140 6. MULTIMODAL SEQUENTIAL LEARNING
Figure 6.7, the new given micro-video belongs to the 55th category, and the attention
mainly focuses on the micro-videos of the same category in the historical sequence. is
demonstrates the attention layer can help obtain improved features according to different
new micro-videos.
Clicked Micro-videos
New Micro-videos
55 55 101 233
55
0.0125
0.0100
0.0075
0.0050
0.0025
Figure 6.7: Visualization of the attention mechanism in the prediction layer.
6.6 SUMMARY
In this chapter, we present a temporal graph-based LSTM model to intelligently route micro-
videos to the target users. To capture the users’ dynamic and diverse interest, we encode their
historical interaction sequence into a temporal graph and then design a novel temporal graph-
based LSTM to model it. As different interactions reflect different degrees of interest, we build
a multi-level interest modeling layer to enhance users’ interest representation. Moreover, our
model extracts uninterested information from true negative samples to improve the recommen-
dation performance. To justify our scheme, we perform extensive experiments on two public
datasets, and the experimental results demonstrate the effectiveness of our model.