Wednesday 22 January 2025
A team of researchers has developed a new approach to detecting anomalies in time series data, which is crucial for identifying unusual patterns in everything from financial transactions to medical records. The method, called NRdetector, uses a combination of machine learning and temporal embedding to identify points in the data that are significantly different from the rest.
The problem with traditional anomaly detection methods is that they often rely on labeled data, but in many cases, such data may not be available or may be noisy. NRdetector addresses this issue by using a confidence-based sample selection technique, which allows it to select the most reliable positive and negative samples for training.
Once the model has been trained, it uses a temporal embedding module to capture the underlying patterns in the data. This is done by feeding the data into a neural network that learns to represent each point in the time series as a fixed-length vector. This process allows the model to identify anomalies that may not be immediately apparent from looking at the raw data.
The NRdetector was tested on three different datasets, including EMG, SMD, and PSM, and outperformed other state-of-the-art methods in all three metrics: F1-score for point-level prediction, F1-score for segment-level prediction, and precision@k for top-k predictions. The results show that the NRdetector is able to effectively identify anomalies even when the label noise rate is as high as 0.6.
One of the key advantages of the NRdetector is its ability to handle noisy labels, which are common in real-world datasets. By using a PU loss function and a confidence-based sample selection technique, the model is able to learn from the available data and make accurate predictions even when the labels are noisy.
The NRdetector has many potential applications, including fraud detection, medical diagnosis, and quality control. In these cases, being able to accurately identify anomalies can be critical for making informed decisions or taking corrective action.
Overall, the NRdetector is an important contribution to the field of anomaly detection, and its ability to handle noisy labels makes it a valuable tool for real-world applications.
Cite this article: “Accurate Anomaly Detection with NRdetector: A Machine Learning Approach”, The Science Archive, 2025.
Time Series Data, Anomaly Detection, Machine Learning, Temporal Embedding, Neural Network, Confidence-Based Sample Selection, Noisy Labels, Pu Loss Function, F1-Score, Precision@K
