首页 > AI前沿 > TabDAG: Structured Evolutionary Search for Tabular Feature Engineering

TabDAG: Structured Evolutionary Search for Tabular Feature Engineering

arXiv机器学习 2026-11-17 22:14 5 阅读 查看原文
用户现在需要把给定的英文新闻正文排版成语义化的HTML片段,首先得看内容,是一段一段的?不对,首先看内容,首先开头的段落,然后有没有需要分小节的?不对,先看内容,整个正文其实是一段?不对,不对,再仔细看,哦,不对,用户给的正文是一整段?不对,不对,再看,哦,不对,首先,规则里说每个自然段落用p包裹,长文用h2/h3划分,这里的内容是介绍TabDAG的,首先,首先第一句?不对,不对,先理清楚内容的结构: 首先,第一句是“Feature engineering can improve tabular prediction, but useful transformations are compositional, data-dependent, and subject to leakage constraints.” 不对,不对,整个正文的逻辑是: 1. 先讲特征工程的作用和存在的问题,然后引出TabDAG框架。 2. 然后讲TabDAG的图表示:节点、边,支持的功能。 3. 然后讲变异操作的约束和评估检查。 4. 然后讲实验结果:结构无效的程序情况,跨数据集和模型的性能提升,排名情况。 5. 然后讲LLM搜索的设置,消融实验的内容和结果。 6. 最后讲结果的可分析性。 不对,但用户要求保留所有信息,不得增删改,所以不能随便加小标题?哦,不对,规则里说“长文用

/

划分小节”,那如果原文没有小标题的话,是不是不需要硬加?不对,不对,再看规则,规则说“保留原文所有信息,不得增删改内容”,所以不能自己加标题内容,因为标题也是内容的一部分?哦对,不能增内容,所以如果没有现成的小节标题,就不需要加h2/h3,对吧?因为加h2的话需要有文字,但是原文没有对应的小标题文字,加了就属于新增内容了,违反规则4。 那接下来,首先,每个自然段落用p包裹,那原文的自然段落是怎么分的?哦,用户给的正文是一整段?不对,再看用户给的正文: 正文如下: Feature engineering can improve tabular prediction, but useful transformations are compositional, data-dependent, and subject to leakage constraints. We present TabDAG, a framework that combines evolutionary search with a graph representation of feature-engineering programs. Nodes implement row-wise or data-fitted transformations; edges encode dependencies among raw variables, intermediate values, and final features. The representation supports interactions, statistical encodings, clustering- and neighborhood-based features, and multi-stage compositions in a single executable program. Mutations are structured edits constrained by a parent-specific schema and checked before evaluation. In our experiments, no structurally invalid program reached evaluation, although proposals could be rejected and valid graphs could still fail at runtime. Across ten regression and classification datasets and three tabular model families, evolved features improve predictive performance over raw inputs. TabDAG and its combination with OpenFE achieve the two best mean ranks among the compared methods; pairwise rank differences between feature-engineering methods do not survive multiple-testing correction. The LLM-based searches use the same backbone and candidate-evaluation budget. A controlled ablation compares free-form Python, whole-graph JSON, and constrained graph edits with the search engine, budget, and backbone held fixed. Constrained edits yield fewer failed evaluations than either rewrite baseline. The resulting graphs expose the transformations and dependencies selected by search, enabling feature-level analysis of the observed gains. 哦,原来用户给的是一整段的正文?那不对,是不是要按语义拆分段落?不对,不对,规则说“每个自然段落用

