Section: New Results

Scalable Data Analysis

Parallel Mining of Maximally Informative k-Itemsets in Big Data

Participants : Saber Salah, Reza Akbarinia, Florent Masseglia.

The discovery of informative itemsets is a fundamental building block in data analytics and information retrieval. While the problem has been widely studied, only few solutions scale. This is particularly the case when i) the data set is massive, and/or ii) the length K of the informative itemset to be discovered is high. In [45] , we address the problem of parallel mining of maximally informative k-itemsets (miki) based on joint entropy. We propose PHIKS (Parallel Highly Informative K-itemSets) a highly scalable, parallel mining algorithm. PHIKS renders the mining process of large scale databases (up to terabytes of data) succinct and effective. Its mining process is made up of only two compact, yet efficient parallel jobs. PHIKS uses a clever heuristic approach to efficiently estimates the joint entropies of miki having different sizes with very low upper bound error rate, which dramatically reduces the runtime process. PHIKS has been extensively evaluated using massive, real-world data sets. Our experimental results confirm the effectiveness of our approach by the significant scale-up obtained with high featuresets length and hundreds of millions of objects.

Frequent Itemset Mining in Massively Distributed Environments

Participants : Saber Salah, Reza Akbarinia, Florent Masseglia.

While the problem of Frequent itemset mining (FIM) has been thoroughly studied, few solutions scale. This is mainly the case when i) the amount of data tends to be very large and/or ii) the minimum support (MinSup) threshold is very low. In [46] , we study the effectiveness and leverage specific data placement strategies for improving parallel FIM (PFIM) performance in MapReduce, a highly distributed computation framework. By offering a clever data placement and an optimal organization of the extraction algorithms, we show that the itemset discovery effectiveness does not only depend on the deployed algorithms. We propose ODPR (Optimal Data-Process Relationship), a solution for fast mining of frequent itemsets in MapReduce. Our method allows discovering itemsets from massive datasets, where standard solutions do not scale.

In [44] , we propose a highly scalable PFIM algorithm, namely Parallel Absolute Top Down (PATD). PATD renders the mining process of very large databases (up to Terabytes) simple and compact. Its mining process is made up of only one parallel job, which dramatically reduces the mining runtime, communication cost and energy power consumption overhead, in a distributed computational platform. Based on a clever and efficient data partitioning strategy, namely Item Based Data Partitioning (IBDP), PATD mines each data partition independently, relying on an absolute minimum support (AM inSup) instead of a relative one. Through an extensive experimental evaluation using real-world data sets, we show that PATD is significantly more efficient and scalable than alternative approaches.

Scalable Mining of Closed Frequent Itemsets

Participants : Mehdi Zitouni, Reza Akbarinia, Florent Masseglia.

Mining big datasets poses a number of challenges which are not easily addressed by traditional mining methods, since both memory and computational requirements are hard to satisfy. One solution is to take advantage of parallel frameworks, such as MapReduce, using ordinary machines. In [48] , we address the issue of mining closed frequent itemsets (CFI) from big datasets in such environments. We introduce a new parallel algorithm, called CloPN , for CFI mining. One important feature of CloPN is to use a prime number based approach to transform the data into numerical form, and then to mine closed frequent itemsets by using only multiplication and division operations. We carried out exhaustive experiments over big real world datasets to assess the performance of CloPN. The results show that our algorithm is very efficient in CFI mining from large real world datasets with up to 53 million articles.

Chiaroscuro

Participants : Tristan Allard, Florent Masseglia, Esther Pacitti.

The advent of on-body/at-home sensors connected to personal devices leads to the generation of fine grain highly sensitive personal data at an unprecendent rate. However, despite the promises of large scale analytics there are obvious privacy concerns that prevent individuals to share their personnal data. In [30] , we propose Chiaroscuro, a complete solution for clustering personal data with strong privacy guarantees. The execution sequence produced by Chiaroscuro is massively distributed on personal devices, coping with arbitrary connections and disconnections. Chiaroscuro builds on our novel data structure, called Diptych, which allows the participating devices to collaborate privately by combining encryption with differential privacy. Our solution yields a high clustering quality while minimizing the impact of the differentially private perturbation. Our study show that Chiaroscuro is both correct and secure.

