SlideShare a Scribd company logo
1 of 70
Download to read offline
Recommender Systems: The Art and Science of Matching Items to Users,[object Object],Deepak Agarwal,[object Object],dagarwal@yahoo-inc.com,[object Object],LinkedIn, 7th July, 2011 ,[object Object]
 Recommender Systems,[object Object],Serve the “right” item to users in an automated fashion to optimize long-term business objectives,[object Object]
Content Optimization: Match articles to users,[object Object]
   Advertising: Recommend Ads on Pages,[object Object],Display/Graphical Ad,[object Object],Contextual  Advertising,[object Object]
Shopping: Recommend Related Items to buy ,[object Object]
Recommend Movies,[object Object]
Recommend People,[object Object]
Problem Definition,[object Object],Item Inventory,[object Object],Articles, web page, ,[object Object],ads, …,[object Object], Example applications,[object Object], Content, Movie,,[object Object],Advertising,,[object Object],Shopping,,[object Object],…..,[object Object],Construct an automatedalgorithm ,[object Object],to select item(s) to show,[object Object],Get feedback ,[object Object],(click, time-spent,rating, buy,…) ,[object Object],Refine parameters of the algorithm,[object Object],Repeat (large number of times),[object Object],Optimize metric(s) of interest,[object Object],(Total clicks, Total revenue,…),[object Object],Low Marginal cost per serve, ,[object Object],         Efficient and intelligent systems can,[object Object],   provide significant improvements,[object Object],Context,[object Object],page, ,[object Object],  previous item viewed,,[object Object],…,[object Object],USER,[object Object]
Data Mining -> Clever Algorithms,[object Object],So much data, enough to process it all and process it fast? ,[object Object],Ideally, we want to learn every user-item interaction,[object Object],Number of things to learn increases with data size ,[object Object],Dynamic nature exacerbates the problem,[object Object],We want to learn things quickly in order to react fast,[object Object]
Simple Approach: Segment Users/Items,[object Object],Estimate CTR of items in each user segment,[object Object],j,[object Object],Serve most popular item in segment,[object Object],Item/item segments,[object Object],Users,[object Object],i,[object Object],CTRij = clicksij/viewsij,[object Object],User segments,[object Object]
Example Application: Yahoo! front page ,[object Object],Recommend most popular article on slot F1 (out of 30-40, editorially programmed),[object Object],Can collect data every 5 minutes,[object Object],Should be simple, just count clicks and views, right?,[object Object],Not quite!,[object Object],Today module,[object Object],F1,[object Object],F2,[object Object],F3,[object Object],F4,[object Object],NEWS,[object Object]
Simple algorithm we began with,[object Object],Initialize CTR of every new article to some high number,[object Object],This ensures a new article has a chance of being shown,[object Object],Show the most popular CTR article (randomly breaking ties) for each user visit in the next 5 minutes,[object Object],Re-compute the global article CTRs after 5 minutes,[object Object],Show the new most popular for next 5 minutes,[object Object],Keep updating article popularity over time,[object Object],Quite intuitive. Did not work! Performance was bad. Why? ,[object Object]
Bias in the data: Article CTR decays over time,[object Object],This is what an article CTR curve looked like,[object Object],We were computing CTR by cumulating clicks and views. ,[object Object],Missing decay dynamics? Dynamic growth model using a Kalman filter. ,[object Object],New model tracked decay very well, performance still bad,[object Object],And the plot thickens, my dear Watson! ,[object Object]
Explanation of decay: Repeat exposure,[object Object],User Fatigue-> CTR Decay,[object Object]
Clues to solve the mystery ,[object Object],Users seeing an article for the first time have higher CTR, those being exposed have lower,[object Object],but we use the same CTR estimate for all ?,[object Object],Other sources of bias? How to adjust for them?,[object Object],A simple idea to remove bias ,[object Object],Display articles at random to a small randomly chosen population,[object Object],Call this the Random bucket,[object Object],Randomization removes bias in data ,[object Object],(Charles Pierce,1877; R.A. Fisher, 1935),[object Object],Some other observations,[object Object],Sticking with an article for complete 5 minutes was degrading performance, many bad articles got displayed too many times,[object Object],Reaction time to display good articles was slower,[object Object]
CTR of same article with/without randomization,[object Object],Serving bucket,[object Object],Random bucket,[object Object],Decay,[object Object],Time-of-Day,[object Object]
CTR of articles in Random bucket,[object Object],Track,[object Object],Unbiased CTR, but it is dynamic. Simply counting clicks and views still didn’t won’t work well.,[object Object]
