Gfm4Mtsad
A graph foundation model framework for multivariate time series anomaly detection that bridges continuous time series signals and the space of pretrained graph models through three modules, SICA, MAPS, and TGEC, enabling transferable reuse of topological priors and reliable window-level anomaly determination.
Timeline (last 90 days)
The Gfm4Mtsad framework attempts to use a pretrained graph foundation model (GFM) to address the problem that prior knowledge cannot be reused in multivariate time series anomaly detection.
The SICA (Structure-and-Information Contract Alignment) module consists of three parts: principal-component subspace token projection, normal-prior-guided graph construction, and residual recovery, and is used to bridge continuous time series and the pretrained GFM space.
The MAPS (Mechanism-Aligned Pseudo-Anomaly Screening) strategy imposes local dependency breakage while preserving empirical residuals, selecting hard negative samples near the normal boundary for targeted supervision.
The TGEC (Tail-Gated Evidence Calibration) mechanism calibrates multi-granularity graph responses with normal-operation references to refine extreme anomaly candidates in the tail distribution while preserving the ranking of benign scores.
The paper evaluates with two GFM backbones, and the results show that Gfm4Mtsad outperforms existing baselines in detection performance.
The material does not disclose the specific datasets, metric values, or a detailed list of comparison methods, so only qualitative conclusions can be given.
The Gfm4Mtsad pipeline corresponds, through the three steps of SICA (spatial alignment), MAPS (supervision signal quality), and TGEC (output aggregation), to the input, training, and output stages of the adaptation pipeline respectively.
All Facts (7)
The Gfm4Mtsad framework attempts to use a pretrained graph foundation model (GFM) to address the problem that prior knowledge cannot be reused in multivariate time series anomaly detection.
50%UnverifiedThe SICA (Structure-and-Information Contract Alignment) module consists of three parts: principal-component subspace token projection, normal-prior-guided graph construction, and residual recovery, and is used to bridge continuous time series and the pretrained GFM space.
50%UnverifiedThe MAPS (Mechanism-Aligned Pseudo-Anomaly Screening) strategy imposes local dependency breakage while preserving empirical residuals, selecting hard negative samples near the normal boundary for targeted supervision.
50%UnverifiedThe TGEC (Tail-Gated Evidence Calibration) mechanism calibrates multi-granularity graph responses with normal-operation references to refine extreme anomaly candidates in the tail distribution while preserving the ranking of benign scores.
50%UnverifiedThe paper evaluates with two GFM backbones, and the results show that Gfm4Mtsad outperforms existing baselines in detection performance.
50%UnverifiedThe material does not disclose the specific datasets, metric values, or a detailed list of comparison methods, so only qualitative conclusions can be given.
50%UnverifiedThe Gfm4Mtsad pipeline corresponds, through the three steps of SICA (spatial alignment), MAPS (supervision signal quality), and TGEC (output aggregation), to the input, training, and output stages of the adaptation pipeline respectively.
50%