Digital Attribution Modeling Using Apache Spark-(Anny Chen and William Yan, Adobe)
6 likes6,313 views
This document summarizes an algorithmic digital attribution model implemented using Spark. It begins with an overview of digital attribution and how algorithmic models can determine attribution weights. It then discusses how the model was implemented using Spark, including data processing, model building, and attribution calculations. Key lessons learned are around memory management, iterative computation, and error handling when working with Spark.
Digital Attribution Modeling Using Apache Spark-(Anny Chen and William Yan, Adobe)
- 1. Algorithmic Digital Attribution Using Spark Anny (Yunzhu) Chen and William (Zhenyu) Yan Adobe
- 2. Agenda • Digital attribution • Algorithmic digital attribution • Spark implementation • Lessons learned
- 3. Path of media touches to conversion Display" • See a display ads of Adobe Creative Cloud" Paid Search" • Search the term “Creative Cloud” and click the promoted link" Organic Search" • Search the term “Adobe Creative Cloud” and click an organic link" Signup" • Signup for a free trial" Email" • Receive an email from Adobe" Subscribe to Adobe Creative Cloud (positive) Not subscribe to Adobe Creative Cloud (negative) A customer may receive various kinds of media touch points before deciding whether to subscribe to a product (conversion) or not
- 4. What’s digital attribution? 4 • A digital attribution model determines how credits for conversions are assigned to media touch points • It is quite important in performance monitoring and budget planning Display" • Saw a display ads of Adobe Creative Cloud" • 0.1" Paid Search" • Search the term “Creative Cloud” and click the promoted link" • 0.2" Organic Search" • Search the term “Adobe Creative Cloud” and click an organic link" • 0.2" Signup" • Signup for a free trial" • 0.2" Email" • Received an email from Adobe" • 0.3" Convert!
- 5. Digital attribution model at a glance Models! Consider all media touch points! Time decay! Data driven! (vs. rule based)! Last interaction, first interaction, last paid search click! no" no" no" Linear! yes" no" no" Time decay! yes" yes" no" Position based! yes" no" no" Algorithmic! yes" yes " yes"
- 6. Algorithmic Attribution • Rule based: predetermined weights based on rules • Algorithmic: machine learning or statistical models are used to determine the weights 6
- 7. Algorithmic Attribution Modeling The attribution model is based on a combination of • Distributed-lag econometric model • Discrete-time survival model Some highlights: • The basic idea is to compare the media touches in positive paths vs. those in negative paths and thus infer their effects • Logistic link function • Time decay parameters to address time-decaying of media effects • Fit media touch effects and decay parameters simultaneously • Constraints on coefficients (combining with rules) • Stratified sampling • Bias reduction: – Using control variables, such as duration of exposure – Causal modeling • Maximum likelihood estimation
- 8. Tokenization • Tokenization is used to group neighboring events together • why is tokenization needed?
- 14. Tokenization DISPLAY" DISPLAY" DISPLAY" DISPLAY" DISPLAY" DISPLAY" • The effect of a media touch is hardly doubled or tripled if it is sent repeatedly to a customer twice or three times in a short period • Tokenization is used to group neighboring events together • If you see the same advertisement 5 times in one hour, will you be impressed 5 times or just have the impression “I saw that ad”?"
- 15. Why Spark? Raw data" Big data tools" Hive, Pig, Impala, Presto…" Big data machine learning tool" Spark" Mahout" Python, R, Java, C++ …" Big data Small data Rule based Algorithmic Iterative update
- 16. Data process and stitching Event" • Media touch events; conversion events; other user behavior events" • Data fields: customer id , time stamp, event type, campaign id and etc." Path" • Stitch the events of the same customer by “id” " • Generate both positive paths and negative paths" Model" • Each path is a record" • Positive/negative label is the outcome"
- 17. Algorithm evolution First model" • Fix time decay" • Logistic regression using Mllib logistic regression with SGD" Second model" • Include both regression coefficients and decay parameters in the model" • Optimize alternatively in each iteration" • Customize and extend the original Mllib logistic regression" Third model " • Second model + Causal modeling (matched sample)"
- 18. Implementation architecture Data storage: s3" Computing platform: Spark standalone cluster on AWS (Amazon Web Service)" Monitor: web UI provided by Spark standalone cluster" Server side encryption Bastion host
- 19. Model building" Data processing and tokenization" Generate paths" Parallel algorithm" Save model to S3" Attribution" Data processing and tokenization" Retrieve model from S3" Generate positive paths" Attribution with model" Implementation Detailed Workflow
- 20. Lessons learned • Memory management – Each record was transformed from String-based to Byte-based using a hash function • Tremendously increased the speed • Reduced shuffle size in the next groupBy step
- 21. Lessons learned • Memory management – Write separate classes for model training and attribution computation • Model training needs complex transformation and intensive iteration • Attribution computation needs more information from each objects
- 22. Lessons Learned (cont’d) • Cache before iterative computation – Only cache the RDDs right before entering the model – Unpersist unused RDDs to save space • Clear jobs of stopped apps in workers from time to time – Check and clean ~/spark/work folder in workers – Check and clean /mnt/spark folder in workers if necessary • Be careful when dealing with unserialized Java objects
- 23. Lessons Learned (cont’d) • Errors of one line of code doesn’t necessarily come from that line – Spark is lazy evaluation – Errors arise at action step, but may come from the previous transformation steps • Adjustment of step size – Time decay must be between 0 and 1 – After the matching, sample size decreased dramatically, the step size has to be adjusted accordingly
- 24. Thank you!