BEHAVIORAL TARGETING IN ON-LINE ADVERTISING: AN EMPIRICAL STUDY
AUTHORS:JOANNA JAWORSKAMARCIN SYDOWIN DEFENSE: XILING SUN & ARINDAM PAUL
INTRODUCTION Internet Economy is driven by Advertising
Search-based Ads(40%) Display Ads (22%) Classifieds (17%)
The revenue comes from whether user is to click on a ad or not Depends on degree of match between ad and
user' s context This kind of matching is called “targeting”
and forms a motivation for this paper
BEHAVIORAL TARGETING We need to automatically decide based
on the statistics of the users' web browsing history
Behavioral Targeting has a great potential in improving the performance of ad system
Experiments in this paper do not constitute any serious threat on users' privacy User represented by cookies
The General Model Each user is identified by a cookie and a
set of attributes U U: 13 different web page categories Each visit of the web page will increase the
corresponding category by 1
The format of some rows of profile data:
The General Model A model that can be represented as function
fc(U) = p [0, 1]∈ The potential relevance of the ad c presented to the user
described by the profile U. Decision whether to present the ad c to a user visiting the page fc(U) > θc, for some threshold θc which can be tuned
experimentally. Current model is simple. Only a single ad is considered at a
time
CTR (click-through rate) is used to evaluate performance higher CTR of the presented ad, the higher revenue of the ad-
serving system
Design of Experiments data comes from real impressions of ads
different data processing
Design of Experiments different Machine-Learning algorithms
different evaluation metrics
Design of Experiments Recall and Precision
Consider an example information request I (of a test reference collection) and its set R of relevant documents.
Let A be the answer set generated by retrieval strategy. Let |Ra| be the number of documents in the intersection of the sets
R and A Recall is the fraction of the relevant documents (the set R) which
has been retrieved, i.e. Recall = |Ra| / |R| Precision is the fraction of the retrieved documents (the set A)
which is relevant, i.e. Precision = |Ra| / |A|
Experimental Results Comparison of Various Algorithms and
Attribute Transformations
Experimental Results The Choice of the Training Sample
10%all − 1 − smp0 10%all 20%all
Experimental Results Observations
it is hard to find any clear relationship between the classification algorithm or data preprocessing technique applied and the performance.
the applied model of adaptive behavioral targeting seems to be generally successful
Different training set did not influence result
Contributions present an experimental framework for testing
and evaluating various factors propose a general adaptive behavioral
targeting model which is generally successful in practice
a preliminary comparison between a couple of classification algorithms and attribute-preprocessing techniques is made and reported
the evaluation is made on unique, large industrial datasets, the first reported evaluations made on real datasets
Conclusions although a very simple model, this model is nonetheless
successful It generally increase the precision value (hit rate).
no clear conclusion about which algorithms are better this is the initial work at this area
decide whether to present a single ad an obvious simplification of the real situation plan to extend the model to take into account multiple as
candidates
this work provides clear directions which all have formed foundations for future work
Further Work introduce temporal dimension additional category-based attributes
specifying the times spent on each of the categories (work-days and week-days)
introduce 2-fold profile : long & short term clustering users or advertising different (larger and balanced) training set extend the model such that it endlessly
adapts to the users and their behavior
Impact on future research This is kind of a seminal work in the area of
Behavioral Targeting in Advertising. It has motivated many future works in this direction
Tomarchio et al.'s work on developing data-driven behavioral algorithms for online ads is directly inspired from this work.
Trzcinski et al. also took cue from this paper on their work on analyzing privacy in mobile ads.
Wang et al.'s work on“Understanding Network and User-Targeting Properties of Web Advertising Networks”is also inspired from this work.
Thank you