dataiku at sf datamining meetup - kaggle yandex challenge
DESCRIPTION
This is a presentation made on the 13th August 2014 at the SF Data Mining Meetup at Trulia. It's about Dataiku and the Kaggle Personalized Web Search Ranking challenge sponsored by YandexTRANSCRIPT
write your own data story!
short story
Founded January 2013
January 2014A Data Science Studio
powered team wins a Challenge
Founded January 2013
January 2014A Data Science Studio
powered team wins a Challenge
Founded January 2013
January 2014A Data Science Studio
powered team wins a Challenge
Data Science Studio’s GA February 2014
Founded January 2013
January 2014A Data Science Studio
powered team wins a Challenge
Data Science Studio’s GA February 2014
July 2014Data Science Studio
Available for Free with a
Community Edition !!
Founded January 2013
January 2014A Data Science Studio
powered team wins a Challenge
Data Science Studio’s GA February 2014
15 People Now
July 2014Data Science Studio
Available for Free with a
Community Edition !!
!BI
Developer
Data Preparation
Build Algorithm
Build Application
Run Application
Business Analyst
DataScientist
I don’t want to be a data cleaner anymore“
Finding Leaks in my Data Pipelines“
Waiting for the
(gradient boosted) trees
to grow“
MPP Databases
Statistical Software Machine Learning
No-SQL Hadoop
Demo Time
Challenge
Using Historical Logs of a search engine QUERIES RESULTS CLICKS !and a set of new QUERIES and RESULTS !rerank the RESULTS in order to optimize relevance
Personalized Web SearchYandexFri 11 Oct 2013 – Fri 10 Jan 2014 194 Teams $9,000 cash prize
No researcher. No experience in reranking.
Not much experience in ML for most of us. Not exactly our job. No expectations.
Kenji Lefevre 37
Algrebraic Geometry Learning Python
Christophe Bourguignat 37
Signal Processing Eng. Learning Scikit
Mathieu Scordia 24
Data Scientist
Paul Masurel 33
Soft. Engineer
The Team
A-Team?
“HOBBITS"
YANDEX SUPPLIED 27 DAYS OF ANONYMOUS LOG
Challenge Data
34,573,630 Sessions with user id 21,073,569 Queries 64,693,054 Clicks
~ 15GB
Example
Relevance?
A METRIC FOR RELEVANCE RIGHT FROM THE LOG? ASSUMING WE SEARCH FOR "FRENCH NEWSPAPER", WE TAKE
A LOOK AT THE LOGS.
WE COMPUTE THE SO CALLED DWELL TIME OF A CLICK I.E. THE TIME ELAPSED BEFORE THE NEXT ACTION
DWELL TIME
DWELL TIME HAS BEEN SHOWN TO BE CORRELATED WITH THE RELEVANCE
GOOD WE HAVE A MEASURE OF RELEVANCE ! CAN WE GET AN OVERALL SCORE FOR OUR SEARCH ENGINE
NOW?
Emphasis on relevant documents
Discount per ranking
Discount Cumulative Gain
Normalized Discount Cumulative Gain
Just Normalize Between 0 and 1
PERSONALIZED RERANKING IS ABOUT REORDERING THE N-BEST RESULTS BASED ON
THE USER PAST SEARCH HISTORY
Results Obtained in the contest: !
Original NCDG 0.79056 !
ReRanked NCDG 0.80714 !!
~ Raising the rank of a relevant ( relevancy = 2) result from Rank #6 to Rank #5 on each query
~ Raising the rank of a relevant ( relevancy = 2) result from Rank #6 to Rank #2 in 20% of the queries
Equivalent To
How they did it
Simple, point wise approach
Session 1 Session 2 ....0
1
2
For each (URL, Session) predict relevance (0,1 or 2)
Supervised Learning on History
We split 27 days of the train dataset 24 (history) + 3 days (annotated). !
