- 多塔模型效果比单塔模型有明显的提升
- 不采用FM,所以embedding可以有不同的dimension。
model_config: {
model_class: 'MultiTower'
feature_groups: {
group_name: 'user'
feature_names: 'user_id'
feature_names: 'cms_segid'
...
feature_names: 'new_user_class_level'
wide_deep: DEEP
}
feature_groups: {
group_name: 'item'
feature_names: 'adgroup_id'
feature_names: 'cate_id'
...
feature_names: 'price'
wide_deep: DEEP
}
feature_groups: {
group_name: 'combo'
feature_names: 'pid'
feature_names: 'tag_category_list'
feature_names: 'tag_brand_list'
wide_deep: DEEP
}
multi_tower {
towers {
input: "user"
dnn {
hidden_units: [256, 128, 96, 64]
}
}
towers {
input: "item"
dnn {
hidden_units: [256, 128, 96, 64]
}
}
towers {
input: "combo"
dnn {
hidden_units: [128, 96, 64, 32]
}
}
final_dnn {
hidden_units: [128, 96, 64, 32, 16]
}
l2_regularization: 1e-6
}
embedding_regularization: 1e-4
}
- feature_groups: 不同的特征组,如user feature为一组,item feature为一组, combo feature为一组
- group_name: 可以根据实际情况取
- wide_deep: 必须是DEEP
- towers:
- 每个feature_group对应了一个tower, tower的input必须和feature_groups的group_name对应
- dnn: 深度网络
- hidden_units: 定义不同层的channel数目,即神经元数目
- final_dnn 整合towers和din_towers的输入
- hidden_units: dnn每一层的channel数目,即神经元的数目
- l2_regularization: L2正则,防止overfit
- embedding_regularization: embedding的L2正则
自研模型,暂无参考论文