New algorithm,[object Object],Create a small random bucket which selects one out of K existing articles at random for each user visit,[object Object],Learn unbiased article popularity using random bucket data by tracking (through a non-linear Kalman filter),[object Object],    Serve the most popular article in the serving bucket,[object Object],Override rules: Don’t show an article to a user after few previous exposures, other rules (diversity, voice),….,[object Object]
Other advantages,[object Object],The random bucket ensures continuous flow of data for all articles, we quickly discard bad articles and converge to the best one,[object Object],This saved the day, the project was a success!,[object Object],Initial click-lift 40% (Agarwal et al. NIPS 08) ,[object Object],after 3 years it is 200+% (fully deployed on Yahoo! front page and elsewhere on Yahoo!), we are still improving the system,[object Object]
More Details,[object Object],Agarwal, Chen, Elango, Ramakrishnan, Motgi, Roy, Zachariah. Online models for Content Optimization, NIPS 2008,[object Object],Agarwal, Chen, Elango. Spatio-Temporal Models for Estimating Click-through Rate, WWW 2009,[object Object]
Lessons learnt,[object Object],It is ok to start with simple models that learn a few things, but beware of the biases inherent in your data,[object Object],E.g. of things gone wrong,[object Object],Learning article popularity ,[object Object],Data used from 5am-8am pst, served from 10am-1pm pst,[object Object],Bad idea if article popular on the east, not on the west,[object Object],Randomization is a friend, use it when you can. Update the models fast, this may reduce the bias,[object Object],User visit patterns close in time are similar  ,[object Object],What if we can’t afford complete randomization?,[object Object],Learn how to gamble,[object Object]
Why learn how to gamble?,[object Object],Consider a slot machine with two arms,[object Object],(unknown payoff probabilities),[object Object],p2,[object Object],p1      >,[object Object],The gambler has 1000 plays, what is the best way to experiment ?,[object Object],                       (to maximize total expected reward),[object Object], This is called the “bandit” problem, have been studied for a long time.,[object Object],Optimal solution: Play the arm that has maximum potential of being good,[object Object]
Recommender Problems: Bandits?,[object Object],Two Items: Item 1 CTR= 2/100 ; Item 2 CTR= 250/10000,[object Object],Greedy: Show Item 2 to all; not a good idea,[object Object],Item 1 CTR estimate noisy; item could be potentially better,[object Object],Invest in Item 1 for better overall performance on average,[object Object],This is also referred to as Explore/exploit problem,[object Object],Exploit what is known to be good, explore what is potentially good,[object Object],Article 2,[object Object],Article 1,[object Object],Probability density,[object Object],CTR,[object Object]
Bayes optimal solution in next 5 mins 2 articles, 1 uncertain,[object Object],Optimal allocation to uncertain article ,[object Object],Uncertainty in CTR: pseudo #views,[object Object]
More Details on the Bayes Optimal Solution,[object Object],Agarwal, Chen, Elango. Explore-Exploit Schemes for Web Content Optimization, ICDM 2009 ,[object Object],(Best Research Paper Award),[object Object]
Recommender Problems: bandits in a casino,[object Object],Items are arms of bandits, ratings/CTRs are unknown payoffs,[object Object],Goal is to converge to the best CTR item quickly,[object Object],But this assumes one size fits all (no personalization),[object Object],Personalization,[object Object],Each user is a separate bandit,[object Object],Hundreds of millions of bandits (huge casino),[object Object],Rich literature (several tutorials on the topic),[object Object],Broadly : Clever/adaptive randomization,[object Object],Our random bucket is a solution, often a good one in practice.,[object Object]
Back to the number of things to learn (curse of dimensionality),[object Object],Pros of learning things at granular resolutions,[object Object],Better estimates of affinities at event level ,[object Object],(ad 77 has high CTR on publisher 88, instead of ad 77 has good CTR on sports publisher),[object Object],Bias becomes less problematic,[object Object],The more we chop, less prone we are to aggregating dissimilar things, less biased our estimates.,[object Object],Challenges,[object Object],Too much sparsity to learn everything at granular resolutions,[object Object],We don’t have that much traffic,[object Object],E.g. many ads are not even shown on many publishers,[object Object],Explore/exploit helps but cannot do so much experimentation,[object Object],In advertising, response rates (conversion, click) are too low, further exacerbates the problem,[object Object]
