+
Practical Lessons from Predicting
Clicks on Ads at Facebook
2014/1/27 (Tue.)
Chang Wei-Yuan @ MakeLab Lab Meeting
ADKDD‘14
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
2
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
4
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
5
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
7
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
9
+Logistic regression for linear
classifier
Stochastic Gradient Descent (SGD)
algorithm
the tunable parameters are optimized by grid
search
Bayesian online learning scheme for profit
regression
10
+
One advantages of LR over BOPR is that the
model size is half
the smaller model size may lead to better cache
locality and thus faster cache lookup
11
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
12
+Data freshness
Click prediction systems are often deployed
in dynamic environments where the data
distribution changes over time.
13
+
Batch
one option would be to have a recurring daily job
that retrains the models, possibly in batch
Concurrency
the training can be done via concurrency in a multi-
core machine with large amount of memory
16
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
17
+Online data joiner
The boosted decision trees can be trained
daily, but the linear classier can be trained in
near real-time by online learning.
Online Joiner
generates real-time training data used to train the
linear classifier via online learning
18
+Online data joiner
The boosted decision trees can be trained
daily, but the linear classier can be trained in
near real-time by online learning.
Online Joiner
generates real-time training data used to train the
linear classifier via online learning
How to label for a new instance ?
19
+Online data joiner
The boosted decision trees can be trained
daily, but the linear classier can be trained in
near real-time by online learning.
Online Joiner
generates real-time training data used to train the
linear classifier via online learning
perform a distributed stream-to-stream join on ad
impressions and ad clicks
20
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
22
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
28
+Conclusion
This has inspired a promising hybrid model
architecture for click prediction.
boosted decision trees and a linear classier
online learning method with real-time training data
29
+Outline
Introduction
Method Decision tree feature transforms
Logistic regression for linear classifier
Data freshness
Online data joiner
Experiment
Conclusion
Thought
30