为什么魔像算法的可能性降低
我正在使用python应用魔像算法进行休闲推断，但由于某种原因，我的可能性得分正在降低。为什么会发生这种情况？
我有一个非常大的数据集，其中包含310127104行。我无法执行此操作，因为此方法对于此特定任务不当吗？
输出看起来像这样：
[2022-06-28 19:18:02,137][golem.py - line 204] - INFO - Started training for 1200 iterations.
[2022-06-28 19:18:04,794][golem.py - line 217] - INFO - [Iter 0] score=142.987, likelihood=142.987, h=0.0e+00
[2022-06-28 20:46:41,095][golem.py - line 217] - INFO - [Iter 100] score=141.733, likelihood=141.655, h=9.5e-04
[2022-06-28 22:07:21,093][golem.py - line 217] - INFO - [Iter 200] score=140.727, likelihood=140.475, h=2.0e-02
[2022-06-28 23:27:34,371][golem.py - line 217] - INFO - [Iter 300] score=139.868, likelihood=139.517, h=3.2e-02
[2022-06-29 00:48:43,334][golem.py - line 217] - INFO - [Iter 400] score=138.987, likelihood=138.560, h=4.0e-02
[2022-06-29 02:06:52,349][golem.py - line 217] - INFO - [Iter 500] score=137.903, likelihood=137.324, h=4.9e-02
[2022-06-29 03:28:45,860][golem.py - line 217] - INFO - [Iter 600] score=136.425, likelihood=135.597, h=7.0e-02
[2022-06-29 04:53:14,669][golem.py - line 217] - INFO - [Iter 700] score=135.297, likelihood=134.317, h=8.3e-02
[2022-06-29 06:19:03,769][golem.py - line 217] - INFO - [Iter 800] score=134.337, likelihood=133.297, h=8.3e-02
预先感谢您！
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（1）
我会继续前进，假设您是从 Ignavier等人。 2020 。
在优化中，它们最大程度地减少了由负模样和额外的惩罚条款组成的损失（例如，请参见等式3）。使用损失（必须最小化）而不是必须最大化的分数是约束优化的常见实践。这主要是出于惯例和一致性的原因。
对于Golem，您可以在在第85行中。因此，您的模型似乎是按照预期的（输出中的“得分”和“可能性”一词可能有点不精确）。
注意注意：
鉴于您可能在现实世界中运行魔像，对您来说，重要的是要考虑到魔像的出色经验结果（以及一系列类似的算法）似乎是由于基准和基准中的模式所致如。
I'll go ahead and assume you are refering to the "GOLEM" causal structure learning algorithm from Ignavier et al. 2020.
In their optimization, they minimize a loss consisting of the negative log-likelihood and additional penalty terms (see e.g. equation 3). It is common practice in constraint optimization to use a loss (which has to be minimized) rather than a score which would have to be maximized. This is mostly for reasons of convention and consistency.
For GOLEM, you can see this in the official implementation in line 85. It seems therefore that your model is training as intended (the words "score" and "likelihood" in the output may just be a little imprecise).
Note of Caution:
Given that you may be running GOLEM on real-world data, it could be important for you to consider that the excellent empirical results of GOLEM (and a range of similar algorithms) appear to be due to patterns in the benchmarks and likely do not translate to the real-world as shown in Reisach el al. 2021.