DeepBET Reverse-Engineering the Behavioral Targeting mechanisms of Ad Networks via Deep Learning Sotirios Chatzis Cyprus University of Technology.

Slides:



Advertisements
Similar presentations
Recommender System A Brief Survey.
Advertisements

Yinyin Yuan and Chang-Tsun Li Computer Science Department
Arnd Christian König Venkatesh Ganti Rares Vernica Microsoft Research Entity Categorization Over Large Document Collections.
Supporting End-User Access
A Graph-based Recommender System Zan Huang, Wingyan Chung, Thian-Huat Ong, Hsinchun Chen Artificial Intelligence Lab The University of Arizona 07/15/2002.
Advanced topics.
1.Accuracy of Agree/Disagree relation classification. 2.Accuracy of user opinion prediction. 1.Task extraction performance on Bing web search log with.
Face Recognition Face Recognition Using Eigenfaces K.RAMNATH BITS - PILANI.
D ISCOVERING REGULATORY AND SIGNALLING CIRCUITS IN MOLECULAR INTERACTION NETWORK Ideker Bioinformatics 2002 Presented by: Omrit Zemach April Seminar.
Social Media Mining Chapter 5 1 Chapter 5, Community Detection and Mining in Social Media. Lei Tang and Huan Liu, Morgan & Claypool, September, 2010.
Topology Generation Suat Mercan. 2 Outline Motivation Topology Characterization Levels of Topology Modeling Techniques Types of Topology Generators.
Unsupervised Learning With Neural Nets Deep Learning and Neural Nets Spring 2015.
December Marginal and Joint Beliefs in BN1 A Hybrid Algorithm to Compute Marginal and Joint Beliefs in Bayesian Networks and its complexity Mark.
Web Mining Research: A Survey
Ranking by Odds Ratio A Probability Model Approach let be a Boolean random variable: document d is relevant to query q otherwise Consider document d as.
Microarray analysis 2 Golan Yona. 2) Analysis of co-expression Search for similarly expressed genes experiment1 experiment2 experiment3 ……….. Gene i:
12 -1 Lecture 12 User Modeling Topics –Basics –Example User Model –Construction of User Models –Updating of User Models –Applications.
Overview of Web Data Mining and Applications Part I
Lucent Technologies – Proprietary Use pursuant to company instruction Learning Sequential Models for Detecting Anomalous Protocol Usage (work in progress)
SharePoint 2010 Business Intelligence Module 6: Analysis Services.
Friends and Locations Recommendation with the use of LBSN
Data Management Turban, Aronson, and Liang Decision Support Systems and Intelligent Systems, Seventh Edition.
Machine Learning. Learning agent Any other agent.
Image Recognition using Hierarchical Temporal Memory Radoslav Škoviera Ústav merania SAV Fakulta matematiky, fyziky a informatiky UK.
Genetic network inference: from co-expression clustering to reverse engineering Patrik D’haeseleer,Shoudan Liang and Roland Somogyi.
Adaptive News Access Daniel Billsus Presented by Chirayu Wongchokprasitti.
Bayesian networks Classification, segmentation, time series prediction and more. Website: Twitter:
UOS 1 Ontology Based Personalized Search Zhang Tao The University of Seoul.
Sujayyendhiren RS, Kaiqi Xiong and Minseok Kwon Rochester Institute of Technology Motivation Experimental Setup in ProtoGENI Conclusions and Future Work.
Video Tracking Using Learned Hierarchical Features
©2015 Apigee Corp. All Rights Reserved. Preserving signal in customer journeys Joy Thomas, Apigee Jagdish Chand, Visa.
Some working definitions…. ‘Data Mining’ and ‘Knowledge Discovery in Databases’ (KDD) are used interchangeably Data mining = –the discovery of interesting,
Lecture 10: 8/6/1435 Machine Learning Lecturer/ Kawther Abas 363CS – Artificial Intelligence.
Presented By :Ayesha Khan. Content Introduction Everyday Examples of Collaborative Filtering Traditional Collaborative Filtering Socially Collaborative.
Friends and Locations Recommendation with the use of LBSN By EKUNDAYO OLUFEMI ADEOLA
Reasoning in Uncertain Situations
Ihr Logo Chapter 5 Business Intelligence: Data Warehousing, Data Acquisition, Data Mining, Business Analytics, and Visualization Turban, Aronson, and Liang.
Mingyang Zhu, Huaijiang Sun, Zhigang Deng Quaternion Space Sparse Decomposition for Motion Compression and Retrieval SCA 2012.
Personalized Interaction With Semantic Information Portals Eric Schwarzkopf DFKI
WEB MINING. In recent years the growth of the World Wide Web exceeded all expectations. Today there are several billions of HTML documents, pictures and.
Unconstrained Endpoint Profiling Googling the Internet Ionut Trestian, Supranamaya Ranjan, Alekandar Kuzmanovic, Antonio Nucci Reviewed by Lee Young Soo.
Working Memory and Learning Underlying Website Structure
Collaborative Filtering Zaffar Ahmed
Search Engine using Web Mining COMS E Web Enhanced Information Mgmt Prof. Gail Kaiser Presented By: Rupal Shah (UNI: rrs2146)
Chong Ho Yu.  Data mining (DM) is a cluster of techniques, including decision trees, artificial neural networks, and clustering, which has been employed.
What It Is To Be Conscious: Exploring the Plausibility of Consciousness in Deep Learning Computers Senior Project – Philosophy and Computer Science ID.
Identifying “Best Bet” Web Search Results by Mining Past User Behavior Author: Eugene Agichtein, Zijian Zheng (Microsoft Research) Source: KDD2006 Reporter:
Unsupervised Streaming Feature Selection in Social Media
WHAT IS DATA MINING?  The process of automatically extracting useful information from large amounts of data.  Uses traditional data analysis techniques.
Multi-label Prediction via Sparse Infinite CCA Piyush Rai and Hal Daume III NIPS 2009 Presented by Lingbo Li ECE, Duke University July 16th, 2010 Note:
A field of study that encompasses computational techniques for performing tasks that require intelligence when performed by humans. Simulation of human.
Analytical Tools. 2 Analytical Tools Business Goals Tool for SDR Workload Planning & Resolution – Manage and track open/awaiting disposition SDRs Review.
(def functor BeeSpace v3). Core BSv3 Features Personalized Collections –All functions operate on virtual collections. Gene Analysis Functions –Gene Annotation,
Deep Belief Network Training Same greedy layer-wise approach First train lowest RBM (h 0 – h 1 ) using RBM update algorithm (note h 0 is x) Freeze weights.
哈工大信息检索研究室 HITIR ’ s Update Summary at TAC2008 Extractive Content Selection Using Evolutionary Manifold-ranking and Spectral Clustering Reporter: Ph.d.
Location-based Social Networks 6/11/20161 CENG 770.
Profiling: What is it? Notes and reflections on profiling and how it could be used in process mining.
1 DATA-DRIVEN SOLUTIONS. 2 KEYWORD-LEVEL SEARCH RETARGETING TARGET USERS BASED ON THEIR RECENT SEARCH HISTORY AND SEARCH QUERIES. A user performs a search.
Computational Challenges in BIG DATA 28/Apr/2012 China-Korea-Japan Workshop Takeaki Uno National Institute of Informatics & Graduated School for Advanced.
Differential Analysis on Deep Web Data Sources Tantan Liu, Fan Wang, Jiedan Zhu, Gagan Agrawal December.
Collaborative Deep Learning for Recommender Systems
HIERARCHICAL TEMPORAL MEMORY WHY CANT COMPUTERS BE MORE LIKE THE BRAIN?
Introduction to Machine Learning, its potential usage in network area,
Unsupervised Learning of Video Representations using LSTMs
Automated Experiments on Ad Privacy Settings
Unsupervised Learning and Autoencoders
Christopher A. Penfold Vicky Buchanan-Wollaston Katherine J. Denby And
Supporting End-User Access
Microarray Data Set The microarray data set we are dealing with is represented as a 2d numerical array.
Machine Learning for Space Systems: Are We Ready?
Presentation transcript:

DeepBET Reverse-Engineering the Behavioral Targeting mechanisms of Ad Networks via Deep Learning Sotirios Chatzis Cyprus University of Technology

Motivation The capability to analyze and reverse-engineer the behavioral targeting practices of online ad networks is a significant challenge. Existing approaches suffer from specific limitations: o They collect data via artificial user profiles pertaining to single interest types. However, real-world users may have diverse interest type combinations; users with multiple interests may be presented with a set of ads different from the union of the sets of ads pertaining to their individual interests. o User interests may change over time. Existing approaches do not examine whether the output of an ad network depends only on the most recently visited website, or on long temporal dynamics.

Our approach Resolves these two issues by: 1.Efficiently analyzing a very large subspace of the user behavioral patterns space. 2.Employing inference algorithms capable of capturing underlying temporal dynamics. To this end, artificial user profile creation comprises: 1.Intelligent selection of websites to populate user profiles with. 2.Creation of website sequences that represent very diverse user behavioral patterns.

Website Selection Based on the top 100 websites of various Adwords Ad Planner categories. System selects a subset of them by: 1.Utilizing a Deep Boltzmann machine (DBM) to generate high-level website content representations in an unsupervised fashion. 2.Presenting them to a nonparametric Bayesian hierarchical clustering model, namely a Bayesian Rose Tree (BRT). BRT infers both the number of clusters and their hierarchy (tree). 3.Retaining representatives of clusters at only some tree layer(s).

Profile Creation 1.Creates thousands of random sequences comprising the retained representative items. 2.Employs an algorithm for (unsupervised) extraction of high-level representations from these sequential data, namely Long Short- Term Memory Auto-Encoders (LSTM-AE). 3.Clusters the so-obtained representations via a BRT model. 4.Retains representative sequences at only some tree layer(s). Each retained (cluster) representative is utilized as a user profile to query ad nets with.

Inference The selected website sequences constitute a small set that comprises highly diverse user behavioral patterns. Thus, this set allows for effectively and efficiently exploring the space of possible user behavioral patterns, so as to robustly infer ad network behavioral targeting mechanisms. We resort to the dynamic Bayesian Probabilistic Matrix Factorization (dBPMF) algorithm. It infers: o The correlations between interest type combinations and ad network outputs. o How temporal dynamics affect the generated outputs.

Questions? Q&A after lunch…