Stop randomly in the last 3 days at a “test" session (like Yandex)
Train Set (24 history)
Train Set (annotation) Test Set
Working with a ML workflow collaboratively
Features Construction : Team Member work independantly
Learning : Team Member work independantly
Split Train & Validation
Features on 30 days
Labelled 30 days data
!regression : we keep the hierarchy between the classes, but optimizing NDCG is cookery. classification : we lose the hierarchy but we can optimize the NDCG (more and that later)
REGRESSION or CLASSIFICATION
According to P. Li, C. J. C. Burges, and Q. Wu. Mcrank: Learning to rank using multiple classification and gradient boosting. In NIPS, 2007. Classification outperforms regression.
!
Compute the probabilities that P(relevance = X)
Build a sorted list
!
Sort by !
P(Relevance=1) + 3 P (Relevance=2)
Hence order by decreasing
Hence order by P(Relevance=1) + 3 P (Relevance=2)
P. Li, C. J. C. Burges, and Q. Wu. Mcrank: Learning to rank using multiple classification and gradient boosting. In NIPS, 2007.
get slightly better results with linear weighting.
Features
FIRST OF ALL THE RANK In this contest, the rank is both
The rank that has been displayed to the user THE DISPLAY RANK
!The rank that is computed by Yandex using
PageRank, non-personalized log analysis?, TF-IDF, and machine learning etc.
THE NON-PERSONALIZED RANK
RANK AS feature
Digression
THE PROBLEM!WITH RERANKING
53% OF THE COMPETITORS COULD NOT IMPROVE THE BASELINE
Worse 53%
Better 47%
1. compute non-personalized rank 2. select 10 best hits and serves them in order 3. re-rank using log analysis. 4. put new ranking algorithm in prod (yeah right!) 5. compute NDCG on new logs 6. … 7. Profits !!
IDEAL
1. compute non-personalized rank 2. select 10 bests hits 3. serve 10 bests hits ranked in random order 4. re-rank using log analysis, including non-personalized rank as a
feature 5. compute score against the log with the former rank
REAL
Users tend to click on the first few urls. User satisfaction metric is influenced by the display rank. Our score is not aligned with our goal.
PROBLEM
We cannot discriminate the effect of the signal of the non-personalized rank from effect of the display rank
PROMOTES OVER CONSERVATIVE RE-RANKING POLICY
Even if we know for sure that the url with rank 9 would be clicked by the user if it was presented at rank 1, it would be probably a bad idea to rerank it to rank 1 in this contest.
Average per session of the max position jump
end digression
Revisits (Query-(User)-URL) features and variants Query Features Cumulative Features User Click Habits Collaborative Filtering Seasonality
FEATURES
!In the past, when the user was displayed this url, with the exact same query
what is the probability that :
REVISITS
• satisfaction=2 • satisfaction=1 • satisfaction=0 • miss (not-clicked) • skipped (after the last click)
5 Conditional Probability Features
1 An overall counter of display 4 mean reciprocal rank (kind of the harmonic mean of the rank) 1 snippet quality score (twisted formula used to compute snippet quality)
11 Base Features
• (In the past|within the same sesssion), • (with this very query | whatever query | a subquery | a super query) • and was offered (this url/this domain)
MANY VARIATIONSX2X 3X 2
12 variants
With the same user
Without being the same user ( URL - query features)
• Same Domain • Same URL • Same Query and Same URL
3 variants
15 Variants X 11 Base Features
165 Features
ADDITIVE SMOOTHINGhttp://fumicoton.com/posts/bayesian_rating
• book A : 1 rating of 5. Average rating of 5. • book B : 50 ratings. Average rating of 4.5
In our case to evaluate the probability that a (URL|query) should have a label l, under predicate P:
CUMULATIVE FEATURES
Aggregate the features of the URL above in the ranking list
Rationale : If a URL above is likely to be clicked, those below are likely to be missed
QUERY FEATURES
Click entropy number of time it has been queried for number of terms average position within in session average number of occurences in a session MRR of its clicks
How complex and ambiguous is a query ?
