Computational statistics and deep learning to strengthen families and reduce violence towards children
Lead supervisor: Dr Seth Flaxman
Co-supervisors: Prof. Lucie Cluver, Dr Jamie Lachman
Commercial partner: IDEMS International Community Interest Company: Innovations in Development, Education and the Mathematical Sciences
Additional partner: Parenting for Lifelong Health
This studentship builds on a six-year-long collaboration between Oxford researchers and IDEMS International, including the recent successful randomized controlled field trial in Tanzania on a hybrid-digital parenting app to reduce violence towards children (Awah et al, 2021; Baerecke et al 2022).
The student will develop computational statistical and deep learning methods to harness the power of digital tools to strengthen families and protect children. A central pillar of this work is that interventions be designed for low- and medium-resource global health settings (e.g., cheap mobile phones without constant internet access) and co-created with in-country experts so that local knowledge is embedded from day one.
Three proposed projects will form chapters of the thesis:
1) machine learning to automatically classify records of parent/child interactions (audio, video, or text) and recommend tailored parenting support. On-device processing power and upload bandwidth are both limited, meaning that new computational approaches (such as knowledge distillation, lowering precision, and layer reuse) will be necessary for low and medium-resource settings.
2) using large language models (LLMs) to aid in the development stage of very large but deterministic offline parenting chatbots in multiple languages. Chatbot safety is of paramount importance, but we believe LLMs can play an important role in building out a very large set of questions and answers to be verified by experts before deployment.
3) beyond "A/B testing": causal inference to understand heterogeneous treatment effects of parenting app design and content choices during deployment in the real world. Machine learning has contributed to an explosion of new methods for causal inference, including doubly robust methods and causal forests (Athey et al, 2019). Further developing these new approaches in quasi-experimental settings will lead to major advances in the external generalizability of controlled trials.
Data availability: Data from the ParentApp randomised trial is already shared between the Universities of Oxford, Cape Town, Tanzania National Institute for Health Research, and IDEMS. This will be made available to the student and team working on this project. In addition, all users of the ParentApp program agree to the non-profit use of their data by the development and research team with the goal of improving the effectiveness of the program - again this will be available for the student and team to use.
Ethical approval is underway, as amendments to broader ethical approvals for the development and testing of hybrid-digital approaches to delivery of parenting programs.
Apply using course: DPhil in Computer Science
