展开查看详情
1 .Matchmaking Audiences
to Content
Lucy X. Wang, BuzzFeed
#AI8SAIS
4 .A vast network of social channels
• 10+ platforms
• 400+ accounts
• 1200+ pieces of
content published a
week
#AI8SAIS 4
5 .But how do we scale our curation?
#AI8SAIS 5
6 .By mimicking human curation
#AI8SAIS 6
7 .3 use cases for AI
1. Surface relevant content for each channel
2. Surface evergreen content
3. Automate publishing
#AI8SAIS 7
8 .Case 1: How do we recommend
relevant content for each channel?
#AI8SAIS 8
9 .Content too often slips through the
cracks
#AI8SAIS 9
10 .Training on human curation data
#AI8SAIS 10
11 .A language-based relevance model
#AI8SAIS 11
12 .Predicting relevance from titles
#AI8SAIS 12
13 .Consolidated into one model
#AI8SAIS 13
15 .Impact unlike rate remains stable
#AI8SAIS 15
16 .Case 2: How do we find evergreen
content to re-circulate?
#AI8SAIS 16
17 .Taking advantage of evergreen posts
2016 2018
55K clicks 36K clicks
#AI8SAIS 17
18 .A model based on content and traffic
performance
#AI8SAIS 18
19 .Not much human curation data
#AI8SAIS 19
20 .Our labelling proxy
timely
evergreen
#AI8SAIS 20
21 .Handling poorly-labeled data
timely?
#AI8SAIS 21
22 .Modeling with uncertainty
! 1
!"#$"%&'() = +
+ 234 ,(.( )+
+,(.( ) +
#AI8SAIS 22
24 .Impact
median clicks per post doubled
(+108%) with addition of
evergreen recommendations unlike rate remains stable
#AI8SAIS 24
25 .Case 3: How do we automate
publishing?
#AI8SAIS 25
26 .Mixing human and AI curation
#AI8SAIS 26
27 .Optimal scheduling model
#AI8SAIS 27
28 .Trying all scheduling times
#AI8SAIS 28
29 .How do we validate these models?
#AI8SAIS 29
热门城市
