Paper notes-deep Interest Network for Click-through rate prediction

Focus: Think different ads will trigger the user's point of interest is different, resulting in user embedding changes.

DIN network structure as right

The starting point of Din: Think different ads will trigger the user's point of interest is different, resulting in user embedding change.

It is considered that the user embedding vector is a function of the recommended ad vector, and the ad vector can be represented as the attention weighting of the historical behavior ID vector by the attention associated with the historical behavior-related ID vectors.

Personal understanding of training: the parameters of vector u are represented by the parameters of vector I and vector A, and the learning parameters are obtained by this constraint (relation).

Paper notes-deep Interest Network for Click-through rate prediction