Intelligent Search
Intelligent Search(or at least really clever)
Some Preliminaries
• Text retrieval = matrix multiplication
A: our corpusdocuments are rowsterms are columns
Some Preliminaries
• Text retrieval = matrix multiplication
for each document d:for each term t:sd += adt qt
A: our corpusdocuments are rowsterms are columns
Some Preliminaries
• Text retrieval = matrix multiplication
A: our corpusdocuments are rowsterms are columns
sd = Σt adt qt
Some Preliminaries
• Text retrieval = matrix multiplication
A: our corpusdocuments are rowsterms are columns
s = A q
More Preliminaries
• Recommendation = Matrix multiply
A: our users’ historiesusers are rowsitems are columns
More Preliminaries
• Recommendation = Matrix multiply
A: our users’ historiesusers are rowsitems are columns
Users who bought itemsin the list h also bought items in the list r
More Preliminaries
• Recommendation = Matrix multiply
for each user u:for each item t1:for each item t2:
rt1 += au,t1 au,t2 ht2
A: our users’ historiesusers are rowsitems are columns
More Preliminaries
• Recommendation = Matrix multiply
A: our users’ historiesusers are rowsitems are columns
sd = Σt2 Σu au,t1 au,t2 qt2
More Preliminaries
• Recommendation = Matrix multiply
A: our users’ historiesusers are rowsitems are columns
s = A’ (A q)
More Preliminaries
• Recommendation = Matrix multiply
A: our users’ historiesusers are rowsitems are columns
s = (A’ A) q
More Preliminaries
• Recommendation = Matrix multiply
A: our users’ historiesusers are rowsitems are columns
s = (A’ A) q ish!
Why so ish?
• In real life, ish happens because:
• Big data ... so we selectively sample
• Sparse data ... so we smooth
• Finite computers ... so we sparsify
• Top-40 effect ... so we use some stats
The same in spite of ish
• The shape of the computation is unchanged
• The cost of the computation is unchanged
• Broad algebraic conclusions still hold
Back to recommendations ...
Dyadic Structure● Functional
– Interaction: actor -> item*● Relational
– Interaction ⊆ Actors x Items● Matrix
– Rows indexed by actor, columns by item– Value is count of interactions
● Predict missing observations
Fundamental Algorithmics● Cooccurrence
● A is actors x items, K is items x items● Product has general shape of matrix ● K tells us “users who interacted with x also
interacted with y”
Fundamental Algorithmic Structure● Cooccurrence
● Matrix approximation by factoring
● LLR
But Wait ...
But Wait ...
Does it have to be that way?
What we have:
For a user who watched/bought/listened to this
What we have:
For a user who watched/bought/listened to this
Sum over all other users who watched/bought/...
What we have:
For a user who watched/bought/listened to this
Sum over all other users who watched/bought/...
Add up what they watched/bought/listened to
What we have:
For a user who watched/bought/listened to this
Sum over all other users who watched/bought/...
Add up what they watched/bought/listened to
And recommend that
What we have:
For a user who watched/bought/listened to this
Sum over all other users who watched/bought/...
Add up what they watched/bought/listened to
And recommend that
ish
What we have:
Add up what they watched/bought/listened to
What we have:
Add up what they watched/bought/listened to
But wait, we can do that faster
What we have:
Add up what they watched/bought/listened to
But wait, we can do that faster
But why not ...
But why not ...
But why not ...
Why just dyadic learning?
But why not ...
Why just dyadic learning?
Why not triadic learning?
But why not ...
Why just dyadic learning?
Why not p-adic learning?
For example● Users enter queries (A)
– (actor = user, item=query) ● Users view videos (B)
– (actor = user, item=video)● AʼA gives query recommendation
– “did you mean to ask for”● BʼB gives video recommendation
– “you might like these videos”
The punch-line● BʼA recommends videos in response to a query
– (isnʼt that a search engine?)– (not quite, it doesnʼt look at content or meta-data)
Real-life example● Query: “Paco de Lucia”● Conventional meta-data search results:
– “hombres del paco” times 400– not much else
● Recommendation based search:– Flamenco guitar and dancers– Spanish and classical guitar– Van Halen doing a classical/flamenco riff
Real-life example
Real-life example
System Diagram
Viewing Logs t user video
Search Logs t user query-term
selective sampler
selective sampler
count
count
join on user
count
Related videos v => v1 v2...
Related terms v => t1 t2...
llr + sparsify
Hadoop
Indexing
Related videos v => v1 v2...
Video meta v => url title...
join on video
Lucene Index
Related terms v => t1 t2...
Hadoop Lucene (+Katta?)
Hypothetical Example● Want a navigational ontology?● Just put labels on a web page with traffic
– This gives A = users x label clicks● Remember viewing history
– This gives B = users x items● Cross recommend
– BʼA = click to item mapping● After several users click, results are whatever
users think they should be
Resources● My blog
– http://tdunning.blogspot.com/● The original LLR in NLP paper
– Accurate Methods for the Statistics of Surprise and Coincidence (check on citeseer)
● Source code– Mahout project– contact me ([email protected])