总结
加序列emb,multi-head self-attention/transformer
细节
当输入list排序变化后,用rank模型输出不变的排序list。multi-head self-attention堆叠解决。
representation-encoding-ranking
先用现有的ranking model做出来init ranking,再multi-head attention做encode,最后fnn做ranking。
交叉熵损失
实验
dataset
- Istella LETOR:http://blog.istella.it/istella-learning-to-rank-dataset/
- Microsoft LETOR 30K:http://research.microsoft.com/en-us/projects/mslr/
- Yahoo! LETOR:http://learningtorankchallenge.yahoo.com
baseline:rankSVM, rankBoost, MART, LambdaMart, DLCM, GSF
评估指标:ndcg@1,3,5,10