Ingmar Weber

Personalized, interactive tag recommendation for flickr

We study the problem of personalized, interactive tag recommendation for Flickr: While a user enters/selects new tags for a particular picture, the system suggests related tags to her, based on the tags that she or other people have used in the past along with (some of) the tags already entered. The suggested tags are dynamically updated with every additional tag entered/selected. We describe a new algorithm, called Hybrid, which can be applied to this problem, and show that it outperforms previous algorithms. It has only a single tunable parameter, which we found to be very robust.n Apart from this new algorithm and its detailed analysis, our main contributions are (i) a clean methodology which leads to conservative performance estimates, (ii) showing how classical classification algorithms can be applied to this problem, (iii) introducing a new cost measure, which captures the effort of the whole tagging process, (iv) clearly identifying, when purely local schemes (using only a users tagging history) can or cannot be improved by global schemes (using everybodys tagging history).

Purely URL-based topic classification

Given only the URL of a web page, can we identify its topic? This is the question that we examine in this paper. Usually, web pages are classified using their content, but a URL-only classifier is preferable, (i) when speed is crucial, (ii) to enable content filtering before an (objection-able) web page is downloaded, (iii) when a pages content is hidden in images, (iv) to annotate hyperlinks in a personalized web browser, without fetching the target page, and (v) when a focused crawler wants to infer the topic of a target page before devoting bandwidth to download it. We apply a machine learning approach to the topic identification task and evaluate its performance in extensive experiments on categorized web pages from the Open Directory Project (ODP). When training separate binary classifiers for each topic, we achieve typical F-measure values between 80 and 85, and a typical precision of around 85. We also ran experiments on a small data set of university web pages. For the task of classifying these pages into faculty, student, course and project pages, our methods improve over previous approaches by 13.8 points of F-measure.

Rethinking the ESP game

The ESP Game was designed to harvest human intelligence to assign labels to images - a task which is still difficult for even the most advanced systems in image processing. However, the ESP Game as it is currently implemented encourages players to assign obvious labels, which can be easily predicted given previously assigned labels.n We present a language model which can assign probabilities to the next label to be added. This model is then used in a program, which plays the ESP game without looking at the image. Even without any use of the actual image, the program manages to agree with the randomly assigned human partner on a label for 69% of all images, and for 81% of images which have at least one off-limits term assigned to them.n We discuss how the scoring system and the design of the ESP game can be improved to encourage users to add less predictable labels, thereby improving the quality of the collected information.

Personalized tag suggestion for flickr

We present a system for personalized tag suggestion for Flickr: While the user is entering/selecting new tags for a particular picture, the system is suggesting related tags to her, based on the tags that she or other people have used in the past along with (some of) the tags already entered. The suggested tags are dynamically updated with every additional tag entered/selected. We describe three algorithms which can be applied to this problem. In experiments, our best-performing method yields an improvement in precision of 10-15% over a baseline method very similar to the system currently used by Flickr. Our system is accessible at http://ltaa5.epfl.ch/flickr-tags/.n To the best of our knowledge, this is the first study on tag suggestion in a setting where (i) no full text information is available, such as for blogs, (ii) no item has been tagged by more than one person, such as for social bookmarking sites, and (iii) suggestions are dynamically updated, requiring efficient yet effective algorithms.

The stability of the h-index

Over the last years the h-index has gained popularity as a measure for comparing the impact of scientists. We investigate if ranking according to the h-index is stable with respect to (i) different choices of citation databases, (ii) normalizing citation counts by the number of authors or by removing self-citations, (iii) small amounts of noise created by randomly removing citations or publications and (iv) small changes in the definition of the index. In experiments for 5,283 computer scientists and 1,354 physicists we show that although the ranking of the h-index is stable under most of these changes, it is unstable when different databases are used. Therefore, comparisons based on the h-index should only be trusted when the rankings of multiple citation databases agree.

Sponsored search, market equilibria, and the Hungarian Method

Matching markets play a prominent role in economic theory. A prime example of such a market is the sponsored search market. Here, as in other markets of that kind, market equilibria correspond to feasible, envy free, and bidder optimal outcomes. For settings without budgets such an outcome always exists and can be computed in polynomial-time by the so-called Hungarian Method. Moreover, every mechanism that computes such an outcome is incentive compatible. We show that the Hungarian Method can be modified so that it finds a feasible, envy free, and bidder optimal outcome for settings with budgets. We also show that in settings with budgets no mechanism that computes such an outcome can be incentive compatible for all inputs. For inputs in general position, however, the presented mechanism-as any other mechanism that computes such an outcome for settings with budgets-is incentive compatible.

Camera brand congruence in the Flickr social graph

Given that my friends on Flickr use cameras of brand X, am I more likely to also use a camera of brand X? Given that one of these friends changes her brand, am I likely to do the same? These are the kind of questions addressed in this work. Direct applications involve personalized advertising in social networks.n For our study we crawled a complete connected component of the Flickr friendship graph with a total of 67M edges and 3.9M users. Camera brands and models were assigned to users and time slots according to the model specific meta data pertaining to their images taken during these time slots. Similarly, we used, where provided in a users profile, information about a users geographic location and the groups joined on Flickr.n Our main findings are the following. First, a pair of friends on Flickr has a significantly higher probability of being congruent, i.e., using the same brand, compared to two random users (27% vs. 19%). Second, the degree of congruence goes up for pairs of friends (i) in the same country (29%), (ii) who both only have very few friends (30%), and (iii) with a very high cliqueness (38%). Third, given that a user changes her camera model between March-May 2007 and March-May 2008, high cliqueness friends are more likely than random users to do the same (54% vs. 48%). Fourth, users using high-end cameras are far more loyal to their brand than users using point-and-shoot cameras, with a probability of staying with the same brand of 60% vs 33%, given that a new camera is bought. Fifth, these expert users brand congruence reaches 66% (!) for high cliqueness friends.n To the best of our knowledge this is the first time that the phenomenon of brand congruence is studied for hundreds of thousands of users and over a period of two years.

Bidder optimal assignments for general utilities

We study the problem of matching bidders to items where each bidder i has general, strictly monotonic utility functions ui,j(pj) expressing his utility of being matched to item j at price pj. For this setting we prove that a bidder optimal outcome always exists, even when the utility functions are non-linear and non-continuous. We give sufficient conditions under which every mechanism that finds a bidder optimal outcome is incentive compatible. We also give a mechanism that finds a bidder optimal outcome if the conditions for incentive compatibility are satisfied. The running time of this mechanism is exponential in the number of items, but polynomial in the number of bidders.

Sponsored Search, Market Equilibria, and the Hungarian Method

Two-sided matching markets play a prominent role in economic theory. A prime example of such a market is the sponsored search market where

Bidder Optimal Assignments for General Utilities

n