SHF: Small: S3: Statistical and Structural Analysis for Spreadsheets

电子表格是世界上最常用的编程环境；仅微软Excel的用户就超过7.5亿。电子表格广泛应用于政府、科学和金融领域；超过95%的美国公司将其用于财务报告，85%的公司将其用于预算和预测。不幸的是，错误是电子表格特有的；最近的一项研究发现错误率超过95%。电子表格错误造成了灾难性的后果，导致了数十亿美元的损失。该项目使用专门为电子表格设计的自动分析技术来(a)自动检测并帮助防止电子表格中的错误，大大提高其计算的可靠性，(b)减少严重错误的风险，以及(c)潜在地节省数百万美元（如果不是数十亿美元的话）。该项目为电子表格（S3）开发统计和结构分析。电子表格具有不同于标准编程语言的独特特性，因此需要利用其特性的新程序分析。S3对电子表格公式的空间和深层结构进行统计分析，以识别高度异常的电子表格单元格，从而可能是错误的。S3通过结合空间和结构信息（包括依赖关系模式）的新颖向量表示，将查找数据和公式错误的问题减少到查找异常结构的问题。对这些向量进行统计分析，可以识别出在任何维度上都非常不寻常的公式，因此可能是错误的。S3既可以在单个电子表格级别操作，也可以从大量现有电子表格中学习电子表格使用模型，以调整分析并进一步降低误报率。

Spreadsheets are the most commonly used programming environment in the world; there are more than 750 million users of Microsoft Excel alone. Spreadsheets are widely used in government, scientific, and financial settings; over 95% of US firms use them for financial reporting and 85% use them for budgeting and forecasting. Unfortunately, errors are endemic to spreadsheets; a recent study found an error rate of over 95%. Spreadsheet errors have had catastrophic consequences, leading to losses of billions of dollars. This project uses automatic analysis techniques designed specifically for spreadsheets to (a) automatically detect and help prevent errors in spreadsheets, dramatically increasing the reliability of their calculations, (b) reduce the risks of serious mistakes, and (c) potentially save the economy millions if not billions of dollars.This project develops statistical and structural analyses for spreadsheets (S3). Spreadsheets have unique features that make them different from standard programming languages, and thus demand new program analyses that exploit their characteristics. S3 employs statistical analyses over the spatial and deep structure of spreadsheet formulas to identify spreadsheet cells that are highly anomalous and thus likely to be wrong. S3 reduces the problem of finding data and formula errors to that of finding anomalous structures via a novel vector representation that combines spatial and structural information (including patterns of dependencies). Applying statistical analyses across these vectors can then identify formulas that are highly unusual in any dimension, and thus likely to be wrong. S3 operates both at an individual spreadsheet level and also incorporates learned models of spreadsheet usage from large bodies of existing spreadsheets to condition the analysis and further reduce false positive rates.