2. At the beginning there were Facebook Apps
Will they match ?
Will he buy ?
The Mood Weather Report
Buy icons to be displayed on your Facebook timeline
BeMyGentleman
Dating WebSite
3. The multi-object graph-based recommender
T
T
Diet
Beauty
T
Mortgage
T
Health
TSport
T
White color
Tag
Content
Product
User
Clicked
Watched
Viewed
Read
Bought
Added to
basket
Collect data and keep it in its original
graph structure
- Any type of object can be used &
recommended (tag, user, product, content … )
- Edges are build from tracked actions : buyings,
readings, tagging…)
Compute click predictions from the
paths linking users to objects we
could recommend
4. General workflow
Graph processor :
- Aggregates edges
- Trims irrelevant ones
Reduced Graph
Complete Graph
Learning Base Recommender
- Suggests objects
- Collects successes & failures
ParametersMachine learning
- Set rules for graph processing
- Optimize real-time decisions
5. Use cases
Sélectionné pour
vous
Sélectionné pour
vous
Recommendation : venues on TickeTac
Edges : tracking on website only
Recommendation : advertisment for Delamaison
furnitures on Prisma Group
Eges : readings on Femme Actuelle & tracking on
website
Recommendation : One site, advertisment
Edges : tracking on website, content views
6. Using media content to find targets
Read
Bought
Read
Read
Watched
Watched
LEA, ID 456
Lose weight before summer
FemmeActuelle content
Rome : the Eternal City
Geo content
Real Estate : interest rate
Capital content
TéléLoisir, Top
Chef, cuisine,
M6, télé réalité
Would Lea be interested by
l’Oréal ?
- Problem : the campaign has never
been launched : no natural « L’Oréal »
node connected to the general graph
First idea : use editorial tags
- Low quality
- Not exhaustive
- Not standardized among publications
Second idea : tag content with
a semantic algorithm
7. Automatic tagging of media content
𝐴~𝑇𝑘 𝑆 𝑘 𝐷 𝑘
𝑇
The corpus is composed of media
contents
The thesaurus contains 40 000
expressions a marketing operator
would be willing to search.
1) Corpus Processing : lematization, entropy computation
2) Truncated Singular value decomposition of the Document / Term matrix
Document / Term matrix
Term / Semantic space matrix Singular Values (we keep the 250
highest values)
Document / Semantic space matrix
3) Projection of the thesaurus onto the semantic space
𝑑 𝑘~𝑎 𝑇
𝑇𝑘 𝑆 𝑘
−1
3) Tag attribution by cosinus computation
8. Connecting user to campaign
User – Content : collected on the web
Shortest path
from user to campaign
Content – Topic : semantic algorithm
Topic – Product : chosen by the operator
Can be very subjective !
Mosturizer ?
Skin health ?
Cosmetic ?
Dry skin ?
9. Behavioral & Semantic extension
Pure semantic extension : a lot of articles are tagged with both topics
Semantic link
User action
Behavioral extension : a lot of users are interrested in both topics
Mosturizer Hydration
Mosturizer Food Supplements
10. And more cool projects …
Recommendation Target Generator Digital Advertising
Blouse col ruché
Dernière tendance, la blouse
se révèle pratique et chic [...)
25,90€ 18,20€
-
3O
%
Sélectionné
pour vous sur
Un produit
identifié pour
vous sur
Mini Pipistrello
Lampe noire Touch Led H35
cm
[...)
729,00
€
495,00
0€
-
32
%
VIDEOS – Jade et Arnaud Lagar-
dère : avant « Envoyé Spécial »,
ils s'étaient illustrés dans des
séquences inoubliables
ARTICLES – Lola Dewaere, la fille
de Patrick Dewaere, refuse de
participer à DALS avec beaucoup
d’humour et de [...]
- Find users reading about topics linked to
your product.
- Choose topics among 40k possibilities.
- Real-time personalized advertising
- Target new high-potential users on a
display network
- Improve the user experience on your
website.
- Increase your conversion rate.
11. Join us !
We have money … 2,8 M€ raised
More and more clients …
We need talents !
To grow To lauch new products To internationalize
simon@antvoice.com