Simply Data driven behavioural algorithms
2 likes1,151 views
This document presents a novel data-driven behavioral algorithm for optimizing online advertising through simply.com, an ad network developed by Dada Spa. It emphasizes the importance of conversion rates and behavioral targeting, proposing a method that utilizes user clustering and real-time campaign analysis to improve ad performance without relying on predefined segments. The results indicate significant improvements in conversion rates compared to traditional methods, showcasing the effectiveness of a purely data-driven approach.
![Data-driven behavioural algorithms for online advertising
Antonio Tomarchio
Simply.com, Dada Spa
Via della Braida 5
20100 Milano
+393351605432
Antonio.tomarchio@dada.net
Francesco Bellacci
Simply.com, Dada spa
Largo Annigoni
Firenze
Telephone number, incl. country code
Francesco.bellacci@dada.net
Filippo Privitera
Simply.com, Dada spa
Via della Braida 5
20100 Milano
Telephone number, incl. country code
Filippo.privitera@dada.net
ABSTRACT
In this paper, we describe an innovative data-driven behavioural
approach that we developed for the optimization of performance
online advertising on Simply, the new international adnetwork
developed by Dada spa.
Categories and Subject Descriptors
I.5.3 [Clustering and similarity algorithms]: Clustering
algorithms , similarity measures, multivariate statistics
General Terms
Algorithms
Keywords
Online performance advertising, conversion optimization,
advertising performance optimization, clickstream analysis, user
clustering, yield optimization, real time bidding
1. INTRODUCTION
Performance advertising is becoming the most successful
advertising model on internet and mobile internet. In this model
advertisers do not pay the visualization of their marketing
message but just a direct response from the users who visualize
the ad. In particular, the last years the CPA (cost per action)
model had strong increase. In this latter model the advertiser will
pay a commission to publisher if and only if the user who clicked
on his ad will perform a specific target action on the advertiser’s
web site such as a purchase or the subscription to a newsletter. It
is then clear that in this kind of market it is absolutely necessary a
capability to optimize advertising visualization. Publishers usually
subscribe to adnetworks which will take care of optimizing their
inventory, acting as intermediaries between them and advertisers
with the aim to maximize revenues for publishers and ROI for
advertisers. The term “conversion rate” indicates the rate of
actions on 1000 users who clicks on the campaign ads, it is the
most important measure of performance for the advertiser.
Performance advertising model is financing several internet
services such as search engines and social networks, with big
benefits for users who can use these products for free..
Simply.com is a new adnetwork developed by Dada spa based on
a proprietary set of user-centric optimization algorithms.
2. BEHAVIOURAL TARGETING
TECHNIQUES
One of most widespread buzzwords of the last two years in the
Internet advertising market is “behavioural targeting. The basic
idea is that if an advertiser is able to reach a user who seems to be
interested in the same products or who is searching for a good
which is sold by the advertiser, the probability of a click and of a
conversion on the advertiser web site will be much higher. By
reaching the interested users, advertisers will be able to increase
the ROI of their campaigns and publishers will be able to increase
the monetization of their traffic. Behavioural targeting is quite
different than more classical social-demographic targeting where
advertisers try to reach users in specific ranges of ages, gender or
annual income. Behavioural targeting can integrate this kind of
information but their primary goal is to identify in real time “the
intents” of users: what they are interested in buying.
The landscape of behavioural targeting solutions is huge, but the
general strategy of most of these platforms is to provide to
publishers a mean to create different “audience pixels” to segment
the users on their web sites. For example a publisher can create a
“sport” pixel and implement it in his sport sections. Users are then
segmented by these pixels and when they visit another page in the
network will receive campaign matching with their profiles. The
core aspect of all these solutions is that the publisher affiliated to
the behavioural targeting adnetwork will create audience pixels by
“a priori” human defined tree categories, not based on
performance data
3. A PURE DATA-DRIVEN
BEHAVIOURAL APPROACH
Our idea is to create data-driven clusters of users and then to
develop a learning algorithm where each new campaign is at the
beginning delivered randomly to all clusters until the system
“learns” which are the user clusters where the campaign is getting
the highest conversion rate.