Solution: Go granular but with back-off,[object Object],Too little data at granular level, need to borrow from coarse resolutions with abundant data (smoothing, shrinkage),[object Object],200/5000,[object Object],400/10000,[object Object],CTR(1) = w1(0/5),[object Object],+ w11(2/200)  +w12(40/1000),[object Object],+w121(200/5000)  +w111(400/10000),[object Object],121. Adv-id=9,[object Object],111. Bay Area,[object Object],40/1000,[object Object],2/200,[object Object],12. Pub-id=88, adv-id=9,[object Object],11.  Palo Alto,[object Object],0/5,[object Object],1.  Pub-id=88, ad-id=77, zip=Palo Alto,[object Object]
Sometimes too much data at granular level,[object Object],No need to back-off ,[object Object], CTR(1) = 100/50000,[object Object],……,[object Object],….,[object Object],12. Pub-id=88, adv-id=8,[object Object],11.  Arizona,[object Object],100/50000,[object Object],1.  Pub-id=88, ad-id=80, zip=Arizona,[object Object]
How much to borrow from ancestors?,[object Object],Learning the weights when there is little data,[object Object],Depends on heterogeneity in CTRs of small cells ,[object Object],Ancestors with similar CTR child nodes are more credible,[object Object],E.g. if all zip-codes in Bay Area have similar CTRs, more weights given to  Bay Area node,[object Object],Pool similar cells, separate dissimilar ones,[object Object],Palo Alto,[object Object],Bay Area,[object Object],Mtn View,[object Object],Las Gatos,[object Object]
Crucial issue,[object Object],Obtain grouping structures to perform effective back-off,[object Object],BUT,[object Object],How do we detect such groupings when dealing with high dimensional data?,[object Object],Billions/trillions of possible attribute combinations,[object Object],Statistical modeling to the rescue,[object Object],Art and science, requires experience. ,[object Object],Important to understand the business, the problem, the data. ,[object Object]
How do we estimate heterogeneity for a group?,[object Object],Simple example: CTR of an ad in different zip-codes,[object Object],(si, ti): i=1,…,K;  emCTRi = si /ti,[object Object],Var(emCTRi ) good measure of heterogeneity?,[object Object],Not quite, empirical estimates not good for small ti and(or) si,[object Object],Use a model,[object Object],Variance among true CTRs can be estimated in a better way using MLE/MOM ,[object Object],(Agarwal  & Chen, Latent OLAP, SIGMOD 2011),[object Object]
Two Examples of learning granular MODELS withback-off,[object Object]
Online Advertising: Matching ads to opportunities,[object Object],Pick best ads,[object Object],Ads,[object Object],Advertisers,[object Object],Ad Network,[object Object],Page,[object Object],User,[object Object],Examples:Yahoo, Google, MSN, ,[object Object],Ad exchanges(network of “networks”) …,[object Object],Opportunity,[object Object],Publisher,[object Object]
How to Select “Best” ads,[object Object],Pick best ads,[object Object],Ads,[object Object],Ad Network,[object Object],Page,[object Object],User,[object Object],Publisher,[object Object],Response rates,[object Object],(click, conversion,,[object Object],ad-view),[object Object],Bids,[object Object],conversion,[object Object],Auction,[object Object],Statistical,[object Object],model,[object Object],Select argmax f(bid,rate),[object Object],Click,[object Object],Advertisers,[object Object]
The Ad Exchange: Unified Marketplace,[object Object],Bids $0.75 via Network…,[object Object],Bids $0.50,[object Object],Bids $0.60,[object Object],Ad.com,[object Object],AdSense,[object Object],Bids $0.65—WINS!,[object Object],Has ad impression to sell --,[object Object],AUCTIONS,[object Object],… which becomes $0.45 bid,[object Object],Transparency and value,[object Object]
Advertising example ,[object Object],f(bid, rate) ---- rate is unknown, needs to be estimated,[object Object],Goal: maximize revenue, advertiser ROI,[object Object],High dimensional rate estimation,[object Object],Response obtained through interaction among few heavy-tailed categorical variables (pub, user, and ad),[object Object],#levels : could be millions and changes over time,[object Object],ad,[object Object],( pub, user),[object Object]
Data,[object Object],Features available for both opportunity and ad,[object Object],Publisher: Publisher content type ,[object Object],User: demographics, geo,…,[object Object],Ad: Industry, text/video, text (if any),[object Object],Hierarchically organized,[object Object],Publisher hierarchy: URL -> Domain -> Publisher type,[object Object],Geo hierarchy for users,[object Object],Ad hierarchy: Ad -> Campaign -> Advertiser,[object Object],Past empirical analysis (Agarwal et al, KDD 2007),[object Object],Hierarchical grouping provides homogeneity in rates,[object Object],Here, groupings available through domain knowledge,[object Object]
Model Setup,[object Object],baseline,[object Object],),[object Object],λij,[object Object],B(,[object Object],Piuj=,[object Object],xi,,[object Object],xj,[object Object],xu,,[object Object],residual,[object Object],i,[object Object],j,[object Object],Eij= ∑uB(xi ,xu,xj)  (Expected Success),[object Object],Sij~ Poisson(Eij λij),[object Object],MLE ( Sij /Eij) does not work well,[object Object],,,[object Object]
