Possible mixup in metrics for binary classification (leads to wrong model selected)

Start by reproducing the data-classification scenario in Model Builder with two categorical classes and Macro-Accuracy optimization, then compare the reported metrics and selected algorithm across binary and multiclass variants. Done means the metric behavior is verified and corrected consistently, with the expected warning behavior addressed if the issue’s proposed case is confirmed.

System Information (please complete the following information):

• Model Builder Version (available in Manage Extensions dialog): 17.19.2.211501
• Visual Studio Version 2022

Describe the bug
Training a multi-category model with just a 2 classes with Macro-Accuracy optimization metric.

I get:
46 FastTreeOva 0.5500 1.2160 |
|47 FastTreeOva 0.5259 1.1710 |
|6 LightGbmBinary 0.6117 1586.0270 |
|17 LightGbmMulti 0.5003 560.4780 |
|7 SdcaLogisticRegressionBinary 0.6583 2.0900 |
|18 FastTreeOva 0.5179 1.3650 |
|19 FastTreeOva 0.5097 1.0210 |
|8 LbfgsLogisticRegressionBinary 0.6583 1.5410 |
|20 FastTreeOva 0.4868 1.2010 |

LbfgsLogisticRegressionBinary gets selected, but it is exactly at the null accuracy 0.658287% for the most common category so my suspect it is just selecting true every time. If that is the case, then macro-accuracy should be 0.5%. I believe that is true, because the multiclass algorithms have much lower metrics.

Although I haven't confirmed the issue, I think the fact that binary variants report much higher accuracy point to the possibility of calculating the metric differently.

To Reproduce

1. Scenario: data classification
2. Set label to categorical . Make sure Boolean is not selected as type, but it just has 2 categories. For example, set 1/0 to numeric, true/false to text so binary algorithm variants will activated
3. Train

Expected behavior
Use macro-accuracy / balanced accuracy as the metric for all algorithms.

Ideally, in binary classification we should also see macro accuracy, or at least warning if it is exactly at null accuracy or always predicting one label.