We propose a term-level simplification method that first precisely identifies domain-specific complex terms within the input sentence, then employs coordinated lightweight Gemini and OpenAI models for controlled, semantics-preserving rephrasing—avoiding global rewriting to prevent information distortion. Our approach integrates rule-based terminology detection with large language models’ semantic fidelity in generation. Evaluated on CLEF SimpleText 2025 Task 1.1, it significantly improves readability and accuracy for low-context scientific sentences. Experiments demonstrate superiority over baselines across three dimensions: plausibility of term replacement, preservation of original meaning, and comprehension by non-experts. The method thus enhances public accessibility of scientific knowledge without compromising technical integrity.
0 citationsRead paper
Recent publications
Latest Papers
Software Frameworks for Explainable AI in Time Series Classification: A Systematic Review
本文系统地回顾了时间序列分类中可解释AI的软件框架，比较了不同维度的方法，指出了现有框架的局限性，并探讨了未来研究方向。
0 citationsRead paper
Beyond Local Inspection: Global, Guideline-Grounded Evaluation of Post-hoc XAI Methods for ECG Classification
This work addresses the limitations of existing post-hoc explainable AI (XAI) methods in electrocardiogram (ECG) classification, which often produce explanations misaligned with clinically relevant features, obscuring whether models rely on genuine pathological signals. The authors propose the first guideline-based global evaluation framework, leveraging cardiologist-defined critical ECG regions from clinical guidelines as ground truth to systematically assess the reliability of 13 gradient-based XAI methods across four binary classification models on the PTB-XL dataset. By aggregating multi-beat explanations and employing Spearman correlation alongside region-wise attribution ratios, they reveal that most methods disproportionately emphasize high-amplitude QRS complexes while neglecting diagnostically crucial low-amplitude ST segments—for instance, LRP-ε assigns only 4.6% attribution to ischemic ST segments compared to LRP-SIGN’s 63.8%. Alarmingly, nine out of thirteen methods perform below random baseline for at least one pathology, exposing a systemic failure of vision-domain XAI techniques when naively transferred to medical time-series signals.
0 citationsRead paper
Breadth-First Search in Succinct Planar Graphs
This work addresses the challenge of efficiently performing breadth-first search (BFS) and related graph algorithms directly on highly compressed representations of planar graphs while strictly limiting auxiliary space usage. The authors propose the first succinct encoding of planar graphs that can be constructed in linear time, supports native BFS traversal, and requires only o(n) additional bits of space. This representation enables constant-time queries on both the BFS tree and the original graph structure. Furthermore, it extends to several advanced applications, including computing balanced separators of size O(√n), diameter-based tree decompositions, triangulation, and bipartiteness testing. By integrating techniques from succinct data structures, planar graph embedding theory, and dual-graph traversal, the method achieves sublinear space complexity without sacrificing linear-time efficiency.
0 citationsRead paper
Sim4IA-Bench: A User Simulation Benchmark Suite for Next Query and Utterance Prediction
Existing user simulation studies lack reproducible and interpretable evaluation benchmarks, particularly for next-query and next-utterance prediction tasks, due to the absence of authentic conversational context. To address this, we propose Sim4IA-Bench—the first benchmark for user behavior simulation in information retrieval—built upon 160 real search sessions (including a publicly available subset of 70). It establishes the first systematic alignment between authentic interaction trajectories and multi-turn simulated predictions. We introduce a dual-task evaluation framework: Task A (next-query prediction) and Task B (next-utterance prediction), incorporating novel metrics tailored to intent drift and query reformulation. The benchmark leverages high-fidelity simulation environments derived from CORE search engine data. Sim4IA-Bench provides a unified, reproducible platform for evaluating user simulators and significantly advances interaction-aware retrieval evaluation.
0 citationsRead paper
THM@SimpleText 2025 -- Task 1.1: Revisiting Text Simplification based on Complex Terms for Non-Experts
This study addresses the challenge of low comprehensibility of context-poor, ultra-short scientific sentences—such as abstracts and titles—for non-expert readers. We propose a term-level simplification method that first precisely identifies domain-specific complex terms within the input sentence, then employs coordinated lightweight Gemini and OpenAI models for controlled, semantics-preserving rephrasing—avoiding global rewriting to prevent information distortion. Our approach integrates rule-based terminology detection with large language models’ semantic fidelity in generation. Evaluated on CLEF SimpleText 2025 Task 1.1, it significantly improves readability and accuracy for low-context scientific sentences. Experiments demonstrate superiority over baselines across three dimensions: plausibility of term replacement, preservation of original meaning, and comprehension by non-experts. The method thus enhances public accessibility of scientific knowledge without compromising technical integrity.
0 citationsRead paper