USER FEATURESWhat are the users habits ?
Click entropy User click rank counters
Rank {1, 2} clicks Rank {3, 4, 5} clicks Rank {6,7,8,9,10 } clicks
Average number of terms Average number of different terms in a session Total number of queries issued by the user
SEASONALITYWhat day is monday ?
COLLABORATIVE FILTERING (ATTEMPT)
User / Domain interaction matrix. FunkSVD Algorithm
Simon Funkhttp://sifter.org/~simon/journal/20061211.html
https://github.com/commonsense/divisi/blob/master/svdlib/_svdlib.pyxCython implementation
Marginal increase 5.10^-5 of the NCDG !
Why ?
learning
Short Story
Point Wise, Random Forest, 30 Features, 4th Place (*)
List Wise , LambdaMART, 90 Features, 1st Place (*)
(*) A Yandex “PaceMaker" Team was also displaying results on the leaderboard and were at the first place during the whole competition even if not officially contestant
Trained in 2 days, 1135 Trees
Optimize & Train in ~ 1 hour (12 cores), 24 trees
Lambda Mart
From RankNet to LambdaRank to LambdaMART: An Overview
Christopher J.C. Burges
Microsoft Research Technical Report MSR-TR-2010-82
LambdaMART = LambdaRank + MART
Lambda RankOriginal Ranking Re Ranked
13 errors 11 errors
High Quality Hit
Low Quality Hit
Rank Net Gradient
LambdaRank "Gradient"
From RankNet to LambdaRank to LambdaMART: An Overview
Christopher J.C. Burges - Microsoft Research Technical Report MSR-TR-2010-82
Grid SearchWe are not doing typical classification here. It is extremely important to perform grid
search directly against NDCG final score.
NDCG “conservatism” end up with large “min samples per leaf” (between 40 and 80 )
Feature Selection
Top-Down approach : Starting from a high number of features, iteratively removed subsets of features. This approach led to the subset of 90 features for the LambdaMart winning solutions
(Similar strategy now implemented by sklearn.feature_selection.RFECV)
! Bottom-up approach : Starting from a low number of features, add the
features that produce the best marginal improvement. Gave the 30 features that lead to the best solution with the point-wise approach.
Top Features
References
http://sourceforge.net/p/lemur/wiki/RankLib/Ranklib ( Implementation of LambdaMART)
These Slides http://www.slideshare.net/Dataiku
Learning to rank using multiple classification and gradient boosting.
P. Li, C. J. C. Burges, and Q. Wu. Mcrank - In NIPS, 2007
From RankNet to LambdaRank to LambdaMART: An Overview
Christopher J.C. Burges - Microsoft Research Technical Report MSR-TR-2010-82
http://fumicoton.com/posts/bayesian_ratingBlog Post About Additive Smoothing
Blog Posts about the solution
Contest Url
Paper with Detailed Description
http://blog.kaggle.com/2014/02/06/winning-personalized-web-search-team-dataiku/http://www.dataiku.com/blog/2014/01/14/winning-kaggle.html
http://research.microsoft.com/en-us/um/people/nickcr/wscd2014/papers/wscdchallenge2014dataiku.pdf
https://www.kaggle.com/c/yandex-personalized-web-search-challenge
Research Papers
References
Random ThoughtsDependancy analysis and comparing rank with predictive “relevance" could help determine general cases where the existing engine is not relevant enough How does it compare to a pure statistical approach ? !Applying personalisation technique this way might not be practical because of the amount of live information to be maintained (in real-time) about users (each query, each click) to perform actionnable predictions How could a machine learning challenge enforce this kind of constraints? Is data science a science, a sport or a hobby. Newcomers can discover a field, improve existing results, and seemingly obtain incrementally more effective results, with little plateau effect ! Are we just at the very beginning non-industrial era of this discipline?