Developing interactive notebooks to support algorithm transparency in data-driven government

数据科学技术越来越多地应用于政府，以支持政策、资源和业务活动的目标。为了真正采用这种活动，应用技术的分析人员和在交流决策或制定政策建议时必须解释其输出的领域实践者之间需要密切合作。例如，为了生成理论上和生态上有效的数据驱动模型，在选择相关输入数据集、训练和评估不同模型规格以及向决策者传达模型输出时，需要特定主题的专门知识。该项目将与莱斯特郡议会的数据科学家和专家决策者目前在成人社会护理、青年服务和交通领域制定的一项工作方案相联系。该计划是开发、构建和评估支持笔记本的可视化工具，将政策制定者插入数据科学过程（参见Wattenberg等人，2019年的典型示例）：使用现代技术进行不确定性可视化，揭示模型背后的潜在机制，并促进对模型概率的直觉。该项目将产生透明且生态有效的数据驱动模型，并将解决学术界和众多政府白皮书中关于数据驱动政府中算法透明度主题的各种挑战。项目产出包括以技术为重点的学术论文，报告揭示和传播数据驱动产出的工具的设计和评价；以及以经验为重点的论文，详细介绍了通过使用这些工具生成的模型。为了建立更广泛的数据科学社区的参与和批评，该项目将附带一个github组织页面，为离散的项目提供单独的存储库。这些可能侧重于不同的领域——成人社会关怀、年轻人、犯罪和交通——或数据分析过程的不同阶段。例如，一些工具将针对模型构建—支持模型参数化和查询—其他工具将针对模型输出的通信—鲁棒性检查和不确定性通信。笔记本设计的通用方面将被打包到软件库中。通过这一点，我们的目标是通过令人信服的例子来展示一个工作流程，以实现有意义的数据驱动的政策制定。

Data Science techniques are increasingly being deployed in government to support the targeting of policy, resources and operational activities. For such activity to be genuinely adopted, close collaboration is needed between analysts applying techniques and domain practitioners who must interpret their outputs when communicating decisions or making policy recommendations. For example, in order to generate data-driven models that are theoretically and ecologically valid, subject-specific expertise is required in the selection of relevant input datasets, when training and evaluating different model specifications and when communicating the outputs of models to decision-makers.This project will connect with a programme of work currently being developed by data scientists and expert policy makers at Leicestershire County Council in the domains of Adult Social Care, Youth Services and Transport. The plan is to develop, build and evaluate notebook-enabled visualization tools that insert policy-makers into the data science process (see Wattenberg et al. 2019 for a characteristic example): exposing the underlying mechanisms behind models and promoting intuition around model probabilities using modern techniques for uncertainty visualization. The project will lead to data-driven models that are transparent and ecologically valid and will address challenges identified variously in academia and numerous government white papers on the theme of algorithm transparency in data-driven government.Project outputs include technically-focused academic papers reporting on the design and evaluation of tools for exposing and communicating data-driven outputs; and empirically-focussed papers detailing models generated through the use of such tools. In order to build engagement and critique from the wider data science community, the project will have an accompanying github organisation page, with separate repositories for discrete projects. These might focus on different domain areas -- Adult Social Care, Young People, Crime and Transport -- or different stages of data analysis process. For example, some tools will be targeted at model-building -- supporting model parameterisation and query - others at the communication of model outputs - robustness checks and uncertainty communication. Generalisable aspects of notebook design will be packaged into software libraries. Through this, the ambition is to demonstrate a workflow, with compelling examples, for effecting meaningful data-driven policy development.