# Report: ML-based model for MTSAD is impossible.

## New:
1. Based on the results, the proposed model did not perform as well as the reliable model, which is the GAT.

2. Since several model of Machine Learning (ML) were used to find the anomaly, combining the result of multiple ML model is the important thing to perform a final result which is really difficult to find a rule to combine them in each scenario.

3. The threshold of each ML model, as well as the rules for preprocessing and postprocessing, are a crucial factor in creating the anomaly detection model. These can be set by analyzing the relationship between historical data and new batch data. However, finding this relationship can be challenging, and there is a low probability of discovering it. Furthermore, there may be hundreds or even more rules that need to be implemented in the preprocessing and postprocessing stages to develop a good model.

## Jessica:
### Introduction
In order to meet the demand for a fast and accurate cloud resource usage monitoring system, we have adopted a statistical and machine learning-based approach to address the task of multivariate time series anomaly detection. However, it is worth noting that currently available unsupervised machine learning algorithms are less stable and detail-oriented compared to their deep learning model counterparts. Therefore, to filter out false positives and false negatives, we need to establish several rule-based systems. Following an analysis of the historical and batch data provided, we have come to the conclusion that creating a multivariate time series anomaly detection model using only machine learning is not feasible.
### Reasons
#### 1. No significant relation between historical and batch data:
For some machines, there is no significant relation found between the given historical and batch data. The historical data of the machines may look very similar to one another, while the batch data in one machine itself varies greatly. One possible reason for this is the use of optimizers in the company's processes. Optimizers can change the machine's behavior, which can result in variations in the batch data while in the long-term conforming to the same pattern. As a result, it is challenging to develop a machine learning-based model that can detect anomalies in such data. The model may fail to capture the changes caused by optimizers, resulting in a high false-positive rate or false-negative rate.

#### 2. High variability in batch data:
Depending on the pattern of each machine, the definition of anomaly changes. Although it is investigated that data from the same machine share a non-linear combination of features that make them distinct from one another, the high variability in batch data makes it impossible to define a generalized rule-based system to perform preprocessing and postprocessing. This variability can make it challenging to identify the relevant features for anomaly detection, which can result in a model that is not accurate enough to be used in practice.

### Conclusion
In conclusion, developing a robust machine learning-based model for multivariate time series anomaly detection is not always possible. When there is no significant relation between historical and batch data, or there is high variability in batch data, it can be challenging to develop an accurate model. In such cases, other techniques such as rule-based systems or expert systems may be more suitable for anomaly detection. However, it is essential to keep in mind that even these methods may not always be successful and may require constant improvement and adaptation to the data.

## Amanda:
When it comes to Multivariate Time Series Abnormality Detection, machine learning methods such as ARIMA and Local Outlier Factor have been commonly used. However, these methods often require manual feature engineering and model architecture design, and are not able to adaptively capture complex dependencies among time series. In contrast, Graph Attention Network (GAT) has shown superior performance in this task by adaptively capturing the dependencies among multiple time series and handling non-linear problems.

The main reason why machine learning methods are less effective than GAT in Multivariate Time Series Abnormality Detection is because they have limited ability to capture the complex interdependencies among multiple time series, and are often based on linear assumptions that may not hold true in real-world scenarios. Additionally, feature engineering and model architecture design require domain knowledge and may not generalize well to other datasets or scenarios.

Furthermore, GAT's ability to adaptively capture complex interdependencies among multiple time series and handle non-linear problems make it a more suitable approach for Multivariate Time Series Abnormality Detection tasks. Therefore, machine learning methods may not be as effective in this task as GAT due to their limitations in capturing the complex dependencies and handling non-linear problems.