研究者针对表格基础模型TabICL与TabPFN和调优XGBoost的性能进行了系统基准测试。1在14个数据集的对比中，TabICL在精确度指标上赢得12场并打成1平1负，在曲线下面积（AUC）指标上实现14连胜。1两个模型的平均精确度差异为0.01，TabICL平均精确度提升达+0.0106。1
推理效率方面，TabICL展现出明显优势。1TabICL的推理时间普遍在1秒以内，而调优XGBoost需耗时0.7至27.2秒。1在32000行数据规模下，TabICL最差种子得分（0.8216）仍优于XGBoost最佳种子（0.7950），两者无重叠区间。1此外，TabICL无需进行超参数调优，进一步简化了使用流程。1
然而，性能优势并非在所有场景下都保持一致。1在列数众多的宽表上，模型性能出现下降，Bioresponse数据集（419列）尤为典型，TabPFN 2.2.1版本在该数据集上的得分跌至0.7567，耗时18.1秒。1
测试使用了14个截至3000行的数据集，采用5个随机种子取中位数的方式，硬件配置为RTX 4070 Ti SUPER GPU。1在许可证方面，TabICL采用三条款BSD许可证可无障碍下载，而TabPFN 8.3.0版本因许可证障碍无法获取，测试采用了2.2.1版本。1
A comparative benchmark study has demonstrated the superior performance of TabICL, a foundation model for tabular data, over carefully tuned XGBoost across multiple metrics.1 In tests conducted on 14 datasets, TabICL won 12 comparisons and tied 1 while losing only 1 on accuracy measures, and achieved a perfect 14-win record on area under the curve (AUC) scoring.1 The model also demonstrated significantly faster inference speeds, with most predictions completed in under one second, compared to XGBoost's 0.7 to 27.2 seconds per run, all without requiring hyperparameter tuning.1
The performance advantage of TabICL proved robust even as data volume scaled substantially.1 At 32,000 rows, TabICL's worst-performing seed variant scored 0.8216 while XGBoost's best seed achieved 0.7950, with no overlap between the ranges.1 However, the model showed notable limitations on wide datasets with many columns; the Bioresponse dataset containing 419 columns represented an exception where TabPFN version 2.2.1 dropped to 0.7567 and required 18.1 seconds for inference.1 The benchmark testing was performed on 14 datasets each truncated to 3,000 rows, using five random seeds with median values reported, and ran on an RTX 4070 Ti SUPER GPU.1
Additional differences emerged regarding licensing and accessibility between the models tested.1 TabICL operates under a three-clause BSD license with unrestricted download availability, while TabPFN version 8.3.0 faced licensing barriers preventing download; consequently, testing used the earlier 2.2.1 version and TabPFN requires account authentication for access.1
评论
还没有评论，欢迎留下第一条。