Audience Selection for On-Line Brand Advertising: Privacy-Friendly Social Network Targeting

Provost, Foster; D'Alessandro, B; Hook, Rod; Zhang, Xiaohan; Murray, Alan T.

doi:10.2139/ssrn.1852644

Cited by 51 publications

(69 citation statements)

References 20 publications

(28 reference statements)

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…We conjecture that these techniques also will be quite useful in other high-dimensional classification problems, which are becoming increasingly important to modern business. For example, it may not be obvious, but classifying web users based on the web pages they visit (Provost et al 2009) could be cast in the same framework as document classification.…”

Section: Discussion and Limitationsmentioning

confidence: 99%

Explaining Data-Driven Document Classifications

Martens

Provost

2014

MISQ

Self Cite

239

188

View full text Add to dashboard Cite

Many document classification applications require human understanding of the reasons for data-driven classification decisions: by managers, client-facing employees, and the technical team. Predictive models treat documents as data to be classified, and document data are characterized by very high dimensionality, often with tens of thousands to millions of variables (words). Unfortunately, due to the high dimensionality, understanding the decisions made by document classifiers is very difficult. This paper begins by extending the most relevant prior theoretical model of explanations for intelligent systems to account for some missing elements.The main theoretical contribution of the work is the definition of a new sort of explanation as a minimal set of words (terms, more generally), such that removing all words within this set from the document changes the predicted class from the class of interest. We present an algorithm to find such explanations, as well as a framework to assess such an algorithm's performance. We demonstrate the value of the new approach with a case study from a real-world document classification task: classifying web pages as containing objectionable content, with the goal of allowing advertisers to choose not to have their ads appear there. A second empirical demonstration on news-story topic classification uses the 20 Newsgroups benchmark dataset. The results show the explanations to be concise and document-specific, and to be capable of providing better understanding of the exact reasons for the classification decisions, of the workings of the classification models, and of the business application itself. We also illustrate how explaining documents' classifications can help to improve data quality and model performance.

show abstract

Section: Discussion and Limitationsmentioning

confidence: 99%

Explaining Data-Driven Document Classifications

Martens

Provost

2014

MISQ

Self Cite

239

188

View full text Add to dashboard Cite

show abstract

“…Earlier work on privacy breaches has shown how people can be uniquely identified using information such as postal codes, gender, and dates and places of birth (12)(13)(14), as well as the contents of search engine queries and online reviews and discussion (15)(16)(17). Other work has shown how social network structure can be exposed by analyzing anonymized versions of it (18,19) or by looking for commonalities in online behavior, such as covisitations to web sites (20) and tagging shared content with similar textual keywords (21). Our findings here differ from these studies by establishing a strong form of leakage from sparse individual information about activities in the physical world into pairwise information about links in the underlying social network.…”

Section: Discussionmentioning

confidence: 99%

Inferring social ties from geographic coincidences

Crandall

Bäckström

Cosley

et al. 2010

Proc. Natl. Acad. Sci. U.S.A.

408

314

View full text Add to dashboard Cite

We investigate the extent to which social ties between people can be inferred from co-occurrence in time and space: Given that two people have been in approximately the same geographic locale at approximately the same time, on multiple occasions, how likely are they to know each other? Furthermore, how does this likelihood depend on the spatial and temporal proximity of the co-occurrences? Such issues arise in data originating in both online and offline domains as well as settings that capture interfaces between online and offline behavior. Here we develop a framework for quantifying the answers to such questions, and we apply this framework to publicly available data from a social media site, finding that even a very small number of co-occurrences can result in a high empirical likelihood of a social tie. We then present probabilistic models showing how such large probabilities can arise from a natural model of proximity and co-occurrence in the presence of social ties. In addition to providing a method for establishing some of the first quantifiable estimates of these measures, our findings have potential privacy implications, particularly for the ways in which social structures can be inferred from public online records that capture individuals' physical locations over time.computer science | privacy | probabilistic models | social networks E very day, we make inferences about the social world from incomplete observations of events around us. A particular category of such inferences draws on co-occurrences in space and time-basing estimates of a social tie between two people on the fact that they were in the same geographic locale at roughly the same time. In addition to its intuitive accessibility, such reasoning has been employed in psychological studies of urban life (1) and legal analyses of the dangers of "guilt by association" (2, 3). These issues also arise naturally in online domains, including those that reflect spatio-temporal traces of their users' activities in the physical world. Despite the broad relevance of the underlying questions, however, there has been essentially no precise basis for quantifying the significance of these effects. Here we study this issue in an online setting and find that geographic co-occurrences can in fact have significant power in forming inferences about social ties: The knowledge that two people were proximate at just a few distinct locations at roughly the same times can indicate a high conditional probability that they are directly linked in the underlying social network, in the data we consider. Our results use publicly accessible spatial and temporal information from a large social media site to derive estimates of links in the online social network of the site. We also develop a probabilistic model to account for the high probabilities that are observed. In addition to providing a quantitative basis for the power of these inferences, our results have implications for the unintended leakage of private information via participation in such sites.Our analysis uses...

show abstract

“…The use of high dimensional raw user log data as features in classification models was introduced in earlier work by Provost et al [17] as well as Chen et al in [6]. Further work, such as that by [15] and [13] describes in finer detail how features can be constructed from raw user event data.…”

Section: Background On M6d Display Advertising and Related Workmentioning

confidence: 99%

“…Each of these criteria may be approximated with a good ranking of potential purchasers in terms of their likelihood of purchasing. These problems have been described in detail previously [17,16,19]. A primary source of feature data is a consumer's browsing history, captured as a collection of anonymized (hashed) URLs that the consumer has visited in the past.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Machine learning for targeted display advertising: transfer learning in action

Perlich¹,

D'Alessandro²,

Raeder³

et al. 2013

Mach Learn

140

View full text Add to dashboard Cite

This paper presents a detailed discussion of problem formulation and data representation issues in the design, deployment, and operation of a massive-scale machine learning system for targeted display advertising. Notably, the machine learning system itself is deployed and has been in continual use for years, for thousands of advertising campaigns (in contrast to simply having the models from the system be deployed). In this application, acquiring sufficient data for training from the ideal sampling distribution is prohibitively expensive. Instead, data are drawn from surrogate domains and learning tasks, and then transferred to the target task. We present the design of this multistage transfer learning system, highlighting the problem formulation aspects. We then present a detailed experimental evaluation, showing that the different transfer stages indeed each add value. We next present production results across a variety of advertising clients from a variety of industries, illustrating the performance of the system in use. We close the paper with a collection of lessons learned from the work over half a decade on this complex, deployed, and broadly used machine learning system.

show abstract

Audience Selection for On-Line Brand Advertising: Privacy-Friendly Social Network Targeting

Cited by 51 publications

References 20 publications

Explaining Data-Driven Document Classifications

Explaining Data-Driven Document Classifications

Inferring social ties from geographic coincidences

Machine learning for targeted display advertising: transfer learning in action

Contact Info

Product

Resources

About