6.5. EXPERIMENTS 139
1 2 3 4 5
0.700
0.702
0.704
0.706
0.708
0.710
0.712
0.714
AUC
L
1 2 3 4 5
0.3500
0.3525
0.3550
0.3575
0.3600
0.3625
0.3650
F@50
L
1 2 3 4 5
0.286
0.288
0.290
0.292
0.294
0.296
0.298
0.300
0.302
P@50
L
1 2 3 4 5
0.450
0.452
0.454
0.456
0.458
0.460
R@50
L
(a) AUC (b) P
@50
(c) R
@50
(d) F
@50
Figure 6.6: Illustration of the neighbor size L of the temporal graph on our recommendation
performance.
6.5.6 ATTENTION VISUALIZATION
As analyzed before, we fed the interested feature sequence
F
in
and a new micro-video’s em-
bedding x
new
into a vanilla attention layer to obtain the improved interested representation. To
intuitively illustrate the attention results, we randomly selected some new micro-videos from
the test data and visualized the attention scores in Figure 6.7. Several interesting observations
stand out.
• For each new micro-video, the attention scores of its historical clicked micro-videos are
different, which indicates that different micro-videos in the historical sequence contribute
differently.
• By and large, the earlier a micro-video locates in the sequence, the smaller the attention
score is, which indicates that the latter clicked micro-videos contribute more to the rec-
ommendation. is observation strongly supports that the user’s interest is dynamic.
• By visualizing the categories of micro-videos, we noticed that, micro-videos from the same
category contributes more to the recommendation results. As shown in the sub-figure of
..................Content has been hidden....................

You can't read the all page of ebook, please click here login for view all page.
Reset