Contrastive Masked Autoencoding for Self-Supervised Anomaly Representation Learning in Distributed Task Systems

Abstract

Addressing the challenge of representing anomalous states in cloud computing, microservice collaboration, and large-scale task scheduling environments, this paper proposes a distributed task anomaly representation learning method based on contrastive learning and masked autoencoders. Considering the characteristics of distributed task execution, such as multi-source heterogeneity, temporal coupling, complex dependencies, and hidden anomaly patterns, the method adopts a unified representation modeling approach, jointly designing the task context recovery process and semantic discrimination process to enhance the model's ability to capture deep features of anomalous states. Specifically, firstly, the multi-dimensional execution signals of the distributed task are temporally organized and mapped to a unified task representation; then, a random masking mechanism is introduced to mask some input information, and the missing content is recovered through collaborative learning between the encoder and decoder, thereby prompting the model to mine global contextual relationships and potential structural information in the task state; based on this, semantically consistent multi-view samples are further constructed, and contrastive constraints are used to strengthen the representational distinction between normal and anomalous states, so that the resulting latent space simultaneously possesses good completeness, robustness, and discriminability. To achieve effective synergy between the two learning objectives, this paper constructs a unified optimization framework that combines reconstruction and contrastive objectives, enabling anomaly representations to balance local detail recovery and global semantic separation. Experiments are conducted using publicly available open-source datasets and compared with related methods under consistent settings. Results show that the proposed method can learn higher-quality distributed task anomaly representations and exhibits better overall performance in anomaly recognition evaluations. This research demonstrates that combining mask modeling and contrastive learning for distributed task state representation construction can provide effective support for anomaly detection, fault analysis, and task state understanding in complex operating environments.