2. • Duplicate results for
the same product
• Poor categorization
• Poor customer
experience
• Poor success rates
and time-on-site
metrics relative to
curators
• Limited to linear UI
• Limited potential for
personalization
Example: Commerce Search With a
Key Value Index
3. Commerce Search With a
Contextual Index
• Easy to collapse and
group by categories
• Rich user
experiences are
possible
• Ranking can be
highly personalized
• Enables highly
targeted, rich
advertisements
4.
5. • Remember the semantic web?
• The amount of training data you need to train
existing algorithms grows exponentially with the
number of nodes
• Can cost millions of dollars to train a 5,000 node graph
• However, lots of recent advances
• Ranking technology is less mature
Problem: Scale
6. • Limit the total number of nodes in a vertical by building
custom vocabularies
• Don’t try to span the entire space early on: a 1000 node
graph will span 95% for most verticals
• Metrics and ranking models need to be tuned to
requirements of the vertical – don’t build one-size-fits-all
applications
Solution: Custom Ontologies
8. Yes: Well defined ontologies
- Commerce search, job search, music search
Maybe: Undefined, open ontologies, but an ontology exists that
would span enough to be useful
- Health care search, code queries, app search
- Important: First find the spanning set, then build the index and ranker
No: Unbounded ontologies
- Web search, natural language search, etc.
- Google, IBM, and Microsoft have failed
- Exception: when your customers map the data for you (Facebook,
LinkedIn)
Final Considerations: not a good
option for every solution!