Machine Learning and Planning Research: How Each Can Push the Other’s Frontiers

By Kshitiz Khanal

Planning and social science research communities are increasingly adopting machine learning techniques in their research. Machine learning (ML) represents a broad range of techniques that uses insights gained from data for prediction and other tasks as opposed to hard-coded rules. Even quantitative planning and social science researchers are still catching up to the (mostly technological) developments in computer science and business applications.

Here I discuss some reasons why planning and other social science domains are lagging behind technological developments in computer science and applications of those developments in businesses, how each domain can help push the boundaries of the other and some possible future actions that emerge from those discussions.

Catching up to the sprawl of techniques

Technological developments are happening at a myriad of frontiers in the field of machine learning, so much so that it’s hard to even for domain specialists to keep up. People pushing these frontiers are mostly working in big technology companies and universities from select regions of the world. 

Compared to the expansion of machine learning science mostly by computer scientists and business applications of those developments by the likes of big technology companies, the use of machine learning in social science domains such as planning research remains low, albeit growing. Following are some factors that hinder the use of machine learning for planning research:

  • Funding: Computer science and AI related domains are some of the best funded research domains. Similarly, companies (not only primarily technology companies) invest in ML/AI resources because of the potential return on investments. Comparatively, research funding in planning and social science domains as well as the capacity of local governments and nonprofits to carry out research is lower. 
  • Skills gap: Planning and social science researchers are not typically trained in machine learning. Although the realization of the utility of machine learning and adoption is increasing, the gap in skills among most planning researchers looking to use machine learning is a challenge.
  • Datasets: There are many well-known benchmark datasets for AI/ML research. Benchmark datasets are popular datasets on which the performance of new machine learning models are tested for standardized comparison with other models. Similarly, businesses generate datasets as part of their operations. There are limited datasets amenable for planning research in comparison. Data generation is a resource-intensive task, and it is not surprising that the amount of datasets available is lower where the allocation of resources is lower.

How ML can help planning research beyond predictions

With the broad variety of techniques available for prediction, causal analysis, data generation, and other tasks, it is more about how not if machine learning is useful in planning research. Let’s look at a few interesting applications.

  • Making sense of non-traditional data sources: Making sense of a lot of data sources that can be useful such as newspaper archives, social media, satellite images, online forums, and listservs can be cumbersome with traditional approaches. Using machine learning techniques such as image segmentation, optical character recognition, natural language processing, etc. can help gain insights from a large volume of data.
  • Causal reasoning: The emerging field of causal machine learning can be used in evaluating policies, creating better programs by targeting heterogeneous effects, and gathering insights from natural experiments that are not practical or possible from traditional social science research designs [1].
  • Creating synthetic datasets: ML models such as Generative Adversarial Networks (algorithms that can generate data such as images and texts strikingly similar to provided examples) can be used to create synthetic data that can help reduce bias in unbalanced datasets[2].
  • Theory building: Machine Learning can also guide theory building in planning and the social sciences. Theory building includes extensively testing the robustness of hypotheses. The suite of machine learning tools and developments in ML based causal reasoning can help guide theory building by uncovering novel and robust patterns in data [3].

How planning research can help ML

The field of AI/ML draws frequent criticism (deservedly) about the associated ethical and social justice issues. Planning research can help push those frontiers of machine learning and some more. Some of them are discussed below.

  • Exploration of potential ethical and social justice issues: The field of planning has forever been concerned with ethics and social justice. Planning scholars can help explore potential ethical and social justice related harms and biases from AI/ML [4].
  • Expanding social applications of machine learning: There is much to be gained by applying machine learning beyond building new machine learning architectures and improving the profitability of technology and other businesses. The techniques can be used in research where insights gained about improving people’s lives through planning are more straightforward.
  • Explaining the explanations of the black box of machine learning: AI/ML practitioners are pushing towards more explainability in terms of how the models came up with their predictions or outputs. Knowledge of social systems for which machine learning was used can help put those outputs in context [5].
  • Moving beyond benchmark datasets: Many machine learning researchers being concerned only with performance on benchmark datasets is a common criticism of machine learning models. Applications to planning problems can help the domain of machine learning move towards goals that are more directly beneficial to humanity.

The way forward

The application of machine learning techniques to planning problems can advance both of these fields. Increased collaboration, increased access to funding, increased institutional support, creation of learning materials, incentives for cross-disciplinary research projects and publications, push for more open data, etc. can help the two domains tango for increased social good.


References

[1] S. Athey and G. W. Imbens, “Machine learning methods for estimating heterogeneous causal effects,” stat, vol. 1050, no. 5, pp. 1–26, 2015.

[2] M. Hittmeir, A. Ekelhart, and R. Mayer, “On the utility of synthetic data: An empirical evaluation on machine learning tasks,” in Proceedings of the 14th International Conference on Availability, Reliability and Security, 2019, pp. 1–6.

[3] P. Choudhury, R. Allen, and M. Endres, “Developing theory using machine learning methods,” Available SSRN 3251077, 2018.

[4] A. Hagerty and I. Rubinov, “Global AI ethics: a review of the social impacts and ethical implications of artificial intelligence,” ArXiv Prepr. ArXiv190707892, 2019.[5] U. Bhatt, M. Andrus, A. Weller, and A. Xiang, “Machine learning explainability for external stakeholders,” ArXiv Prepr. ArXiv200705408, 2020.


Kshitiz Khanal is a PhD candidate at the Department of City and Regional Planning. His current research focuses on the application of emerging machine learning techniques in energy planning. He studied engineering and energy planning. Before coming to UNC, he co-founded an open technology advocacy non-profit in Nepal and was involved in energy as well as open data for development research. He enjoys playing and watching football (soccer), calligraphy, and sipping the Himalayan silver tips tea.


Edited by Jo Kwon, Managing Editor

Featured image courtesy of Cyberpunk style AI generated image using text submission “urban planning”