Section: New Results

Scalable Data Analysis

StreamCloud

Participants : Vincenzo Gulisano, Patrick Valduriez.

Recent years have witnessed the growth of a new class of data-intensive applications that do not fit the DBMS query paradigm. Instead, the data arrive at high speeds taking the form of an unbounded sequence of values (data streams) and queries run continuously returning new results as new data arrive. Examples of data streams are sensor data (e.g. in environmental applications) or IP packets (e.g. in a network monitoring application). The unbounded nature of data streams makes it impossible to store the data entirely in bounded memory. Current research efforts have mainly focused on scaling in the number of queries and/or query operators having overlooked the scalability with respect to the stream volume.

Current Stream Processing Engines do not scale with the input load due to single-node bottlenecks. Additionally, they are based on static configurations that lead to either under or over-provisioning. In [21] , [22] , we present StreamCloud, a scalable and elastic stream processing engine for processing large data stream volumes. StreamCloud uses a novel parallelization technique that splits queries into subqueries that are allocated to independent sets of nodes in a way that minimizes the distribution overhead. Its elastic protocols exhibit low intrusiveness, enabling effective adjustment of resources to the incoming load. Elasticity is combined with dynamic load balancing to minimize the computational resources used. We present the system design, implementation and a thorough evaluation of the scalability and elasticity of the fully implemented system.

Mining Uncertain Data Streams

Participants : Reza Akbarinia, Florent Masseglia.

Dealing with uncertainty by using probabilistic approaches has gained increasing attention these past few years. One of the main requirements for uncertain data mining is the ability to discover Probabilistic Frequent Itemsets (PFI). However, PFI mining, particularly in uncertain data streams, is very challenging and needs the development of new techniques, since approaches designed for deterministic data are not applicable in this context. In [29] , we propose an efficient solution for exact PFI mining over data streams with sliding windows. Our proposal includes efficient solutions for updating frequentness probability of itemsets and thus fast extraction of PFI, whenever transactions are added or removed from the sliding window. To the best of our knowledge, this is the first efficient solution for data stream PFI mining. We have conducted an extensive experimental evaluation of our approach over synthetic and real-world data sets; the results illustrate its very good performance.

Detecting Rare Events in Massive Datasets

Participant : Florent Masseglia.

In this work, we consider that rare events are very small clusters typically representing less than 0.01% of the entire dataset. Finding these abnormal events allows to identify the emergence of pos- sible anomalies in their very early stages. Such a scenario is generally difficult to handle as it lies at the frontier between outlier detection and clustering and is characterized by a clear challenge to avoid false nega- tives. To address this challenge, we take a backward approach and pro- pose RARE, a framework that identifies and isolates the abnormal/rare regions. The dense regions are identified using a radius-limited density- driven variant of k-means and adjacent regions are merged to form new regions. These newly formed regions are gradually augmented as long as a density-driven condition is respected. When no more dense regions are observed, the remaining data is clustered and presented for further analysis to human experts. The framework is tested on a medical appli- cation and compared against human analysis. The experiments show that rare events that were missed during human analysis because of the multivariate character of the data can be discovered by our approach.

This work is funded by the labex NUMEV and a patent application involving Inria, CNRS, UM2 and INSERM has been filled.

Highly Informative Feature Set Mining

Participant : Florent Masseglia.

For many textual collections, the number of features is often overly large. As these features can be very redundant, it is desirable to have a small, succinct, yet highly informative collection of features that describes the key characteristics of a dataset. Information theory is one such tool for us to obtain this feature collection. In [48] , we mainly contribute to the improvement of efficiency for the process of selecting the most informative feature set over high-dimensional unlabeled data. We propose a heuristic theory for informative feature set selection from high dimensional data. Moreover, we design data structures that enable us to compute the entropies of the candidate feature sets efficiently. We also develop a simple pruning strategy that eliminates the hopeless candidates at each forward selection step. We test our method through experiments on real-world data sets, showing that our proposal is very efficient.

Clustering Users with Evolving Profiles in Usage Streams

Participant : Florent Masseglia.

Existing data stream models commonly assume that users' records or profiles in data streams will not be updated once they arrive. In many applications such as web usage, however, the users' records/profiles may evolve along time. This kind of streaming transactions are referred to as bi-streaming data (i.e. the data evolves temporally in two dimensions, the flowing of transactions as with the traditional data streams, and the evolving of users' profiles inside the streams, which makes bi-streaming data different from traditional data streams). The two-dimensional evolving of bi-streaming data brings difficulties on modeling and clustering for exploring the users' behaviors. In [49] , we propose three models to summarize bi-streaming data, which are the batch model, the Evolving Objects (EO) model and the Dynamic Data Stream (DDS) model. Through creating, updating and deleting user profiles, the models summarize the behaviors of each user as an object. Based on these models, clustering algorithms are employed to identify the user groups. The proposed models are tested on a real-world data set showing that the DDS model can summarize the bi-streaming data efficiently and effectively, providing better basis for clustering user profiles than the other two models.

Scalable Mining of Small Visual Objects

Participants : Pierre Letessier, Julien Champ, Alexis Joly.

Automatically linking multimedia documents that contain one or several instances of the same visual object has many applications including: salient events detection, relevant patterns discovery in scientific data or simply web browsing through hyper-visual links. Whereas efficient methods now exist for searching rigid objects in large collections, discovering them from scratch is still challenging in terms of scalability, particularly when the targeted objects are rather small. In this work [40] , we formally revisit the problem of mining or discovering such objects, and then generalized two kinds of existing methods for probing candidate object seeds: weighted adaptive sampling and hashing based methods. We then introduce a new hashing strategy, working first at the visual level, and then at the geometric level. Experiments conducted on millions of images show that our method outperforms state-of-the-art.

This method was integrated within a visual-based media event detection system in the scope of a French project called the transmedia observatory. It allows the automatic discovery of the most circulated images across the main news media (news websites, press agencies, TV news and newspapers). The main originality of the detection is to rely on the transmedia contextual information to denoise the raw visual detections and consequently focus on the most salient trans-media events. This work was presented at ACM Multimedia Grand Challenge 2012 [39] . The movie presented during this event is available at http://www.otmedia.fr/?p=217 .