Building data-driven decision-support systems to assist organizations operating in high-stakes domains requires fundamental advances in machine learning. Indeed, to be deployed, the tools created need to: 1) be interpretable, taking e.g., the form of decision trees or linear scoring rules; 2) be fair, being able to satisfy arbitrary domain specific requirements, and; 3) be robust to distribution shifts which arise when e.g., the data collection process changes. Importantly, they need to be as accurate as possible such that e.g., a maximum number of people at risk are identified.

Since joining USC, I am leading a research direction that integrates mixed-integer optimization (MIO) and robust optimization with data analytics to design provably optimal, robust, interpretable, and fair machine learning models suitable to deploy in high-stakes settings.

My research in this stream is supported in part by my NSF CAREER award, funds from Schmidt Futures, and my Coordinated Entry System Triage Tool Research & Refinement (CESTTRR) grant (which is funded by the Hilton Foundation and the United Way of Greater LA Home for Good).

Motivated by the popularity of decision trees in high-stakes domains, I have focused my efforts on models, algorithms, and tools for learning optimal decision trees. In work led by my PhD student Sina Aghaei, we leveraged the power of MIO to provide a strong formulation for the problem and exploited its structure to derive a tailored decomposition method [7]. These advances resulted in a 31-fold speed-up relative to the state-of-the-art and improved out-of-sample accuracy by up to 8%. Critically, our method does not rely on warm-starts, enabling us to incorporate arbitrary fairness, interpretability, and other domain constraints.

My lab has launched a new direction leveraging MIO to learn optimal decision trees incorporating fairness penalties [6] and constraints [11], in works led by my PhD student Sina Aghaei and my undergraduate student Nathan Jo, respectively. Notably, [11] can incorporate arbitrary machine learning fairness metrics and is the first completely versatile tool for building fair decision trees. For statistical parity, for which (heuristic) methods exist, our approach results in higher out-of-sample accuracy in 100% of cases. We have leveraged this tool to generate fair couterfactual outcome predictions under different housing assignments as part of our work with LAHSA [26]. We are using it in our collaboration with Prof. Fulginiti, Assistant Professor in the University of Denver School of Social Work, to predict risk of suicidal ideation.

Since 2019, I am the data science lead in an interdisciplinary effort to improve the system that connects homeless persons to services in LA [3]. My team and I are using data from the HMIS to improve the predictions made by the risk assessment tool based on a self-reported survey. The social science side of the team is modifying the data collection processes to make them more sensitive: they are changing the wording, length, and order of the questions and when, where, and how the survey is conducted. These changes will impact responses, resulting in a distribution shift to occur. To address this issue, we have devised the first optimal method for learning decision trees that are robust to distribution shifts [15]. We show an increase of up to 14.16% (resp. 4.72%) in worst-case (resp. average-case) accuracy from using our robust solution. This work is led by my PhD student, Nathan Justin, and his findings were a major factor in earning him the 2022 Viterbi Graduate Student Best RA Award.

Figure. Distribution across problem instances of the gain in worst-case (left) or average case (right) accuracy from using a robust tree versus a non-robust, regularized tree across different values of budget of uncertainty when the data perturbations are either expected or unexpected.

1. Facts About Suicide. URL: https://www.cdc.gov/suicide/facts/index.html.

2. Grand Challenges for Social Work. URL: https://grandchallengesforsocialwork.org.

3. LAHSA CES Triage Tool Research & Refinement webpage. URL: https://www.lahsa.org/ documents?id=4370-ces-triage-tool-research-refinement.pdf.

4. Landslides 101. URL: https://www.usgs.gov/programs/landslide-hazards/landslides-101.

5. Too Many Donor Kidneys Are Discarded in U.S. Before Transplantation. URL:https://www.pennmedicine.org/news/news-releases/2020/december/ too-many-donor-kidneys-are-discarded-in-us-before-transplantation.

6. S Aghaei, M J Azizi, and P Vayanos. Learning optimal and fair decision trees for non- discriminative decision-making.In Proceedings of the 33rd AAAI Conference on Artificial Intelligence, 2019.

7. S Aghaei, A Gómez, and P Vayanos. Strong optimal classification trees. Major revision at Operations Research, 2022. URL: https://arxiv.org/abs/2002.09142.

8. M J Azizi, P Vayanos, B Wilder, E Rice, and M Tambe. Designing fair, efficient, and in- terpretable policies for prioritizing homeless youth for housing resources. In Proceedings of the 15th International Conference on the Integration of Constraint Programming, Artificial Intelligence, and Operations Research, 2018.

9. C Bandi, N Trichakis, and P Vayanos. Robust multiclass queuing theory for wait time esti- mation in resource allocation systems. Management Science, 65(1):152–187, 2018.

