What is the key to the holistic success of the fastest growing and most successful companies of our time globally? Well, often the key is the rapid increase in collected and analysed data. Graph databases provide a way to organise semantically by classes, not tables, are web-aware, and are superior for handling deep, complex relationships than traditional relational or NoSQL data stores. It is these deep, complex relationships that can provide the rich context for hyper-personalising your product offering, inspiring consumers to purchase. In this talk, we describe how we are using artificial intelligence at Farfetch to not only help build a knowledge graph but also to evolve our insights with state-of-the-art graph-based AI.

1. Knowledge Graphs and AI to
Hyper-Personalise the Fashion
Retail Experience at Farfetch
@GeorgeCushen
Connected Data London 2019

2. 2
Image: Kelly Sikkema

3. 3
Outfit available from
https://www.farfetch.com

4. Image: Paramount Pictures

5. Farfetch at a glance
5
> 3,000*
Employees across 13 countries
$1.4 Billion*
Gross Merchandise Value
> 3,000*
Brands available for consumers
to shop
> 1,000**
Luxury sellers on the
Marketplace
$601**
AOV on Marketplace
> 2.9 Million*
Orders on Marketplace
1.7 million**
Active Marketplace consumers
$307 Billion
Size of personal luxury good
industry (Bain estimates)
*Correct for full year 2018 **As at Q1 2019
15**
Marketplace language sites

6. 6
Background

8. 8
Image: Walt Disney Television (Flickr)

9. A New Perspective: Emphasising Relationships
● Businesses and their products/services are all about Entities and Relationships
● Examples of entities and relationships in industry:
Farfetch Consumer searches Product with Terms
Amazon Seller sells Product to Consumer
Uber Driver provides Trip to Rider
Facebook Person shares Status with Friend
● How can we represent, analyse, and visualise this kind of data?

10. 10
What is a knowledge graph?
A knowledge graph can describe
● a collection of nodes (entities) representing business and fashion entities
has_term
has_synonym
has_child
Properties:
Inherit = true
● and with labeled relationships between the nodes
Product
D&G
tote bag
Attribute
Leopard
Print Attribute
Leopard
Spots
Attribute
Animal
Print
Properties:
Language = “EN”
● each containing information (properties)
Properties:
ProductID = 123

11. 11
Dots and Lines

12. 12
Why use a knowledge graph?
● Have naturally highly connected-data
● Derive new insights with Graph Analysis & Graph-based AI
● Enable stakeholders to easily visualise relationships and make informed decisions
● Flexible schema to facilitate evolution to expand business entities
● Optimized for storing and querying graphs
○ Significantly faster than SQL databases for querying relationships
○ Relationships are a fundamental structure, so following relationships is a
single lookup, making this operation blazingly fast

13. Where Business Meets Fashion
A domain specific knowledge graph for fashion.
Business vs Fashion Entities
Business Fashion
Product
Content
Brand
Category
Customer
Season
Gender
...
Occasion
Celebration
Theme
Style
Trend
DNA
Pattern
Colour
Material
Synonym
...
Order
Payment
Promotion
Review
...

14. 📖 Constructs a unified semantic fashion vocabulary
🏷 Connects these fashion entities with business entities in a KG via AI
🧬 Infers DNA from the relationships in the Knowledge Graph (KG)
We’re mapping fashion DNA to decode personal style

15. We’re mapping fashion DNA to decode personal style
Loosely Structured
Data
Data Science Data Science
Powerful fashion
DNA, new
knowledge, and
insights

16. 16
Example Use Cases
Free Text Search
Increase product discovery with
synonyms and rich attributes for
material, occasion (e.g. skiing), etc.
Semantic Search
Increase product discovery based
by using graph to understand
consumer’s intent
Ranking
Leverage rich product connections to
increase relevance on listing pages
Recommendations
Increase relevance based on richer
product attributes and deep graph
relationships

17. 17
Communicating a graph
Product Managers
“How can we improve the
customer experience?”
“How can we increase
GMV/revenue?”
Data Scientists
“Wow, looks like a NN,
hold my Pandas 🐼🐼🐼,
I’m onboard!!”
Backend Engineers
“Why do we need a
graph?”
“Which graph database
meets the requirements?”
Data Engineers
“Is your Airflow
dizzy🥴😵? It’s
traversing through cyclic
connections💫?!”

