Recommendation System - Recall Phase - 2013: DSSM (Twin Towers Model) [Embedding (Semantic Vector) Recall] [Microsoft]

DSSM, the full name of Deep Structured Semantic Models, was published on CIKM in 2013 by the University of Illinois at Urbana-Champaign (UIUC) and Microsoft.

DSSM's influence in industry search and promotion is currently more than any of the previously introduced models, Basically dominates the recall/coarse row stage.

Baidu also made a twin-tower model in 2013 (Baidu was still in its prime at that time) SimNet, but no corresponding paper has been published.

1. DSSM double-tower model structure recommended

The DSSM structure in the recommended field is generally shown in the following figure:

Figure 1. The structure of the commonly used DSSM model in the recommendation field

The result of the model is very simple and mainly consists of two parts: a tower on the user side and a tower on the item side.The user-side features and item-side features pass through their respective DNNs (generally, the two DNNs have the same structure, but of course they can be different) to obtain user embedding and item embedding. It should be noted here that if your user dnn andThe structure of item dnn is different, please ensure that the output dimension is the same, that is, the number of hidden units in the last fully connected layer is the same, which requires