2. Types
Collaborative Filtering
X is frequently bought together with Y
Customers often buy Y after looking at X
Customer who bought X item also bought Y
Content based
Buy X because it has a relationship with Y
Like Robbie Williams, try Take That
Glyn Darkin - @glyndarkin
3. Context
User
Recommendation is determined based on user context
We recommend X product because you purchased Y
Amazon’s
Today’s Recommendations For You
Product
Recommendation is determined based on product context
We recommend X product because other’s purchased Y
Amazon’s
Frequently Bought Together
Glyn Darkin - @glyndarkin
4. Collaborative Filtering
Focus on statistical analysis of the relationship
between products and people
No knowledge of product domain required in
analysis
Technology of note
R
statistics programming language
Mahout
Machine learning and data mining
Standard Amazon Recommendations
Glyn Darkin - @glyndarkin
5. Content Based
Focus on product graph and defining the
relationships between products
Some domain knowledge of the products is required
Dependency on quality external metadata
If you want to cross sell red house hold products you will
need a good data source to provide it
Technology of note
Neo4J
Graph database
Lucene / SolR
Full text search
Basic More episodes recommendation on the BBC
iPlayer
Glyn Darkin - @glyndarkin
6. Summary
Both Collaborative & Content based
recommendations can be of a user or product
context.
Context is important as it defines the schema of data
capture required to deliver the recommendation
The sweet spot is probably in a hybrid approach to a
recommendation
We must not forget the Social recommendation
where a 3rd party body of trust recommends a
product
Glyn Darkin - @glyndarkin
7. Delivery Mechanism
Targeted
Email – could be either User or Product Context
Tweet – should be User Context
Personalised Homepage
Product page cross sell/upsell
Landing page merchandising
The delivery mechanism dictates the type of
technology required
Glyn Darkin - @glyndarkin
8. Data Capture Techniques
Batched export
Orders/Baskets
People
Product metadata
Real-time – Analytics packages
Pages
Transactions
Customer interaction
Customer Surveys
Not everybody will be able to capture
everything, therefore there maybe technology
requirements to capture particular data points
Glyn Darkin - @glyndarkin
9. Trends in recommendations
Amazon started recommendations trend
Everybody is tired of getting recommended products
that are not relevant to them caused by gifting or
one-off purchases
Upsurge in “Curated” sites
www.Etsy.com
www.shoedazzle.com
Glyn Darkin - @glyndarkin