MINOR: Multivariate Time Series Iterative Cleaning Algorithm

Errors are common in time series data, such as sensor measurements. Existing methods tend to focus on single errors in univariate data, but do not provide satisfactory results for consecutive errors, especially in the more general multivariate data. Modeling each dimension separately in one run can lead to bias, as the correlation between dimensions and the effects of existing errors are not taken into account. We also note that current methods suffer from the problem of over-repair, i.e. clean observations may be modified after data cleaning. In this paper, we propose MINOR, an iterative cleaning algorithm for multivariate time series with limited labels. We formalize the repair problem and propose a bidirectional validation that uses a local speed constraint and the past and future information to solve the over-repair problem. We also present a unidirectional validation that supports online computations. We analyze the properties of our proposals and compare them with SOTA methods in terms of effectiveness, efficiency, and the impact of applications such as classification. Experiments on real datasets show that MINOR can have higher repair accuracy in various situations and improve time efficiency over the proposed incremental techniques. Interestingly, it can be effective even when there are few labels around 1%.