Hierarchical Smoothing of residuals,[object Object],Assuming two hierarchies (Publisher and advertiser),[object Object],Advertiser,[object Object],Pub-class,[object Object],Conv-id,[object Object],campaign,[object Object],Pub-id,[object Object],cell (i,j),[object Object],Ad-id,[object Object],(Sij, Eij, λij),[object Object]
Back-off Model,[object Object],7 neighbors,[object Object],3 blues, 4 greens,[object Object],Advertiser,[object Object],campaign,[object Object],Pub-class,[object Object],Conv-id,[object Object],Pub-id,[object Object],Ad-id,[object Object],i,[object Object],j,[object Object],(Sij, Eij, λij),[object Object],Back-off is through parameter sharing,[object Object],Blues and greens are neighbors of several reds,[object Object]
Ad- exchange (RightMedia),[object Object],Advertisers participate in different ways,[object Object],CPM (pay by ad-view),[object Object],CPC (pay per click),[object Object],CPA (pay per conversion),[object Object],To conduct an auction, normalize across pricing types,[object Object],Compute eCPM (expected CPM),[object Object],Click-based eCPM= click-rate*CPC,[object Object],Conversion-based eCPM= conv-rate*CPA,[object Object]
Data ,[object Object],Two kinds of conversion rates,[object Object],Post-Click conv-rate = click-rate*conv/click,[object Object],Post-View conv-rate = conv/ad-view,[object Object],Three response rate models,[object Object],Click-rate (CLICK), conv/click (PCC), ,[object Object],post-view conv/view (PVC),[object Object]
Datasets : Right-Media,[object Object],CLICK  [~90B training events, ~100M parameters],[object Object],Post Click Conversion(PCC) (~.5B training events,~81M parameters),[object Object],PVC – Post-View conversions (~7B events, ~6M parameters),[object Object],Cookie gets augmented with pixel, trigger conversion when user visits the landing page,[object Object],Features,[object Object],Age, gender, ad-size, pub-class, user fatigue,[object Object],2 hierarchies (publisher and advertiser),[object Object],Two baselines,[object Object],Pubid x adid [FINE] (no hierarchical information),[object Object],Pubid x advertiser [COARSE] (collapse cells),[object Object]
Accuracy: Average test log-likelihood,[object Object]
More Details,[object Object],Agarwal, Kota, Agrawal, Khanna: Estimating Rates of Rare Events with Multiple Hierarchies through Scalable Log-linear Models, KDD 2010,[object Object]
Back to Yahoo! front page,[object Object],Recommend articles:,[object Object],    Image,[object Object],    Title, summary,[object Object],    Links to other pages,[object Object],For each user visit,,[object Object], Pick 4 out of a pool of K,[object Object],Routes traffic to other pages,[object Object],2,[object Object],3,[object Object],4,[object Object],1,[object Object]
DATA,[object Object],article j with,[object Object],item featuresxj,[object Object],(keywords, content categories, ...),[object Object],Algorithm selects,[object Object],        (i,j) : response yij,[object Object],User i,[object Object],with,[object Object],user featuresxi,[object Object],(demographics,,[object Object],browse history,,[object Object],search history, …),[object Object],visits,[object Object],(rating or click/no-click),[object Object]
Bipartite Graph completion problem,[object Object],Observed Graph,[object Object],no-click,[object Object],Articles,[object Object],Articles,[object Object],Users,[object Object],Predicted,[object Object],CTR Graph,[object Object],Users,[object Object],click,[object Object]
Factor Model to estimate CTR at granular levels,[object Object],ui,[object Object],vj,[object Object],j,[object Object],i,[object Object],Item popularity,[object Object],User popularity,[object Object]
Estimating granular latent factors via back-off,[object Object],If user/item have high degree, good estimates of factors available else we need back-off,[object Object],Back-off: We use user/item features through regressions,[object Object],               Age=old     Geo=Mtn-View   Int=Ski,[object Object],Uik = G1k 1(Agei=old) + G2k 1(Geoi=Mtn-View) + G3k 1(Inti=Ski),[object Object],Weights of 8 different fallbacks using 3 parameters,[object Object]
Estimates with back-off,[object Object],For new user/article, factor estimates based on features,[object Object],For old user/article, factor estimates,[object Object],Linear combination of regression and user “ratings”,[object Object]
Estimating the back-off Regression function,[object Object],Maximize,[object Object],Integral cannot be computed in closed form, ,[object Object],approximated by Monte Carlo using  Gibbs Sampling,[object Object]
Data Example,[object Object],2M binary observations by 30K heavy users on 4K articles,[object Object],Heavy user ---- at least 30 visits to the portal in last 5 months,[object Object],Article features,[object Object],Editorially labeled category information (~50 binary features),[object Object],User features,[object Object],Demographics, browse behavior (~1K features),[object Object],Training/test split by timestamp of events (75/25),[object Object],Methods,[object Object],Factor model with regression, no online updates,[object Object],Factor model with regression + online updates,[object Object],Online model based on user-user similarity (Online-UU),[object Object], Online probabilistic latent semantic index (Online-PLSI),[object Object]
