AI 中文总结
本文提出Astro-Hunters端到端流程，将标签来源作为实验变量，对比不同分类器在TESS测光数据中系外行星凌星探测的性能，发现标签来源对性能影响远大于模型架构，相位折叠是信号可检测的关键。
AI 中文摘要
寻找太阳系外行星如今在很大程度上是一个数据分析问题：测光巡天返回的恒星光变曲线数量远超人工目视检查的能力，因此越来越需要机器学习来识别凌星行星产生的微弱、周期性变暗信号。这类探测器能否有效工作取决于一个很少被报告的选择：其训练标签是如何生成的。没有任何星表提供每个测光时刻（cadence）探测器必须分类的单独测量值，因此标注必须通过推导得到，而本文探讨了这种推导带来的代价。我们提出了Astro-Hunters，这是一个针对TESS两分钟测光数据的端到端流程，包含检索、去趋势化、每个测光时刻的七个滑动窗口统计量，以及梯度提升分类。这些组件均采用常规设计，创新之处在于将标签来源作为实验变量处理：从12颗已确认宿主星中获取的189279个测光时刻样本，是基于已发表的星历表而非测光数据进行标注的，同时为该任务设定了物理边界。我们在星不相交（star-disjoint）协议下对比了6类分类器。在保持特征、模型和协议固定的情况下，不同标签来源的精确率-召回率性能相差29倍，而不同架构间的差距仅为1.8倍。基于分类器自身特征拟合的孤立森林（isolation forest）生成的标签，其表观AUC为0.9915，该值衡量的是信号的循环性；未转换的凌星 epoch 会使性能降至随机水平；正确标注的标签对应的AUC为0.788，是基线流行度的5.3倍。性能上限是观测性的而非架构性的：中值单测光时刻信噪比为2.10，将单测光时刻AUC限制在0.932。Box Least Squares基线从一个天区中恢复了12个轨道周期中的8个。使凌星信号可被检测的是相位折叠，而非分类器的容量。
英文摘要
Finding planets beyond the Solar System is now largely a data-analysis problem: photometric surveys return far more stellar light curves than can be inspected by eye, and machine learning is increasingly asked to recognise the faint, periodic dimming of a transiting planet. Whether such a detector works turns on a choice that is seldom reported: how its training labels were made. No catalogue disposes the individual measurements a per-cadence detector must classify, so the annotation has to be derived, and this paper asks what that costs. We present Astro-Hunters, an end-to-end pipeline over TESS two-minute photometry: retrieval, detrending, seven sliding-window statistics per cadence, and gradient-boosted classification. These components are deliberately conventional. What is new is the treatment of label provenance as an experimental variable, a corpus of 189,279 cadences from twelve confirmed hosts annotated from published ephemerides rather than from the photometry, and a physical bound on what the task permits. Six classifier families are compared under a star-disjoint protocol. Holding features, model and protocol fixed, precision--recall performance spans a factor of 29 across label sources against 1.8 across architectures. Labels from an isolation forest fitted to the classifier's own features give an apparent AUC of 0.9915 that measures circularity; an unconverted transit epoch drives performance to chance; correct annotation gives AUC 0.788 at 5.3 times the prevalence baseline. The ceiling is observational, not architectural: a median single-cadence signal-to-noise ratio of 2.10 caps per-cadence AUC at 0.932. A Box Least Squares baseline recovers eight of twelve orbital periods from one sector. Phase-folding, not classifier capacity, is what makes the transit signal accessible.
