FAMOS is a fault diagnosis approach for microservice systems on multi-modal Data
Existing methods struggle to address the two challenges in multi-modal fault diagnosis: First, how to avoid information loss within each modality during the data fusion process? Second, how to accurately correlate data from different modalities and mine the relationships between them? We propose that due to the significant structure differences among the three modalities of data, a late-fusion-based approach can avoid the intrinsic information loss to each modality. However, it is crucial to consider the relationships between modalities during data processing and fusion. In this paper, we propose FAMOS, a novel fault diagnosis approach for microservices on muti-modal data. For the first challenge, FAMOS employs a late-fusion-based approach and designing specific feature extractors tailored to the structures of each modality, thereby preserving the intrinsic information of each modality. To address the second challenge, we introduce a new Gaussian-attention mechanism, which assigns weights based on the degree of relationships between modalities during data processing.
To begin FAMOS on Train-ticket dataset, run this code:
python main.py --config train-ticket.yamldgl==0.9.1
diskcache==5.6.3
einops==0.8.0
numpy==2.0.1
pandas==2.2.2
pytz==2024.1
PyYAML==6.0.1
PyYAML==6.0.1
scikit_learn==1.4.2
torch==2.3.0
torch_geometric==2.5.3
tqdm==4.66.2
Currently, we provide a processed version of the train-ticket dataset, which will be automatically downloaded during the execution of the code. Upon acceptance of our paper, we will make the original content of the two datasets mentioned in the paper publicly available.
Our configuration files are located in the ./config directory, and you can modify the contents there to perform experiments with different datasets and parameters.