ROC curve,[object Object],Factor model: regression + online updates,[object Object],Factor model: regression only,[object Object]
More Details,[object Object],Agarwal and Chen: Regression Based Latent Factor Models, KDD 2009,[object Object]
Computation,[object Object],Both models run on Hadoop, scalable to large datasets,[object Object],For the factor models, also working on online EM ,[object Object],Collaboration with Andrew Cron, Duke University,[object Object]
Multi-ObjectivesBeyond Clicks,[object Object]
 Post-click utilities,[object Object],Recommender,[object Object],EDITORIAL,[object Object],AD SERVER,[object Object],      PREMIUM DISPLAY,[object Object],         (GUARANTEED),[object Object],      NETWORK PLUS,[object Object],        (Non-Guaranteed),[object Object],Clicks on FP links influence downstream supply distribution,[object Object],content,[object Object],   SPORTS,[object Object],NEWS,[object Object],Downstream engagement,[object Object],(Time spent),[object Object],OMG,[object Object],FINANCE,[object Object]
Serving Content on Front Page: Click Shaping,[object Object],What do we want to optimize?,[object Object],Usual: Maximize clicks (maximize downstream supply from FP),[object Object],But consider the following,[object Object],Article 1: CTR=5%, utility per click = 5 ,[object Object],Article 2: CTR=4.9%, utility per click=10,[object Object],By promoting 2, we lose 1 click/100 visits, gain 5 utils,[object Object],If we do this for a large number of visits --- lose some clicks but obtain significant gains in utility?,[object Object],E.g. lose 5% relative CTR, gain 20% in utility (revenue, engagement, etc),[object Object]
How are Clicks being Shaped ?,[object Object],AFTER,[object Object],BEFORE,[object Object],Supply distribution,[object Object],Changes,[object Object],SHAPING can happen  with respect to  multiple downstream metrics (like engagement, revenue,…),[object Object]
Multi-Objective Optimization,[object Object],n articles ,[object Object],K properties ,[object Object],m user segments,[object Object],A1,[object Object],S1,[object Object],news,[object Object],xij: variables,[object Object],known pij, dij,[object Object],A2,[object Object],S2,[object Object],finance,[object Object],…,[object Object],…,[object Object],…,[object Object],omg,[object Object],An,[object Object],Sm,[object Object],[object Object]
 Time duration of i on j: dij62,[object Object]
63,[object Object],Multi-Objective Program ,[object Object],[object Object]
Goal Programming ,[object Object]
More Details,[object Object],Agarwal, Chen, Elango, Wang: Click Shaping to Optimize Multiple Objectives, KDD 2011 (forthcoming),[object Object]
Can we do it with Advertising Revenue?,[object Object],Yes, but need to be careful.,[object Object],Interventions can cause undesirable long-term impact,[object Object],Communication between two complex distributed systems ,[object Object],Display advertising at Y! also sold as long-term guaranteed contracts,[object Object],We intervene to change supply when contract is at risk of under-delivering,[object Object],Research to be shared in the future,[object Object]
Summary,[object Object],Simple models that learn a few parameters are fine to begin with  BUT beware of bias in data,[object Object],Small amounts of randomization + fast model updates,[object Object],Clever Randomization using Explore/Exploit techniques,[object Object],Granular models are more effective but we need good statistical algorithms to provide back-off estimates,[object Object],Considering multi-objective optimization is often important,[object Object]
A modeling strategy,[object Object],Feature Engineering,[object Object],Content: IR, clustering, taxonomy, entity,.. ,[object Object],User profiles: clicks, views, social, community,..,[object Object],Online,[object Object],(Fine resolution,[object Object],Corrections),[object Object],(item, user level),[object Object],(Quick updates),[object Object],Initialize,[object Object],Offline(Logistic, GBDT,..),[object Object],Coarse and slow changing,[object Object],components,[object Object],Explore/Exploit,[object Object],(Adaptive sampling),[object Object]
Indexing for fast retrieval at runtime,[object Object],Retrieving the top-k when item inventory is large in few a milli-seconds could be challenging for complex models,[object Object],Current work (joint with Maxim Guverich),[object Object],Approximate the model by an index friendly synthetic model,[object Object],Index friendly model retrieves the top-K very fast, a second stage evaluation on top-K retrieves the top-k ( K > k),[object Object],Research to be shared in a forthcoming paper,[object Object]