3.1 Data Gathering
Affiliated publishers on Simply network implements a single
audience pixel. They do not have to create predefined segments
such as sports or entertainment. They just have to implement a
1X1 pixel. By this pixel we insert anonymous cookie in user
browsers and we are able to collect information such as the
campaigns they clicked on, the web pages they visited in simply
network and search queries they executed on simply affiliated](https://image.slidesharecdn.com/data-driven-behavioural-algorithms-101030132025-phpapp02/85/Simply-Data-driven-behavioural-algorithms-1-320.jpg)
![publishers. The data gathering is completely privacy consistent as
we do not track any sensible private information, neither the ip
address. For each publisher or advertiser web page they visit, by
proprietary state of the art information retrieval techniques, we are
able to extract most significant keywords contained in the above
pages.
3.2 Users Profiling
For each user we are able to have information about advertising
campaigns, web pages and search queries of interest. We analyze
the top significant keywords associated to the content he visited
and we are able to extract those ones which have the highest
occurency frequency across the corpus of documents. The User
Profiling algorithms is then able to build a profile with selected
keywords..
3.3 Similarity measure
The user profiles can be used to introduce a mathematical
similarity measure.Each user is represented as a vector in a n-
dimensional space. Each coordinate represents the weight of a
specific keyword in the user profile. This representation is then
very similar to vector space model in information retrieval, where
documents are represented as vectors in a space where each
dimension is a determined word.We can add to our representation
other dimensions given by the demographic information if
available. The vector modeling of users allow to introduce a
standard similarity measures as the Cartesian product between two
vectors.By the Cartesian product operator we can build a
similarity matrix of our users set.
3.4 Clustering Methods
The User Similarity Matrix is used to apply an automatic K-Mean
Clustering Algorithm. It would be computationally difficult to
cluster the huge amount of user profiles built over Simply
Network. As a consequence we developed the following
clustering strategy:
1. We apply the clustering methods just on the most active
users where for active we mean the fact that they clicked
on advertising campaigns and/or launched search
queries.
2. Once the clusters are built , we estimate a set of
centroids
3. We then built a classification algorithm that estimates
the distance between an user and the centroids of the
different cluster and will assign the user to the best
matching cluster
4. We classified all profiled users
A key issue of k-mean based clustering algorithms is to establish
the optimal number of clusters. We applied a feedback algorithm:
we estimated the optimal number of clusters for conversion rate
3.5 Yield Optimization
Once profiled users are clustered, Simply optimization algorithms
implement what is called in the industry “yield optimization”
strategy that can be summarized in the following steps:
1. A new advertiser campaign is uploaded on Simply
Network
2. At the beginning, the campaign is delivered completely
random on all user clusters
3. Algorithm tracks in real time on a hour-basis the
average conversion rate of the campaign across the user
clusters
4. Algorithm identifies clusters where the performance of
the campaign in term of conversion rate are highest
5. Algorithm will then delivery the campaigns just to the
users belonging to the top clusters
This delivery algorithm is completely real time and very effective.
We highlight that this strategy is radical different than standard
behavioural techniques: we do not take care at all about the
content of the advertising campaign and the content of the web
pages the user visited. We just focus on creating a similarity
between users and on clustering them. Once clusters are created ,
the delivery is led uniquely by the performances data. As the
conversion rate is the only driver, this algorithm can provide a
boost in term of performances optimization
3.6 Results
We ran several tests to compare this methodology with
competitors platforms and with non optimized impressions.We
delivered the same campaigns on the same publishers and
simultaneously by three different delivery algorithms:
1. Our cluster yield method described in this paper
2. A random non optimized method
3. A competitor platform based on standard behavioural
techniques
We executed this test on different days and with different
campaigns.We measured an average increase of conversion rate of
150% by the cluster yield method compared with non optimized
delivery.We measured an average increase of conversion rate of
60% by the cluster yield method compared with competitor
platform
3.7 Conclusions
We developed a new performance advertising behavioural method
which is radically different than standard behavioural
solutions.The main and core innovation is that the algorithm is
based on a data-driven user clustering and on a real time analysis
of campaigns across the clusters. This method is content
independent and do not require any segmentation of the web
pages by the publishers.We have strong results that support our
intuition that a pure data-driven strategy can have a much higher
optimization potential than any strategy based on human
classification of content
4. References
[1] De Liung and Jianqing Chen, 2006. Designing auctions with
past performance information. Decision Support Systems 42
(2006)
[2] Liu, C. 2010. When Machine Learning Meets the Web.
