KDSelector: A Framework of Knowledge-Enhanced and Data-Efficient Selector Learning for Anomaly Detection Model Selection in Time Series

Model selection has been raised as an essential problem in the area of time series anomaly detection (TSAD), because there is no single best TSAD model for highly heterogeneous time series in real-world applications. However, despite the success of existing model selection solutions, which usually learn (a.k.a. train) a classification model (especially neural network, NN) using historical data as a selector to predict the correct TSAD model for each time series to detect, the existing NN-based selector learning method cannot utilize the auxiliary knowledge in the historical data and requires iterating over all training samples, which limits the model selection ability and training speed of the selector. The latter data efficiency problem can be partially solved by existing data pruning methods designed for general NN training, but with suboptimal speedup or degraded selection ability due to disregarding intrinsic data properties in TSAD model selector training. To address these limitations, we propose KDSelector, to the best of our knowledge, the first framework customized for knowledge-enhanced and data-efficient learning of NN-based TSAD model selectors, of which we design three plug-and-play modules that are agnostic to NN architectures (e.g., ResNet and Transformer) and can be seamlessly integrated into the existing selector learning framework. Specifically, we propose two knowledge enhancement mechanisms to improve the selection ability of the selector with any architecture by integrating the auxiliary knowledge in a unified way. We further design a novel data pruning framework with theoretical guarantees to achieve state-of-the-art training acceleration for the NN-based selector with almost lossless selection ability. Extensive experiments demonstrate the superior performance of our proposals in terms of model selection ability and selector learning efficiency.