1. 1 Estimating Rates of Rare Events at Multiple Resolutions Deepak Agarwal Andrei Broder Deepayan Chakrabarti Dejan Diklic Vanja Josifovski Mayssam Sayyadian

2. 2 Estimation in the “tail” Contextual Advertising  Show an ad on a webpage (“impression”)  Revenue is generated if a user clicks  Problem: Estimate the click-through rate (CTR) of an ad on a page Most (ad, page) pairs have very few impressions, if any, and even fewer clicks  Severe data sparsity

3. 3 Estimation in the “tail” Use an existing, well-understood hierarchy  Categorize ads and webpages to leaves of the hierarchy  CTR estimates of siblings are correlated  The hierarchy allows us to aggregate data Coarser resolutions  provide reliable estimates for rare events  which then influences estimation at finer resolutions

4. 4 System overview Retrospective data [URL, ad, isClicked] Crawl URLs Classify pages and ads Rare event estimation using hierarchy a sample of URLs Impute impressions, fix sampling bias

5. 5 Sampling of webpages Naïve strategy: sample at random from the set of URLs  Sampling errors in impression volume AND click volume Instead, we propose:  Crawling all URLs with at least one click, and  a sample of the remaining URLs  Variability is only in impression volume

6. 6 Imputation of impression volume Ad classes Page classes sums to #impressions on ads of this ad class [column constraint] sums to ∑n ij + K.∑m ij [row constraint] sums to Total impressions (known) #impressions = n ij + m ij + x ij Clicked pool Sampled Non-clicked pool Excess impressions (to be imputed)

7. 7 Imputation of impression volume Level 0 Level i Page hierarchy Ad hierarchy Region = (page node, ad node) Region Hierarchy  A cross-product of the page hierarchy and the ad hierarchy Page classes Ad classes Region

8. 8 Imputation of impression volume sums to [block constraint] Level i Level i+1

9. 9 Imputing x ij Level i Level i+1 Iterative Proportional Fitting [Darroch+/1972] Initialize x ij = n ij + m ij Iteratively scale x ij values to match row/col/block constraint Ordering of constraints: top- down, then bottom-up, and repeat block Page classes Ad classes

10. 10 Imputation: Summary Given  n ij (impressions in clicked pool)  m ij (impressions in sampled non-clicked pool)  # impressions on ads of each ad class in the ad hierarchy We get  Estimated impression volume Ñ ij = n ij + m ij + x ij in each region ij of every level

11. 11 System overview Retrospective data [page, ad, isclicked] Crawl Pages Classify pages and ads Rare event estimation using hierarchy a sample of pages Impute impressions, fix sampling bias

12. 12 Rare rate modeling 1.Freeman-Tukey transform:  y ij = F-T(clicks and impressions at ij) ≈ transformed-CTR  Variance stabilizing transformation: Var(y) is independent of E[y]  needed in further modeling

13. 13 S ij S parent(ij) Rare rate modeling 2.Generative Model (Tree-structured Markov Model) y ij y parent(ij) covariates β ij variance V ij Unobserved “state” variance W ij V parent(ij) β parent(ij) W parent(ij)

14. 14 Rare rate modeling Model fitting with a 2-pass Kalman filter:  Filtering: Leaf to root  Smoothing: Root to leaf Linear in the number of regions

15. 15 Experiments 503M impressions 7-level hierarchy of which the top 3 levels were used Zero clicks in  76% regions in level 2  95% regions in level 3 Full dataset DFULL, and a 2/3 sample DSAMPLE

16. 16 Experiments Estimate CTRs for all regions R in level 3 with zero clicks in DSAMPLE Some of these regions R >0 get clicks in DFULL A good model should predict higher CTRs for R >0 as against the other regions in R

17. 17 Experiments We compared 4 models  TS: our tree-structured model  LM (level-mean): each level smoothed independently  NS (no smoothing): CTR proportional to 1/Ñ  Random: Assuming |R >0 | is given, randomly predict the membership of R >0 out of R

18. 18 Experiments TS Random LM, NS

19. 19 Experiments Enough impressions  little “borrowing” from siblings Few impressions  Estimates depend more on siblings

20. 20 Related Work Multi-resolution modeling  studied in time series modeling and spatial statistics [Openshaw+/79, Cressie/90, Chou+/94] Imputation  studied in statistics [Darroch+/1972] Application of such models to estimation of such rare events (rates of ~10 -3 ) is novel

21. 21 Conclusions We presented a method to estimate  rates of extremely rare events  at multiple resolutions  under severe sparsity constraints Our method has two parts  Imputation  incorporates hierarchy, fixes sampling bias  Tree-structured generative model  extremely fast parameter fitting

22. 22 Rare rate modeling 1.Freeman-Tukey transform  Distinguishes between regions with zero clicks based on the number of impressions  Variance stabilizing transformation: Var(y) is independent of E[y]  needed in further modeling ~~ # clicks in region r # impressions in region r

23. 23 Rare rate modeling Generative Model  S ij values can be quickly estimated using a Kalman filtering algorithm  Kalman filter requires knowledge of β, V, and W  EM wrapped around the Kalman filter filtering smoothing

24. 24 Rare rate modeling Fitting using a Kalman filtering algorithm  Filtering: Recursively aggregate data from leaves to root  Smoothing: Propagate information from root to leaves Complexity: linear in the number of regions, for both time and space filtering smoothing

25. 25 Rare rate modeling Fitting using a Kalman filtering algorithm  Filtering: Recursively aggregate data from leaves to root  Smoothing: Propagates information from root to leaves Kalman filter requires knowledge of β, V, and W  EM wrapped around the Kalman filter filtering smoothing

26. 26 Imputing x ij Z (i) Z (i+1) Iterative Proportional Fitting [Darroch+/1972] Initialize x ij = n ij + m ij Top-down: Scale all x ij in every block in Z (i+1) to sum to its parent in Z (i) Scale all x ij in Z (i+1) to sum to the row totals Scale all x ij in Z (i+1) to sum to the column totals Repeat for every level Z(i) Bottom-up: Similar block Page classes Ad classes