Improving Data Imputation Through a Tuned Strategy for Dependency Discovery

Data Imputation approaches aim at improving the quality of data trying to infer values often missing into data. Among others, dependency-based imputation approaches exploit data relationships among attributes, such as Relaxed Functional Dependencies relaxing on the attribute comparison $\left(\text{RFD}_{c} \mathrm{s}\right)$, to provide semantically coherent and less biased imputations by exploiting similar candidate tuples. However, according to the possibility of considering variable similarity threshold combinations, existing RFD $_{c}$ discovery algorithms can limit imputation quality or become impractical for big datasets. In this paper, we propose triard, a tuned strategy for dependency discovery that iteratively adjusts similarity thresholds based on imputation performance, thus selecting the most effective $\mathbf{R F D}_{c} \mathbf{s}$ for selecting tuple candidates to impute missing values. Experimental results on 20 real-world datasets show the improvement in both imputation performance and execution time with respect to the $\mathbf{R F D}_{c} \mathbf{s}$ discovered by the algorithm domino. Moreover, triard resulted in a higher imputation performance, mainly in terms of precision, when compared with other Data Imputation approaches.