tdm 运行tree_cluster.py 之后，再次训练，recall变低了很多

第一次训练tdm之后（opoch=5次），在1w数据集上面测试（UserBehavior），recall为4.4%，之后运行tree_clutser.py之后，再次训练tdm，(epoch=5次)，再次测试相同数据集recall变为1.4%，不合理，感觉聚类之后 还变低了。麻烦帮忙看看，谢谢！！聚类之后，userbehavoir_tree.pb 也是更新了。

Reproduce the two five-epoch TDM runs on the 1w UserBehavior dataset, recording recall before and after running tree_cluster.py. Compare the generated userbehavoir_tree.pb and the training inputs to isolate what changes after clustering. Done means identifying why recall falls from 4.4% to 1.4% and documenting a verified correction or reproducible finding.

Written by the indexing model from the issue text.