More Related Content

Recently uploaded

Unleashing the power of AI in UiPath Studio with UiPath Autopilot.
Unleashing the power of AI in UiPath Studio with UiPath Autopilot.Unleashing the power of AI in UiPath Studio with UiPath Autopilot.
Unleashing the power of AI in UiPath Studio with UiPath Autopilot.DianaGray10
 
Empowering Africa's Next Generation: The AI Leadership Blueprint
Empowering Africa's Next Generation: The AI Leadership BlueprintEmpowering Africa's Next Generation: The AI Leadership Blueprint
Empowering Africa's Next Generation: The AI Leadership BlueprintMahmoud Rabie
 
Anypoint Code Builder , Google Pub sub connector and MuleSoft RPA
Anypoint Code Builder , Google Pub sub connector and MuleSoft RPAAnypoint Code Builder , Google Pub sub connector and MuleSoft RPA
Anypoint Code Builder , Google Pub sub connector and MuleSoft RPAshyamraj55
 
COMPUTER 10 Lesson 8 - Building a Website
COMPUTER 10 Lesson 8 - Building a WebsiteCOMPUTER 10 Lesson 8 - Building a Website
COMPUTER 10 Lesson 8 - Building a Websitedgelyza
 
GenAI and AI GCC State of AI_Object Automation Inc
GenAI and AI GCC State of AI_Object Automation IncGenAI and AI GCC State of AI_Object Automation Inc
GenAI and AI GCC State of AI_Object Automation IncObject Automation
 
Linked Data in Production: Moving Beyond Ontologies
Linked Data in Production: Moving Beyond OntologiesLinked Data in Production: Moving Beyond Ontologies
Linked Data in Production: Moving Beyond OntologiesDavid Newbury
 
20200723_insight_release_plan
20200723_insight_release_plan20200723_insight_release_plan
20200723_insight_release_planJamie (Taka) Wang
 
What Developers Need to Unlearn for High Performance NoSQL
What Developers Need to Unlearn for High Performance NoSQLWhat Developers Need to Unlearn for High Performance NoSQL
What Developers Need to Unlearn for High Performance NoSQLScyllaDB
 
Videogame localization & technology_ how to enhance the power of translation.pdf
Videogame localization & technology_ how to enhance the power of translation.pdfVideogame localization & technology_ how to enhance the power of translation.pdf
Videogame localization & technology_ how to enhance the power of translation.pdfinfogdgmi
 
Reference Domain Ontologies and Large Medical Language Models.pptx
Reference Domain Ontologies and Large Medical Language Models.pptxReference Domain Ontologies and Large Medical Language Models.pptx
Reference Domain Ontologies and Large Medical Language Models.pptxChimezie Ogbuji
 
Designing A Time bound resource download URL
Designing A Time bound resource download URLDesigning A Time bound resource download URL
Designing A Time bound resource download URLRuncy Oommen
 
AI Fame Rush Review – Virtual Influencer Creation In Just Minutes
AI Fame Rush Review – Virtual Influencer Creation In Just MinutesAI Fame Rush Review – Virtual Influencer Creation In Just Minutes
AI Fame Rush Review – Virtual Influencer Creation In Just MinutesMd Hossain Ali
 