Microsoft Research Keynote Speech
[3] Jaworska, J. and Sydow, M. 2008. Behavioural Targeting in
Onlina Advertising: an Empyrical Study. Lecture Notes in
Computer Sciences (Springer, Volume 5175/2008, 2008)
[4] Muthukrishnan, S.2008. Internet Ad auctions: insights and
directions. ICALP , 2008
.](https://image.slidesharecdn.com/data-driven-behavioural-algorithms-101030132025-phpapp02/85/Simply-Data-driven-behavioural-algorithms-2-320.jpg)
Simply Data driven behavioural algorithms
- 1. Data-driven behavioural algorithms for online advertising Antonio Tomarchio Simply.com, Dada Spa Via della Braida 5 20100 Milano +393351605432 Antonio.tomarchio@dada.net Francesco Bellacci Simply.com, Dada spa Largo Annigoni Firenze Telephone number, incl. country code Francesco.bellacci@dada.net Filippo Privitera Simply.com, Dada spa Via della Braida 5 20100 Milano Telephone number, incl. country code Filippo.privitera@dada.net ABSTRACT In this paper, we describe an innovative data-driven behavioural approach that we developed for the optimization of performance online advertising on Simply, the new international adnetwork developed by Dada spa. Categories and Subject Descriptors I.5.3 [Clustering and similarity algorithms]: Clustering algorithms , similarity measures, multivariate statistics General Terms Algorithms Keywords Online performance advertising, conversion optimization, advertising performance optimization, clickstream analysis, user clustering, yield optimization, real time bidding 1. INTRODUCTION Performance advertising is becoming the most successful advertising model on internet and mobile internet. In this model advertisers do not pay the visualization of their marketing message but just a direct response from the users who visualize the ad. In particular, the last years the CPA (cost per action) model had strong increase. In this latter model the advertiser will pay a commission to publisher if and only if the user who clicked on his ad will perform a specific target action on the advertiser’s web site such as a purchase or the subscription to a newsletter. It is then clear that in this kind of market it is absolutely necessary a capability to optimize advertising visualization. Publishers usually subscribe to adnetworks which will take care of optimizing their inventory, acting as intermediaries between them and advertisers with the aim to maximize revenues for publishers and ROI for advertisers. The term “conversion rate” indicates the rate of actions on 1000 users who clicks on the campaign ads, it is the most important measure of performance for the advertiser. Performance advertising model is financing several internet services such as search engines and social networks, with big benefits for users who can use these products for free.. Simply.com is a new adnetwork developed by Dada spa based on a proprietary set of user-centric optimization algorithms. 2. BEHAVIOURAL TARGETING TECHNIQUES One of most widespread buzzwords of the last two years in the Internet advertising market is “behavioural targeting. The basic idea is that if an advertiser is able to reach a user who seems to be interested in the same products or who is searching for a good which is sold by the advertiser, the probability of a click and of a conversion on the advertiser web site will be much higher. By reaching the interested users, advertisers will be able to increase the ROI of their campaigns and publishers will be able to increase the monetization of their traffic. Behavioural targeting is quite different than more classical social-demographic targeting where advertisers try to reach users in specific ranges of ages, gender or annual income. Behavioural targeting can integrate this kind of information but their primary goal is to identify in real time “the intents” of users: what they are interested in buying. The landscape of behavioural targeting solutions is huge, but the general strategy of most of these platforms is to provide to publishers a mean to create different “audience pixels” to segment the users on their web sites. For example a publisher can create a “sport” pixel and implement it in his sport sections. Users are then segmented by these pixels and when they visit another page in the network will receive campaign matching with their profiles. The core aspect of all these solutions is that the publisher affiliated to the behavioural targeting adnetwork will create audience pixels by “a priori” human defined tree categories, not based on performance data 3. A PURE DATA-DRIVEN BEHAVIOURAL APPROACH Our idea is to create data-driven clusters of users and then to develop a learning algorithm where each new campaign is at the beginning delivered randomly to all clusters until the system “learns” which are the user clusters where the campaign is getting the highest conversion rate. 3.1 Data Gathering Affiliated publishers on Simply network implements a single audience pixel. They do not have to create predefined segments such as sports or entertainment. They just have to implement a 1X1 pixel. By this pixel we insert anonymous cookie in user browsers and we are able to collect information such as the campaigns they clicked on, the web pages they visited in simply network and search queries they executed on simply affiliated