Large-scale Recognition of Visual and Audio Entities

Participants : Valentin Leveau, Alexis Joly, Patrick Valduriez.

We improved our work on the retrieval of visual identities by introducing a supervised classification layer on top of the large-scale instance-based matching layer. We introduce a new match kernel based on the inverse rank of the Shared Nearest Neighbors (SNN) combined with local geometric constraints [40] . To avoid overfiting and reduce processing costs, the dimensionality of the resulting over-complete representation is further reduced by hierarchically pooling the raw consistent matches according to their spatial position in the training images. The final image representation is obtained by concatenating the resulting feature vectors at several resolutions. Learning from these representations using a logistic regression classifier is shown to provide excellent fine-grained classification performance. In [38] , we transpose our new SNN match kernel to the case of audio contents (applied to bird sounds recognition). Thus, the spatial pooling of geometrically consistent visual matches is replaced by a temporal pooling of temporally consistent audio matches. The resulting classification system obtained the second best results at the LifeCLEF bird identification challenge 2015 [36] , the largest challenge of this kind ever organized (1000 bird species, 33K audio recordings).

Crowd-sourced Biodiversity Data Production through Pl@ntNet

Participants : Alexis Joly, Julien Champ, Jean-Christophe Lombardo, Antoine Affouard.

Initiated in the context of a citizen sciences project with botanists of the AMAP laboratory and the Tela Botanica social network, Pl@ntNet [18] is an innovative collaborative plateform focused on image-based plant identification as a mean to enlist new contributors and boost the production of biodiversity data and knowledge. Since 2010, several hundreds of thousands of geo-tagged and dated plant photographs were collected and revised by tens of thousands of novice, amateur and expert botanists. A content-based identification tool, available as both web and mobile applications, is synchronized with the growing data and allows any user to query or enrich the system with new observations. As a concrete new result, the cumulative number of downloads of the iPhone or Android app did reach 1M in October 2015. One of the main novelty in 2015 was the introduction of deep learning technologies in order to improve classification performance as well as the quality and speed of the content-based image retrieval.

A comparative study that we conducted in the context of the LifeCLEF (www.lifeclef.org) plant identification challenge did actually confirm that deep convolutional neural networks definitely outperforms the best fine-grained classification models on the aggregation of hand-crafted visual features [33] . Thus, we integrated this technology in the Pl@ntNet platform and exploited it in two ways: (i) for extracting more relevant (local and global) visual features to be indexed and searched within our efficient content-based indexing and retrieval framework (SnoopIm software) (ii) for reranking the species returned by the content-based search engine so as to increase the average reciprocal rank of the correct species while keeping a good level of interpretability of the returned results.

Crowd-sourced Biodiversity Data Production through LifeCLE

Participants : Alexis Joly, Julien Champ, Jean-Christophe Lombardo, Antoine Affouard.

We continued sharing the data produced by the Pl@ntNet platform with the international research community through the animation of the LifeCLEF research platform and the set-up of three new challenges, one related to plant images, one to bird sounds and one to fish videos. More than 200 research groups registered to at least one of the challenges and about 20 of them crossed the finish lines by running their system on the final test data. A synthesis of the results is published in the LifeCLEF 2015 overview paper [37] and more detailed analyses are provided in technical reports for the plant task [35] and the bird task [36] . We also report on an experimental study aimed at evaluating how state-of-art computer vision systems perform in identifying plants compared to human expertise [15] ,. A subset of the evaluation dataset used within LifeCLEF 2014 plant identification challenge was shared with volunteers of diverse expertise, ranging from leading experts of the targeted flora to inexperienced test subjects. In total, 16 human runs were collected and evaluated comparatively to the 27 machine-based runs of LifeCLEF challenge. The main outcome of the experiment was that machines are still far from outperforming the best expert botanists but they are clearly competing with some experienced botanists specialists of other floras.