ACM SIGKDD (International Conference on Knowledge Discovery and Data Mining), abbreviated as KDD, is a top international academic conference in the field of data mining. It is organized by the Association for Computing Machinery (ACM) Special Interest Group on Knowledge Discovery and Data Mining (SIGKDD) and is recommended as a Class A international academic conference by the China Computer Federation (CCF). Since 1995, KDD has been successfully held for 26 consecutive years, and the KDD Cup, which was created in response to the conference, has become the most influential event in the field of data mining.

From August 14th to 18th, KDD 2021 was held in Singapore. DeepBlueAI (DBAI), a team from Deep Blue Technology, actively participated in this year's event and became the center of attention by achieving outstanding results, winning the championship in the Multi-dataset Time Series Anomaly Detection competition. It is worth mentioning that the top seven teams also included teams from well-known companies and universities such as Huawei Noah's Ark Lab, Alibaba DAMO Academy, Hikvision, Hitachi, Mitsubishi Electric, the National Institute of Advanced Industrial Science and Technology (AIST) of Japan, and Humboldt University of Berlin. The competition was highly competitive, attracting over 500 teams to participate actively and receiving nearly 2000 valid submissions.

Multi-dataset Time Series Anomaly Detection

The Multi-dataset Time Series Anomaly Detection competition provided 250 time series data, each of which included an anomaly point. The organizers hoped that participants would use unsupervised or self-supervised methods to identify the locations of these anomaly points.

Time series anomaly detection aims to detect unexpected or rare events in data. It is commonly used in many industrial applications, such as operations and maintenance, industry monitoring, and online monitoring of product prices.

Data Analysis

The types of anomaly points in these time series data are diverse and can be individual anomalies or group anomalies, as shown in the figure below. It is challenging for a single method to find anomaly points in all the data, so an applicable framework or a good ensemble method is needed.

For this competition, the DeepBlueAI team developed a highly generalized and flexible anomaly detection framework called TsaDetect. Specifically, each time series goes through three modules: periodicity analysis, multi-model prediction, and evaluation and integration.

Conclusion

In 2019, the DeepBlue team achieved first place in the KDD Cup 2019 AutoML Track Challenge, and their victory in this year's competition demonstrates their commitment to "technology-driven, in-depth basic research" and their leading position in the field of data mining. Their independently developed time series anomaly detection framework also provides new ideas for the practical implementation of AI operations and AI time series monitoring.