18. 18
Building a fashion knowledge graph

19. 19
Perception

20. 20
Subjectivity

21. 21
Building a fashion knowledge graph
Search Recommendations ...
Fashion Knowledge Graph
Associates fashion entities with business entities
AI Knowledge cleaning Entity resolution Schema mapping
Applications
Taxonomy &
Graph
Construction
Knowledge
Collection
Expert Knowledge Data-Driven Insights

22. Techniques
📷 Computer Vision +
📖 NLP +
✔ Conflation +
👙 Inference +
👥 Crowdsourcing
22

23. 23
AI: A Multi-Modal Multi-Task Approach
Images Text
Computer
Vision NLP
Deep
Classifier
Example output
Product Type: Dress
Colour: White
Occasion: Wedding
Theme: Classic
Embeddings?
NER?
Coreference
resolution?
Relationship
extraction?

24. Skinny
24
Universal Fashion Taxonomy
Fashion
Taxonomy
Synonyms
Descriptive
attributes
Brand DNA
Materials
ColoursTrends
Editorial,
emotive,
seasonal
concepts
Textile Cotton Denim
Product
2
Swedish
Design
Acne
Connected
Data
Conferen
ce
Autumn
Product
1
PrintsCircles
Blue
Light
Blue

25. Synonym Enrichment
Padded
coat
Down
coat
Duvet coat
Quilted
coat
Puffer
jacket
Down-filled
jacket
Down
jacket
Quilted
jacket
Duvet
jacket
Down-filled
coat
Padded
jacket
Puffer
coat

26. 26
Richer Product Data
Existing
catalog
External Enrichment
Internal Enrichment

27. 27
Richer Product Data
Existing catalog
data
AI predicts richer and
more diverse attributes to
help construct the graph
Graph based AI and analytics
further enrich attributes and infer
product DNA
Qualityof
ProductDNA
RichproductDNA

28. 28
Deriving new knowledge and insights

30. 30
Discovering the pearl
DELFINA DELETTREZ 'Trillion' earring

31. 31
Features from Graphs
Extract features from the graph such as:
● nodes
○ degree
● pairs
○ number of common neighbours
● groups
○ custer assignments
● Infer DNA
● Link Prediction
● Anomaly Prediction
● Clustering
● ...
Adjacency Matrix

32. 32
360o
Customer View
360o
Customer
View
Social
Email
Call
CentreClick-
stream
PoS
and
ClientelingPurchase
History
Style
Preferences

33. Identity Resolution with Graph Analytics
33
Person A Person BPerson A
Account 1 Account 2 Account 3
Call
Centre
Web/App
Family A
...
...

34. 34
What is Deep Walk?
Learn a latent representation of adjacency matrices
using deep learning based language processing.
● Infer DNA
● Link Prediction
● Anomaly Prediction
● Clustering
● ...
Adjacency Matrix Latent Representation

35. 35
How to perform Deep Walk
Image: Jazeen Hollings

36. 36
How to perform Deep Walk
Image: Perozzi et al.

37. 37
Node2Vec
Images: Semantic Scholar, SNAP Stanford

38. 38
Graph2Vec
Image: Lego
Word (wj)
Document (d)
Document embedding matrix (d-->)
Word embedding matrix (wj
)
Vocab list of words (V)

39. 39
Vertex and Graph Embeddings
Vertex embedding approaches:
DeepWalk, Node2Vec, LLE, Laplacian Eigenmaps, Graph Factorization,
GraRep, HOPE, DNGR, GCN, LINE
Graph embedding approaches:
Graph2Vec, Patchy-san, sub2vec, WL kernel, Deep WL kernels
Image: rocknwool on Unsplash

40. Image: Kim Albrecht

41. 41
Summary

42. 42
Takeaways
● Graphs can offer a new, democratised
perspective on enterprise data
● When graph based analytics and AI
are performed on connected data, we
can derive powerful new knowledge
and insights
● Which can drive hyper-personalisation,
improving the customer experience

43. 43
Questions
@GeorgeCushen
#Farfetch
We’re hiring!