Re-rank as a pattern relationship between the ordination and impact of recommended items, more and more applications in recent years is the recommended system. Conventional recommendation system is only considered <user, item> pair, and will not be considered as a list of mutual influence between the recommended items. Rerank conventional model output rank model as input items and relationships strengthen the relationship between the user and items for one-more-time sorting. For different users, list the items of distribution should have a relatively large difference.
Personalized Re-ranking for Recommendationarxiv.org
Learning a Deep Listwise Context Model for Ranking Refinementarxiv.org
The idea of this paper is largely inherited the idea DLMC (Deep Listwise Context Model) is, DLMC rnn method is used, based on the scene search, transormer this paper fires in multi head structure to fit the relationship between the item and impact, based on the recommended scenario. Two paper emphasized the point is that you can seamlessly migrate to the industry recommendation system, as used in the previous step rank features are used in the feature, we do not need extra feature extraction. Transformer replacement RNN advantage is the more common cognitive 1. Parallelization easier engineering algorithms for item 2 on long-distance relationship portrayed better. Due to different scenarios, DLMC emphasized that the feature distribution under different query is different, for example, search for "Friends", and that the correlation is even more important compared to diversity, we highlight the impact of different users of different intention of the recommendation list , such as a user no obvious intent to buy, it is recommended diversity becomes important.
Model Architecture
FIG PRM is the model structure is a relatively conventional multi head structure, there are input layer, encoding layer and output layer of three parts. Initial List recommendation set input is a row of fine model output outputting Re-ranked List is the result of reordering.
structure of PRM(Personalized Re-ranking Model)
Input Layer: is characterized in that the item row upon finishing the model,
the user of the item and a feature of the cross, obtained by the pre-train, the method used to pre-train a method of dnn, simple and crude, as shown in FIG. the author proved in later experiments, this method does to the overall precision to some extent improved.
Is embedding a position, mainly using fine row ranking model output, through experiments, the authors found
to trainable effect better. Embedding the above-mentioned three groups are part of the input layer.
user-item pretrain method
Encoding Layer: using the Transformer self-attention structure, to achieve any two cross item so as to obtain mutual influence between them. Encoding Layer Transformer is a conventional structure, Q, K, V is the same matrix as shown below.
transformer encoder
Output Layer: front transformer out of the score to be a softmax.
Last but not the least, authors experimentally found that a variety of mature re-rank algorithm are all positive benefits, this is well understood, re-rank is an incremental process, can play a list-wise role. Throughout the paper industry to achieve it should not be a bottleneck, transformer also a good solution to the problem before consuming rnn, providing we do a re-rank ideas.
