To Impute or Not: Recommendations for Multibiometric Fusion

Combining match scores in multibiometrics via fusion is a well-established approach to improving recognition performances. However, when scores are missing, this can degrade performance, as well as limit the possible fusion techniques that can be applied. Imputation is a technique for estimating reasonable values to replace missing data and an approach that has previously shown promise in addressing missing scores in multibiometrics. In this paper, we evaluate various approaches to imputation methods on three multimodal biometric score datasets: NIST BSSR1, BIOCOP2008, and MIT LL TRIMODAL, and investigate the factors which might influence the effectiveness of imputation. Our studies reveal three key observations: (1) Imputation is preferable to not imputing missing scores, even when the fusion rule does not necessitate complete score data. (2) Balancing the classes in the training data is crucial to mitigate biases in the imputation technique and prevent favoritism towards the overrepresented class, even if it involves dropping a substantial number of score vectors. (3) Multivariate imputation approaches exhibit better estimation for genuine scores, while univariate imputation approaches yield stronger results for imputed imposter scores.