IaC & GitOps in a Nutshell - a FridayInANuthshell Episode.pdf
IaC & GitOps in a Nutshell - a FridayInANuthshell Episode.pdfIaC & GitOps in a Nutshell - a FridayInANuthshell Episode.pdf
IaC & GitOps in a Nutshell - a FridayInANuthshell Episode.pdfDaniel Santiago Silva Capera
 
UiPath Studio Web workshop series - Day 8
UiPath Studio Web workshop series - Day 8UiPath Studio Web workshop series - Day 8
UiPath Studio Web workshop series - Day 8DianaGray10
 
Babel Compiler - Transforming JavaScript for All Browsers.pptx
Babel Compiler - Transforming JavaScript for All Browsers.pptxBabel Compiler - Transforming JavaScript for All Browsers.pptx
Babel Compiler - Transforming JavaScript for All Browsers.pptxYounusS2
 
KubeConEU24-Monitoring Kubernetes and Cloud Spend with OpenCost
KubeConEU24-Monitoring Kubernetes and Cloud Spend with OpenCostKubeConEU24-Monitoring Kubernetes and Cloud Spend with OpenCost
KubeConEU24-Monitoring Kubernetes and Cloud Spend with OpenCostMatt Ray
 
Spring24-Release Overview - Wellingtion User Group-1.pdf
Spring24-Release Overview - Wellingtion User Group-1.pdfSpring24-Release Overview - Wellingtion User Group-1.pdf
Spring24-Release Overview - Wellingtion User Group-1.pdfAnna Loughnan Colquhoun
 
Monitoring Java Application Security with JDK Tools and JFR Events.pdf
Monitoring Java Application Security with JDK Tools and JFR Events.pdfMonitoring Java Application Security with JDK Tools and JFR Events.pdf
Monitoring Java Application Security with JDK Tools and JFR Events.pdfAna-Maria Mihalceanu
 
UiPath Studio Web workshop series - Day 7
UiPath Studio Web workshop series - Day 7UiPath Studio Web workshop series - Day 7
UiPath Studio Web workshop series - Day 7DianaGray10
 
Retrofitting for the Built Environment - IES
Retrofitting for the Built Environment - IESRetrofitting for the Built Environment - IES
Retrofitting for the Built Environment - IESIES VE
 

Recently uploaded (20)

Unleashing the power of AI in UiPath Studio with UiPath Autopilot.
Unleashing the power of AI in UiPath Studio with UiPath Autopilot.Unleashing the power of AI in UiPath Studio with UiPath Autopilot.
Unleashing the power of AI in UiPath Studio with UiPath Autopilot.
 
Empowering Africa's Next Generation: The AI Leadership Blueprint
Empowering Africa's Next Generation: The AI Leadership BlueprintEmpowering Africa's Next Generation: The AI Leadership Blueprint
Empowering Africa's Next Generation: The AI Leadership Blueprint
 
Anypoint Code Builder , Google Pub sub connector and MuleSoft RPA
Anypoint Code Builder , Google Pub sub connector and MuleSoft RPAAnypoint Code Builder , Google Pub sub connector and MuleSoft RPA
Anypoint Code Builder , Google Pub sub connector and MuleSoft RPA
 
COMPUTER 10 Lesson 8 - Building a Website
COMPUTER 10 Lesson 8 - Building a WebsiteCOMPUTER 10 Lesson 8 - Building a Website
COMPUTER 10 Lesson 8 - Building a Website
 
GenAI and AI GCC State of AI_Object Automation Inc
GenAI and AI GCC State of AI_Object Automation IncGenAI and AI GCC State of AI_Object Automation Inc
GenAI and AI GCC State of AI_Object Automation Inc
 
Linked Data in Production: Moving Beyond Ontologies
Linked Data in Production: Moving Beyond OntologiesLinked Data in Production: Moving Beyond Ontologies
Linked Data in Production: Moving Beyond Ontologies
 
20200723_insight_release_plan
20200723_insight_release_plan20200723_insight_release_plan
20200723_insight_release_plan
 
What Developers Need to Unlearn for High Performance NoSQL
What Developers Need to Unlearn for High Performance NoSQLWhat Developers Need to Unlearn for High Performance NoSQL
What Developers Need to Unlearn for High Performance NoSQL
 
