Bidirectional Temporal Attention Networks for Missing Data Recovery in Urban Air Quality Monitoring

  This paper presented BiTAN, a Bidirectional Temporal Attention Network for missing data recovery in urban air quality time series. BiTAN combines variable-type-specific learnable missing embeddings with independent forward and backward self-attention encoder stacks fused through a gated cross-directional layer. Comprehensive evaluation on EPA AQS data from 12 Chicago monitoring stations under MCAR, MNAR-block, and MNAR-seasonal missing regimes demonstrates consistent 17–21% RMSE improvements over both recurrent and variational baselines, with increasing advantages at higher missingness rates. Ablation experiments confirm that bidirectionality, the gated fusion layer, learned missing embeddings, and positional encoding all contribute meaningfully to performance. BiTAN occupies a computationally efficient point on the accuracy-efficiency frontier: it is 1.3× faster than competing attention and variational baselines while achieving the best accuracy. Its single-station architecture makes it straightforward to deploy in operational monitoring pipelines and to retrain as sensor networks evolve. Future directions include integration with spatial graph attention for spatiotemporal modelling across station networks, adaptive temporal windowing for extreme outage scenarios, Bayesian uncertainty quantification to attach prediction intervals to imputed values and automatically flag high-uncertainty estimates for regulatory review, transfer learning across metropolitan areas to reduce data requirements for new deployments, and interpretability analysis of attention patterns to support faithful and auditable AI reasoning in compliance-critical monitoring applications [12].

1. V. Kataria and N. Kumar, "SemanticQuark: A Python-Native Framework for Unified Semantic Analytics," in IEEE Access, vol. 14, pp. 66337-66357, 2026, doi: 10.1109/ACCESS.2026.3688764 2. Hadeed, S.J., O'Rourke, M.K., Burgess, J.L., Harris, R.B., Canales, R.A.: Imputation methods for addressing missing data in short-term monitoring of air pollutants. Sci. Total Environ. 730, 139140 (2020). https://doi.org/10.1016/j.scitotenv.2020.139140 3. Che, Z., Purushotham, S., Cho, K., Sontag, D., Liu, Y.: Recurrent neural networks for multivariate time series with missing values. Sci. Rep. 8, 6085 (2018). https://doi.org/10.1038/s41598-018-24271-9 4. Du, W., Côté, D., Liu, Y.: SAITS: self-attention-based imputation for time series. Expert Syst. Appl. 219, 119619 (2023). https://doi.org/10.1016/j.eswa.2023.119619 5. Kim, T., Kim, J., Yang, W., Lee, H., Choo, J.: Missing value imputation of time-series air-quality data via deep neural networks. Int. J. Environ. Res. Public Health 18(22), 12213 (2021). https://doi.org/10.3390/ijerph182212213 6. Hua, V., Nguyen, T., Dao, M.S., Nguyen, H.D., Nguyen, B.T.: The impact of data imputation on air quality prediction problem. PLoS ONE 19(9), e0306303 (2024). https://doi.org/10.1371/journal.pone.0306303 7. Junninen, H., Niska, H., Tuppurainen, K., Ruuskanen, J., Kolehmainen, M.: Methods for imputation of missing values in air quality data sets. Atmos. Environ. 38(18), 2895–2907 (2004). https://doi.org/10.1016/j.atmosenv.2004.02.026 8. Moshenberg, S., Lerner, U., Fishbain, B.: Spectral methods for imputation of missing air quality data. Environ. Syst. Res. 4, 1–13 (2015). https://doi.org/10.1186/s40068-015-0052-z 9. Nazabal, A., Olmos, P.M., Ghahramani, Z., Valera, I.: Handling incomplete heterogeneous data using VAEs. Pattern Recogn. 107, 107501 (2020). https://doi.org/10.1016/j.patcog.2020.107501 10. Chen, Y., Wang, J., Liu, X.: Variational autoencoding with conditional iterative sampling for missing data imputation. Mathematics 12(20), 3288 (2024). https://doi.org/10.3390/math12203288