调参日志:记录参数、训练集得分、测试集（提交答案）得分 - ShenDezhou/rct-intel-hw-recommend GitHub Wiki
- #1#3, #2#4 shows by 有l2_norm(即不设--embed_l2=1e-6参数)更好
- #1#2 shows by 减少epoch+增大lr更好
- #1#5 shows by 减少batch size几乎没影响
- #1#6 shows by 减小embed dim更好
- #5#7 shows by 减少neg来平衡数据更糟
- #1#8 shows by 增加来平衡数据更糟，平衡数据没有收益
1.--epochs=10 --batch_size=1024 --embed_dim=10 --learning_rate=1e-2
线上训练集：
{'read_comment': 0.7933732360130826, 'like': 0.7580773920201094, 'click_avatar': 0.8567413676355881, 'forward': 0.864511659963672} Weighted uAUC: 0.802572
线上测试集： 0.614079
2.--epochs=100 --batch_size=1024 --embed_dim=10 --learning_rate=1e-3
线上Train：
{'read_comment': 0.9296389102634225, 'like': 0.9276057828706135, 'click_avatar': 0.9763918319179635, 'forward': 0.9921951744441916} Weighted uAUC: *0.944635*
线上Test： 0.581364
3.--epochs=10 --batch_size=1024 --embed_dim=10 --learning_rate=1e-2 --embed_l2=1e-6
线上Train：
{'read_comment': 0.5728769691240823, 'like': 0.5701266322666655, 'click_avatar': 0.6763979779112091, 'forward': 0.6589805821336259} Weighted uAUC: *0.601366*
线上Test： 0.572742
4.--epochs=100 --batch_size=1024 --embed_dim=10 --learning_rate=1e-3 --embed_l2=1e-6
线上Train：
{'read_comment': 0.589317704686516, 'like': 0.5791047343396772, 'click_avatar': 0.6916181607551332, 'forward': 0.6714297988740285} Weighted uAUC: *0.614925*
线上Test： 0.550584
5.--epochs=10 --batch_size=8 --embed_dim=10 --learning_rate=1e-2
线上Train：
{'read_comment': 0.7170037704597131, 'like': 0.7179916952979987, 'click_avatar': 0.830932173337868, 'forward': 0.8569557624429204} Weighted uAUC: 0.754081
线上Test： 0.614417
6.--epochs=10 --batch_size=1024 --embed_dim=32 --learning_rate=1e-2
线上Train：
{'read_comment': 0.929857967278868, 'like': 0.9109560878170173, 'click_avatar': 0.9712593062700721, 'forward': 0.9767509915612788} Weighted uAUC: 0.937157
线上Test： 0.602205
7. 平衡数据(减少neg到pos)，--epochs=10 --batch_size=8 --embed_dim=10 --learning_rate=1e-2
线上Train：
{'read_comment': 0.7333809135325238, 'like': 0.7306940034156933, 'click_avatar': 0.8442448791976356, 'forward': 0.8246810086586831} Weighted uAUC: 0.763878
线上Test： 0.592053
8. 平衡数据(复制pos到neg)，--epochs=10 --batch_size=1024 --embed_dim=10 --learning_rate=1e-2
线上Train：
{'read_comment': 0.9661392482912947, 'like': 0.9877125054445267, 'click_avatar': 0.9991942156567207, 'forward': 0.9996995867431433} Weighted uAUC: 0.982578
线上Test： 0.557775