Videogame localization & technology_ how to enhance the power of translation.pdf
Videogame localization & technology_ how to enhance the power of translation.pdfVideogame localization & technology_ how to enhance the power of translation.pdf
Videogame localization & technology_ how to enhance the power of translation.pdf
 
Reference Domain Ontologies and Large Medical Language Models.pptx
Reference Domain Ontologies and Large Medical Language Models.pptxReference Domain Ontologies and Large Medical Language Models.pptx
Reference Domain Ontologies and Large Medical Language Models.pptx
 
Designing A Time bound resource download URL
Designing A Time bound resource download URLDesigning A Time bound resource download URL
Designing A Time bound resource download URL
 
AI Fame Rush Review – Virtual Influencer Creation In Just Minutes
AI Fame Rush Review – Virtual Influencer Creation In Just MinutesAI Fame Rush Review – Virtual Influencer Creation In Just Minutes
AI Fame Rush Review – Virtual Influencer Creation In Just Minutes
 
IaC & GitOps in a Nutshell - a FridayInANuthshell Episode.pdf
IaC & GitOps in a Nutshell - a FridayInANuthshell Episode.pdfIaC & GitOps in a Nutshell - a FridayInANuthshell Episode.pdf
IaC & GitOps in a Nutshell - a FridayInANuthshell Episode.pdf
 
UiPath Studio Web workshop series - Day 8
UiPath Studio Web workshop series - Day 8UiPath Studio Web workshop series - Day 8
UiPath Studio Web workshop series - Day 8
 
Babel Compiler - Transforming JavaScript for All Browsers.pptx
Babel Compiler - Transforming JavaScript for All Browsers.pptxBabel Compiler - Transforming JavaScript for All Browsers.pptx
Babel Compiler - Transforming JavaScript for All Browsers.pptx
 
KubeConEU24-Monitoring Kubernetes and Cloud Spend with OpenCost
KubeConEU24-Monitoring Kubernetes and Cloud Spend with OpenCostKubeConEU24-Monitoring Kubernetes and Cloud Spend with OpenCost
KubeConEU24-Monitoring Kubernetes and Cloud Spend with OpenCost
 
Spring24-Release Overview - Wellingtion User Group-1.pdf
Spring24-Release Overview - Wellingtion User Group-1.pdfSpring24-Release Overview - Wellingtion User Group-1.pdf
Spring24-Release Overview - Wellingtion User Group-1.pdf
 
Monitoring Java Application Security with JDK Tools and JFR Events.pdf
Monitoring Java Application Security with JDK Tools and JFR Events.pdfMonitoring Java Application Security with JDK Tools and JFR Events.pdf
Monitoring Java Application Security with JDK Tools and JFR Events.pdf
 
UiPath Studio Web workshop series - Day 7
UiPath Studio Web workshop series - Day 7UiPath Studio Web workshop series - Day 7
UiPath Studio Web workshop series - Day 7
 
Retrofitting for the Built Environment - IES
Retrofitting for the Built Environment - IESRetrofitting for the Built Environment - IES
Retrofitting for the Built Environment - IES
 

Recommender Systems: The Art and Science of Matching Items to Users - A LinkedIn open data talk by Deepak Agarwal from Yahoo Research!

  • 1.
  • 2.
  • 3.
  • 4.
  • 5.
  • 6.
  • 7.
  • 8.
  • 9.
  • 10.
  • 11.
  • 12.
  • 13.
  • 14.
  • 15.
  • 16.
  • 17.
  • 18.
  • 19.
  • 20.
  • 21.
  • 22.
  • 23.
  • 24.
  • 25.
  • 26.
  • 27.
  • 28.
  • 29.
  • 30.
  • 31.
  • 32.
  • 33.
  • 34.
  • 35.
  • 36.
  • 37.
  • 38.
  • 39.
  • 40.
  • 41.
  • 42.
  • 43.
  • 44.
  • 45.
  • 46.
  • 47.
  • 48.
  • 49.
  • 50.
  • 51.
  • 52.
  • 53.
  • 54.
  • 55.
  • 56.
  • 57.
  • 58.
  • 59.
  • 60.
  • 61.
  • 62.
  • 63.
  • 64.
  • 65.
  • 66.
  • 67.
  • 68.
  • 69.
  • 70.
  • 71.
  • 72.

Editor's Notes

  1. Focus on Today module. Publishes trendy, eclectic articles on a broad range of topics including sports, finance, entertainment etc.For each visit, select 4 to display from an inventory of K. Hundreds of millions of visits/day, ~600M visitors per month.