包裹”,原文的自然段落就是这一段?但不对,可能用户给的是连续的正文,那是不是可以按语义拆分?不对,不对,规则4说“保留原文所有信息,不得增删改内容”,拆分段落算不算改内容?哦,不对,排版的话,自然段落是指原文的段落划分,如果原文是一整段,那是不是就用一个p包裹?不对,但再看,有没有需要用strong的地方?比如TabDAG,是不是可以加strong?对,重要句子或者重要术语可以用strong,规则3说重要句子可用,那TabDAG是核心框架,可以用strong。 等下,再仔细看规则,3里的重要句子可用strong,那关键的术语比如TabDAG可以加吗?对,因为是重要内容。然后,有没有列表?没有,有没有引用?没有,有没有代码?没有,所以不需要那些标签。 那现在,首先,原文是一整段,那用一个p包裹?不对,不对,会不会我理解错了,原文是不是可以按语义拆分成几个自然段落?比如: 第一部分:Feature engineering can improve tabular prediction, but useful transformations are compositional, data-dependent, and subject to leakage constraints. We present TabDAG, a framework that combines evolutionary search with a graph representation of feature-engineering programs. 第二部分:Nodes implement row-wise or data-fitted transformations; edges encode dependencies among raw variables, intermediate values, and final features. The representation supports interactions, statistical encodings, clustering- and neighborhood-based features, and multi-stage compositions in a single executable program. 第三部分:Mutations are structured edits constrained by a parent-specific schema and checked before evaluation. In our experiments, no structurally invalid program reached evaluation, although proposals could be rejected and valid graphs could still fail at runtime. 第四部分:Across ten regression and classification datasets and three tabular model families, evolved features improve predictive performance over raw inputs. TabDAG and its combination with OpenFE achieve the two best mean ranks among the compared methods; pairwise rank differences between feature-engineering methods do not survive multiple-testing correction. 第五部分:The LLM-based searches use the same backbone and candidate-evaluation budget. A controlled ablation compares free-form Python, whole-graph JSON, and constrained graph edits with the search engine, budget, and backbone held fixed. Constrained edits yield fewer failed evaluations than either rewrite baseline. 第六部分:The resulting graphs expose the transformations and dependencies selected by search, enabling feature-level analysis of the observed gains. 不对,但这样拆分的话,算不算改内容?哦,排版的段落拆分是允许的吧?因为规则说“每个自然段落用

包裹”,这里的自然段落应该是指语义上的自然段,原文因为是纯文本,可能没有分段,但语义上可以拆分?不对,不对,再看规则4,“保留原文所有信息,不得增删改内容”,拆分段落不增删改内容,只是排版,没问题。不过要不要?或者是不是原文就是一整段,就用一个p?不对,再想想,用户说“纯文本新闻正文”,新闻正文的段落是自然分隔的,可能用户粘贴的时候把段落合并了?不对,看原文的内容,逻辑上是几个部分: 1. 背景+提出框架 2. 框架的表示 3. 变异操作与评估前置检查 4. 实验结果(性能、排名) 5. LLM搜索设置与消融实验 6. 结果的分析价值 这样拆分是合理的,而且不改变内容。然后,重要的术语比如TabDAG用strong,还有“constrained graph edits”?不对,或者重要的结论?比如“evolved features improve predictive performance over raw inputs”可以用strong?还有“Constrained edits yield fewer failed evaluations than either rewrite baseline”是重要结论,也可以用strong? 等下,再仔细看,不能增删改内容,所以只是加标签,文字本身不变。那现在组织HTML: 首先,有没有需要h2或者h3?原文没有小标题,所以不能加,因为加h2的话需要有文本,原文没有,所以不用,不然就新增内容了,违反规则。所以直接用p包裹每个语义段落。 现在写出来:

Feature engineering can improve tabular prediction, but useful transformations are compositional, data-dependent, and subject to leakage constraints. We present TabDAG, a framework that combines evolutionary search with a graph representation of feature-engineering programs.

Nodes implement row-wise or data-fitted transformations; edges encode dependencies among raw variables, intermediate values, and final features. The representation supports interactions, statistical encodings, clustering- and neighborhood-based features, and multi-stage compositions in a single executable program.

Mutations are structured edits constrained by a parent-specific schema and checked before evaluation. In our experiments, no structurally invalid program reached evaluation, although proposals could be rejected and valid graphs could still fail at runtime.

Across ten regression and classification datasets and three tabular model families, evolved features improve predictive performance over raw inputs. TabDAG and its combination with OpenFE achieve the two best mean ranks among the compared methods; pairwise rank differences between feature-engineering methods do not survive multiple-testing correction.

The LLM-based