SBIR Phase I: Feature Engineering Automation via Human Insight Integration and End to End Optimization

这个小企业创新研究（SBIR）第一阶段项目的更广泛影响是使具有最少数学和编程知识的个人能够学习和构建人工智能，并通过这一成功开辟了通往更大STEM参与和成功的道路。这种创新有助于在完整的生命周期框架内自动化人工智能生命周期的关键组件——特征工程（FE）。行业可以以更低的成本和现有员工更快地将人工智能应用于业务解决方案。通过关注企业数据，该技术可以实现广泛的企业使用和随后的经济效益。这个小企业创新研究（SBIR）第一阶段项目解决了商业人工智能解决方案的一个关键限制，即人类专家驱动和大部分手动特征工程（FE）阶段。该创新是一种在全生命周期环境中实现FE自动化的新方法，有三个不同之处：(a)通过自动生成和人类注释的知识图嵌入人类的洞察力，(b)将复杂的FE过程构建为一系列创新动态组合的子阶段，通过知识图洞察力进一步调整，(c)通过定量的端到端反馈进行迭代优化。该方法集成了专家直觉，可以通过知识图和反馈广泛选择潜在的有限元转换和可探索的状态空间。该研究通过对现实场景进行测试，并利用直接暴露现实FE问题的客户数据集，确保了翻译的可行性，解决了过去研究工作的重大局限性。本项目力求证明该方法在解决生产企业特征工程问题方面的可行性和效益。预期的结果是一个解决方案，可以很容易地商业化，自动化特征工程，用于广泛的商业用途，使用数字、分类、文本和时间序列数据。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project is to enable individuals with minimal mathematics and programming knowledge to learn and build AI, and through that success open a pathway to greater STEM engagement and success. This innovation helps by automating a key component of the AI lifecycle, Feature Engineering (FE), within a full lifecycle framework. Industries can apply AI to business solutions faster, at lower cost, and with existing employees. By focusing on enterprise data, this technology enables broad corporate use and subsequent economic benefit. This Small Business Innovation Research (SBIR) Phase I project addresses a key limitation of commercial AI solutions, the human expert driven and largely manual Feature Engineering (FE) stage. The innovation is a novel approach to FE automation within a full lifecycle context, with three differentiations: (a) embedding human insight via auto-generated and human annotated Knowledge Graphs, (b) structuring the complex FE process as a series of sub-stages for dynamic combination of innovations, further tuned via Knowledge Graph insight and (c) iterative optimization via quantitative end-to-end feedback. The approach integrates expert intuition, enabling a wide selection of potential FE transformations and explorable state space via the Knowledge Graph and feedback. The research ensures translational viability by testing against real-world scenarios and utilizing customer datasets that directly expose realistic FE problems, addressing a significant limitation of past research efforts. The project strives to prove the feasibility and benefits of this approach in tackling Feature Engineering problems in production enterprises. The anticipated result is a solution that can be easily commercialized to automate Feature Engineering for a wide range of business usages using numerical, categorical, text and time series data.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.