Léa Laporte

Query Performance Prediction Focused on Summarized Letor Features

Query performance prediction (QPP) aims at automatically estimating the information retrieval system effectiveness for any users query. Previous work has investigated several types of pre- and post-retrieval query performance predictors; the latter has been shown to be more effective. In this paper we investigate the use of features that were initially defined for learning to rank in the task of QPP. While these features have been shown to be useful for learning to rank documents, they have never been studied as query performance predictors. We developed more than 350 variants of them based on summary functions. Conducting experiments on four TREC standard collections, we found that Letor-based features appear to be better QPP than predictors from the literature. Moreover, we show that combining the best Letor features outperforms the state of the art query performance predictors. This is the first study that considers such an amount and variety of Letor features for QPP and that demonstrates they are appropriate for this task.

TournaRank: When retrieval becomes document competition

Numerous feature-based models have been recently proposed by the information retrieval community. The capability of features to express different relevance facets (query- or document-dependent) can explain such a success story. Such models are most of the time supervised, thus requiring a learning phase. To leverage the advantages of feature-based representations of documents, we propose TournaRank, an unsupervised approach inspired by real-life game and sport competition principles. Documents compete against each other in tournaments using features as evidences of relevance. Tournaments are modeled as a sequence of matches, which involve pairs of documents playing in turn their features. Once a tournament is ended, documents are ranked according to their number of won matches during the tournament. This principle is generic since it can be applied to any collection type. It also provides great flexibility since different alternatives can be considered by changing the tournament type, the match rules, the feature set, or the strategies adopted by documents during matches. TournaRank was experimented on several collections to evaluate our model in different contexts and to compare it with related approaches such as Learning To Rank and fusion ones: the TREC Robust2004 collection for homogeneous documents, the TREC Web2014 (ClueWeb12) collection for heterogeneous web documents, and the LETOR3.0 collection for comparison with supervised feature-based models.

Leveraging Query Sensitivity for Practical Private Web Search

Several private Web search solutions have been proposed to preserve the user privacy while querying search engines. However, most of these solutions are costly in term of processing, network overhead and latency as they mostly rely on cryptographic techniques and/or the generation of fake requests. Furthermore, all these solutions protect all queries similarly, ignoring whether the original request contains sensitive content (e.g., religious, political or sexual orientation) or not. Based on an analysis of a real dataset of Web search requests, we show that queries related to sensitive matters are in practice a minority. As a consequence, protecting all queries similarly results in poor performance as a large number of queries get overprotected. In this paper, we propose a request sensitivity assessment module that we use for improving the practicability of existing private web search solutions. We assess the sensitivity of a request in two phases: a semantic sensitivity analysis (based on the topic of the query) and a request linkability analysis (based on the similarity between the current query and the query history of the requester). Finally, the sensitivity assessment is used to adapt the level of protection of a given query according to its identified degree of sensitivity: the more sensitive a query is, the more protected it will be. Experiments with a real dataset show that our approach can improve the performance of state-of-the-arts private Web search solutions by reducing the number of queries overprotected, while ensuring a similar level of privacy to the users, making them more likely to be used in practice.

Sélection de variables en apprentissage d'ordonnancement : évaluation des SVM pondérés

Selectionner les caracteristiques les plus utiles et les moins redondantes au sein des fonctions d’ordonnancement et reduire les temps d’execution sont des enjeux en apprentissage d’ordonnancement. Les algorithmes de selection de variables bases sur les SVM regularises sont des approches prometteuses dans ce cadre. Dans cet article, nous proposons de nouvelles methodes de selection de variables en apprentissage d’ordonnancement basees sur des approches de ponderation des SVM en norme l2. Nous proposons une adaptation d’une methode l2-AROM qui resout des SVM en norme l0 et un algorithme de ponderation de la norme l2 qui resout les problemes en norme l0 et l1. Nos evaluations sur des jeux de donnees industriels et de reference montrent que les methodes proposees sont jusqu’a 7 fois plus rapides et 10 fois plus parcimonieuses que l’etat de l’art, pour des qualites d’ordonnancement equivalentes.

De l'apprentissage d'ordonnancement à l'adaptation au contexte : Etat de l'art et propositions

Les moteurs de recherche georeferences utilisent des algorithmes d’ordonnancement complexes, prenant en compte le contexte d’utilisation, l’e-reputation et les reseaux sociaux, pour classer pertinemment les lieux vis-a-vis d’une requete. Or, comprendre les criteres de selection des utilisateurs et d’ordonnancement des moteurs est crucial pour les entreprises. Nous presentons le principe de l’optimisation de l’ordonnancement sur les moteurs de recherche et les approches et algorithmes existants. Nous montrons qu’ils sont limites et non adaptes au georeferencement. Nous proposons une amelioration de l’evaluation de la pertinence et une methodologie d’adaptation aux requetes utilisant la selection de variables embarquee.