Using Simple Machine Learning Models in a New Ads Manager
•
0 likes•50 views
Ruth Garcia presented on using simple machine learning models in an ads manager. Online advertising spending for mobile has grown significantly, with a 76.8% compound annual growth rate for mobile compared to 15.4% overall. The ads manager aims to balance increasing revenue and engaging users. Various machine learning models were considered for click prediction, including logistic regression, random forests, and neural networks. Challenges addressed categorical values through one-hot encoding and hashing tricks. Model performance was evaluated offline using metrics like precision at 1, mean reciprocal rank, and AUC. The talk concluded with lessons on starting lean, communicating machine learning requirements upfront, and balancing exploitation and exploration in ads delivery.
Recommended
Recommended
More Related Content
Similar to Using Simple Machine Learning Models in a New Ads Manager
Similar to Using Simple Machine Learning Models in a New Ads Manager (20)
More from Ruth Garcia Gavilanes
More from Ruth Garcia Gavilanes (10)
Recently uploaded
Recently uploaded (20)
Using Simple Machine Learning Models in a New Ads Manager
- 1. Using Simple Machine Learning Models in a New Ads Manager Ruth Garcia The Data Science Summit – Sept 12, 2018 London, UK
- 2. Online Advertising for Mobile $5.7 $5.4 $6.2 $7.7 $8.1 $8.6 $8.9 $9.9 $8.7 $9.4 $0.7 $1.6 $2.8 $4.4 $8.2 $11.4 $5.7 $5.4 $6.2 $7.7 $8.8 $10.3 $11.7 $14.3 $16.9 $20.8 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 Mobile Non-Mobile 15.4% Overall CAGR 76.8% Mobile CAGR – Source: IAB/PwC Internet Ad Revenue Report, HY 2017 – * CAGR: Compound Annual Growth Rate
- 3. Find balance: Increase our revenue and engage users Revenue Engaged user
- 4. Delivery of Ads Data ownership ? Technical difficulties Black box algorithms Skyscanner Ads Manager Solution External Ads Managers
- 5. Click prediction algorithm Goal: Click prediction model Test: Is it better than random? Rankingalgorithm Candidates
- 6. Cloud services and tools at Skyscanner Languages AWS services Batch S3
- 7. Overview of an Advertising System
- 8. Expectation vs. reality Reality Not flexible but fast and easier to implement. Tensorfiow • Optimization technique • Embeddings • Crossed columns • Hashing Features: • User history, • User features, • Route features , • Ad features with, colors, text
- 9. Challenges: Which model to use? Model Possibilities (easy to read in node.js): • Logistic regression • Random Forest : gets lost • Neural networks: too slow hard to put it in json Solvers: • Logistic regression: Liblinear, sag • Train data in batches Train all data at once SGDClassifier Saves memory Gridsearch for hyperparameteres
- 10. Challenges: Categorical values Pros: • No collisions • Inverse mapping Creatives C1 C2 C3 C1 C2 C3 1 0 0 0 1 0 0 0 1 Creatives C1 C2 C3 C4 C1 C2 C3 C4 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 Cons: • Need to know all values in advance • Not good for online learning • Keep dictionary in prod One hot encoding:
- 11. Challenges: Categorical values id features 123 creative1, advertiser2,mobile, etc. 321 creative2, advertiser4,mobile, etc. id Feat_1 Feat_2 Feat_3 …. Feat_k 123 0.1 0 1 …. 0 321 0.5 0 0 …. 1 Hashing Trick: map data of arbitrary sizes to data of a fixed size Pros: • Memory efficient • Online learning • No dictionary Cons: • No inverse mapping • Hash collisions
- 12. Machine Learning Performance: offline Precision at 1: based on target groups. Mean Reciprocal Rank: order of ranked ads AUC: if caring about ranking Log-Loss: if caring about the value of CTR Other metrics :
- 13. Optimizing evaluation metric Updating model based on different sampling methods and training days. 2 3 4 5 6 7 Histogram of training days 6/ 4/ 18 6/ 11/ 18 6/ 18/ 18 6/ 25/ 18 7/ 2/ 18 7/ 9/ 18 7/ 16/ 18 7/ 23/ 18 7/ 30/ 18 8/ 6/ 18 8/ 13/ 18 8/ 20/ 18 8/ 27/ 18 9/ 3/ 18 AUC over time Best AUC Worst AUC
- 14. Satisficing metric: Precision at 1 Choose best AUC conditioned of precision at 1 better than random 6/ 4/ 18 6/ 11/ 18 6/ 18/ 18 6/ 25/ 18 7/ 2/ 18 7/ 9/ 18 7/ 16/ 18 7/ 23/ 18 7/ 30/ 18 8/ 6/ 18 8/ 13/ 18 8/ 20/ 18 8/ 27/ 18 9/ 3/ 18 Precision at 1: Satisficing metric pr ecisi on_at _1 r andom _pr ecisi on_at _1
- 15. The road ahead: Balancing exploitation and exploration Choose ad based on ONLY CTR Choose ad based on OTHER criteria – Most common approaches: – ! − #$%%&' – ! − &%($%)*+,#
- 16. Learnings 1. Start lean to prove the value of your Machine Learning project 2. Speak up front since the beginning about the benefits and requirements of using ML in the product (talk about time and costs) 3. If you have problems with dimensionality, explore different ways of optimizing your resources, e.g., mini batch, hashing trick. 4, Advertising systems are very dynamic so be aware how often you need to update the model. Eng.