- 2. publishers. The data gathering is completely privacy consistent as we do not track any sensible private information, neither the ip address. For each publisher or advertiser web page they visit, by proprietary state of the art information retrieval techniques, we are able to extract most significant keywords contained in the above pages. 3.2 Users Profiling For each user we are able to have information about advertising campaigns, web pages and search queries of interest. We analyze the top significant keywords associated to the content he visited and we are able to extract those ones which have the highest occurency frequency across the corpus of documents. The User Profiling algorithms is then able to build a profile with selected keywords.. 3.3 Similarity measure The user profiles can be used to introduce a mathematical similarity measure.Each user is represented as a vector in a n- dimensional space. Each coordinate represents the weight of a specific keyword in the user profile. This representation is then very similar to vector space model in information retrieval, where documents are represented as vectors in a space where each dimension is a determined word.We can add to our representation other dimensions given by the demographic information if available. The vector modeling of users allow to introduce a standard similarity measures as the Cartesian product between two vectors.By the Cartesian product operator we can build a similarity matrix of our users set. 3.4 Clustering Methods The User Similarity Matrix is used to apply an automatic K-Mean Clustering Algorithm. It would be computationally difficult to cluster the huge amount of user profiles built over Simply Network. As a consequence we developed the following clustering strategy: 1. We apply the clustering methods just on the most active users where for active we mean the fact that they clicked on advertising campaigns and/or launched search queries. 2. Once the clusters are built , we estimate a set of centroids 3. We then built a classification algorithm that estimates the distance between an user and the centroids of the different cluster and will assign the user to the best matching cluster 4. We classified all profiled users A key issue of k-mean based clustering algorithms is to establish the optimal number of clusters. We applied a feedback algorithm: we estimated the optimal number of clusters for conversion rate 3.5 Yield Optimization Once profiled users are clustered, Simply optimization algorithms implement what is called in the industry “yield optimization” strategy that can be summarized in the following steps: 1. A new advertiser campaign is uploaded on Simply Network 2. At the beginning, the campaign is delivered completely random on all user clusters 3. Algorithm tracks in real time on a hour-basis the average conversion rate of the campaign across the user clusters 4. Algorithm identifies clusters where the performance of the campaign in term of conversion rate are highest 5. Algorithm will then delivery the campaigns just to the users belonging to the top clusters This delivery algorithm is completely real time and very effective. We highlight that this strategy is radical different than standard behavioural techniques: we do not take care at all about the content of the advertising campaign and the content of the web pages the user visited. We just focus on creating a similarity between users and on clustering them. Once clusters are created , the delivery is led uniquely by the performances data. As the conversion rate is the only driver, this algorithm can provide a boost in term of performances optimization 3.6 Results We ran several tests to compare this methodology with competitors platforms and with non optimized impressions.We delivered the same campaigns on the same publishers and simultaneously by three different delivery algorithms: 1. Our cluster yield method described in this paper 2. A random non optimized method 3. A competitor platform based on standard behavioural techniques We executed this test on different days and with different campaigns.We measured an average increase of conversion rate of 150% by the cluster yield method compared with non optimized delivery.We measured an average increase of conversion rate of 60% by the cluster yield method compared with competitor platform 3.7 Conclusions We developed a new performance advertising behavioural method which is radically different than standard behavioural solutions.The main and core innovation is that the algorithm is based on a data-driven user clustering and on a real time analysis of campaigns across the clusters. This method is content independent and do not require any segmentation of the web pages by the publishers.We have strong results that support our intuition that a pure data-driven strategy can have a much higher optimization potential than any strategy based on human classification of content 4. References [1] De Liung and Jianqing Chen, 2006. Designing auctions with past performance information. Decision Support Systems 42 (2006) [2] Liu, C. 2010. When Machine Learning Meets the Web. Microsoft Research Keynote Speech [3] Jaworska, J. and Sydow, M. 2008. Behavioural Targeting in Onlina Advertising: an Empyrical Study. Lecture Notes in Computer Sciences (Springer, Volume 5175/2008, 2008) [4] Muthukrishnan, S.2008. Internet Ad auctions: insights and directions. ICALP , 2008 .