Research

I work on machine learning methods that stay reliable when the data change, and on taking these methods out of the lab and into real applications. My main research themes are listed below, with a few representative papers for each. The full list is on my Publications page.

Quantification and learning under distribution shift

Many applications need to know how many items belong to each class, not which class each item belongs to: the proportion of infected mosquitoes in a trap, of malicious devices on a network, or of misleading posts on a platform. This task is called quantification. It is closely related to label shift, where class proportions change between training and deployment. My group develops quantification methods, ways to detect and adapt to label shift, and ways to evaluate them.

• Donyavi, Li, Zhang, Silva and Batista. Match: A maximum-likelihood approach for classification under label shift. KDD 2025.
• Donyavi, Serapião and Batista. MC-SQ and MC-MQ: Ensembles for multi-class quantification. IEEE TKDE, 2024.
• Li, Gharakheili and Batista. Quantification over time. ECML-PKDD 2024.
• Serapião, Donyavi and Batista. MetaQuant: A framework for metaheuristic based quantification. Applied Soft Computing, 2025.

Time series and data streams

I have worked for many years on time series classification, similarity search and data stream mining, including learning when labels arrive late or never, and when the data distribution drifts over time.

• Rakthanmanon, Campana, Mueen, Batista, Westover, Zhu, Zakaria and Keogh. Searching and mining trillions of time series subsequences under dynamic time warping. KDD 2012 (Best Paper Award and 2022 Test of Time Award).
• Batista, Keogh, Tataw and Souza. CID: An efficient complexity-invariant distance for time series. Data Mining and Knowledge Discovery, 2014.
• Souza, Reis, Maletzke and Batista. Challenges in benchmarking stream learning algorithms with real-world data. Data Mining and Knowledge Discovery, 2020.
• Zhang, Batista and Kanhere. Mosaic: An accurate and efficient kernel-based multivariate time series classifier. PAKDD 2026.

Machine learning for IoT, networks and security

With colleagues in networking and cybersecurity, I develop methods to identify IoT devices and classify network traffic, keep these models accurate as device behaviour changes, and make them robust to attacks.

• Babaria, Lyu, Batista and Sivaraman. FastFlow: Early yet robust network flow classification using the minimal number of time-series packets. Proc. ACM on Measurement and Analysis of Computing Systems, 2025.
• Pashamokhtari, Okui, Nakahara, Kubota, Batista and Gharakheili. Dynamic inference from IoT traffic flows under concept drifts in residential ISP networks. IEEE Internet of Things Journal, 2023.
• Azizi et al. Towards label shift adaptation for robust IoT device identification. CPS&IoT Security and Privacy Workshop, 2025.

Machine learning on small devices: insects and satellites

I developed optical sensors that identify flying insect species from their wingbeats, used to monitor disease-carrying mosquitoes, with funding from USAID (Zika) and the Innovative Vector Control Consortium (malaria). More recently, my group has designed object detection models efficient enough to run on board nanosatellites.

• Chen, Why, Batista, Mafra-Neto and Keogh. Flying insect classification with inexpensive sensors. Journal of Insect Behavior, 2014.
• da Silva, Souza and Batista. An open-source tool for classification models in resource-constrained hardware. IEEE Sensors Journal, 2021.
• da Silva and Batista. ContinualCropBank: Object-level replay for semi-supervised online continual object detection. PAKDD 2026.

Data pre-processing and evaluation

My early work studied missing data imputation, class imbalance and model evaluation. Some of these papers are still widely used.

• Batista, Prati and Monard. A study of the behavior of several methods for balancing machine learning training data. ACM SIGKDD Explorations, 2004.
• Batista and Monard. An analysis of four missing data treatment methods for supervised learning. Applied Artificial Intelligence, 2003.

New directions

With colleagues in media and communication at UNSW and PUC Chile, I am starting work on multimodal AI for political communication: methods that combine video, audio, speech and platform metadata to study polarisation, and that generalise across countries and languages. See the advertised PhD scholarship.