10. Q Jin, A Georghiou, P Vayanos, and G Hanasusanto. Distributionally robust optimization with decision-dependentinformation discovery. In preparation for submission to INFORMS Journal on Computing, 2022.

11. N Jo, S Aghaei, J Benson, A Gómez, and P Vayanos. Learning optimal fair classication trees. Under review at second ACM conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO’22), 2022. URL: https://arxiv.org/pdf/2201.09932.pdf.

12. N Jo, S Aghaei, A Gómez, and P Vayanos. Learning optimal prescriptive trees from observa- tional data. Major Revision at Management Science, short version appeared at 2022 AAAI Workshop on AI and Behavior Change, 2021. URL:https://arxiv.org/pdf/2108.13628. pdf.

13. N Jo, B Tang, K Dullerud, S Aghaei, E Rice, and P Vayanos. Evaluating fairness of contextual resource allocationsystems: metrics and impossibility results. In preparation for submission to 37th AAAI Conference on Artificial Intelligence, 2022.

14. C Johnston, S Blessenhohl, and P Vayanos. Preference Elicitation and Aggregation to Aid with Patient Triage during the COVID-19 Pandemic. In preparation for submission to Operations Research; short version appeared in International Conference on Machine Learning (ICML) Workshop on Participatory Approaches to Machine Learning, 2020.

15. N Justin, S Aghaei, A G´omez, and P Vayanos. Optimal robust classification trees. In prepa- ration for submission to Operations Research; short version appeared in AAAI Workshop on Adversarial Machine Learning and Beyond, 2021.

16. A Rahmattalabi, P Vayanos, K Dullerud, and E Rice. Learning resource allocation policies from observational data with an application to homeless services delivery. In 2022 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’22), 2022.

17. A Rahmattalabi, P Vayanos, A Fulginiti, E Rice, B Wilder, A Yadav, and M Tambe. Exploring algorithmic fairness in robust graph covering problems. In Proceedings of the 33rd Conference on Neural Information Processing Systems (NeurIPS), 2019.

18. B Tang, C Koçyigit, and P Vayanos. Learning optimal policies for allocating housing to people experiencinghomelessness from data collected in deployment. In preparation for submission to Management Science, 2022.

19. United-Nations. Resolution adopted by the General Assembly on 6 July 2017, Work of the Statistical Commissionpertaining to the 2030 Agenda for Sustainable Development. 2017.

20. P Vayanos, A Georghiou, and H Yu. Robust optimization with decision-dependent information discovery. Major revisionat Management Science, 2019. URL: https://arxiv.org/pdf/ 2004.08490.pdf.

21. P Vayanos, Q Jin, and G Elissaios. ROC++: Robust Optimization in C++. Forthcoming at INFORMS Journal on Computing, 2022. URL: https://arxiv.org/pdf/2006.08741.pdf.

22. P Vayanos, Q Jin, and G Elissaios. ROCPP Version v2020.0140. INFORMS Journal on Computing, 2022.doi:10.5281/zenodo.6360996.

23. P Vayanos, D Kuhn, and B Rustem. Decision rules for information discovery in multi-stage stochastic programming. InProceedings of the 50th IEEE Conference on Decision and Control, pages 7368–7373, 2011.

24. P Vayanos, D Kuhn, and B Rustem. A constraint sampling approach for multi-stage robust optimization. Automatica, 48(3):459–471, 2012.

25. P Vayanos, W Wiesemann, and D Kuhn. Hedging electricity swing options in incomplete markets. In Proceedings of the 18th IFAC Wold Congress, pages 846–853, 2011.

26. P Vayanos, Y Ye, D McElfresh, J Dickerson, and E Rice. Robust active preference elicitation. Under second round ofreview at Management Science, 2021. URL: https://arxiv.org/pdf/ 2003.01899.pdf.

27. P Vossler, S Aghaei, N Justin, N Jo, A Gómez, and P Vayanos. ODTlearn package v0.1. URL: https://github.com/D3M-Research-Group/odtlearn.

28. P Vossler, S Aghaei, N Justin, N Jo, A Gómez, and P Vayanos. ODTlearn: a Python package for learning optimal decision trees. In preparation for submission to Journal of Machine Learning Research, 2022.

29. Y Ye, C Doehring, A Georghiou, P Vayanos, and H Robinson. Conserving biodiversity via ad- justable robust optimization. In preparation for submission to Management Science; short version appeared in Proc. of the 21st International Conference on Autonomous Agents and Mul- tiagent Systems (AAMAS 2022), Workshop on Autonomous Agents for Social Good (AASG), 2022.