Post on 02-Jul-2015
Making a simple question into a complicated query
Richard BoultonLemur Consulting Ltd
Making a simple question into a complicated query
Richard BoultonLemur Consulting Ltd
Making a simple question into a complicated query
Richard BoultonLemur Consulting Ltd
Only 20 minutes until Lunch
Where shall we have Lunch?
“Lunch”
Assertion
Complicated questions areeasier to answer well.
Assertion
Complicated questions areeasier to answer well accurately.
� Restaurant � Pizza restaurant near Covent Garden, fairly cheap.
Time for a real example
http://mydeco.com/
Interior decoration site
� Users type: “Sofa”
� We'd prefer them to ask questions like: “Red velvet, three seater, sofa, from a supplier who can deliver to central Cambridge at a weekend”.
� How can we move to this kind of search?
Getting more from users
Getting more from users
� Suggested search completions
Getting more from users
� Facets
Getting more from users
� Facets
� Which facets to display?
− Depends on the user.
� Which facet values are interesting?
− A particularly fun problem for continuous numeric values, like price.
� How many values should we display?
− Based on likelihood of any being useful?
Getting more from users
� Personal data
− Using details about the user directly.
� e.g., Postcode
− Grouping users by similarity of interests
Getting more from users
� Similarity search
− “More like this”
− Colour / image-based similarity
Behind the scenes
� Applying our own bias.
− Perhaps we want to push some items
− Perhaps we want to avoid other items
− Perhaps some items go well together
− Behave like a shop assistant
− “Product Rank”
Behind the scenes
� Categorisation
− User asks for “Sofa”.
− We search for “Products categorised as one of the sofa subcategories, based on the output of a machine learning system trained with some human judgements”.
Behind the scenes
� Variety
− Don't display lots of very similar items
− Give the user a choice
− But don't display irrelevant junk, either!
− Need some way to measure variety
Answering complicated questions
Answering complicated questions
� Getting the best answer
− Good models
− Careful design
− Lots of tuning
Answering complicated questions
� Getting an answer quickly
− Good algorithms
− Well matched data-structures
− Plenty of machines
− Plenty of RAM
Questions, and then Lunch!
Richard BoultonLemur Consulting Ltd
richard@lemurconsulting.com