An adaptive ensemble imputer for lightweight streaming imputation in resource constrained environments

Handling of missing values in a resource-constrained settings is more challenging when dealing with streaming because traditional methods (KNN, MICE) are computationally expensive online, while basic approaches (mean imputation) miss temporal dynamics and concept drift. A lightweight per-feature streaming ensemble method Adaptive Ensemble Imputer is proposed combining simple candidate imputers–exponential moving average (EMA), running mean, last observed value, short-window median. The weights are adapted online via occasional masked validation. The method maintains O(d) per-sample complexity and O(kd) memory (four candidates per feature), enabling real-time deployment on constrained settings. Evaluated on five benchmark data streams (Electricity, Poker, Covtype, MovingRBF, Rotating Hyperplane) under controlled MNAR mechanisms with repeated stratified runs, the method achieves \(14.7× \) speedup over KNN, competitive accuracy with mean imputation on temporal data, and 37.5% faster downstream drift detection latency. Performance is mechanism-dependent: temporal-shift MNAR shows \(-\,53\) % MAE improvement (attributed to EMA emphasis when anomalies coincide with missingness), while level-dependent MNAR shows \(+262\) % degradation (systematic selection bias not corrected by adaptive weighting). The method does not provide theoretical MNAR unbiasedness guarantees; practitioners should use inverse-probability weighting for applications requiring MNAR correction. The primary contribution is computationally efficient, drift-responsive imputation suitable for streaming scenarios with temporal autocorrelation and concept drift.