Section: New Results

Data Mining

Participants : Marie-Odile Cordier, Yann Dauxais, Serge Vladimir Emteu Tchagou, Clément Gautrais, Thomas Guyet, Yves Moinard, Benjamin Negrevergne, René Quiniou, Laurence Rozé, Alexandre Termier.

Sequential pattern mining with intervals

In previous work, we developed a framework for sequential pattern mining with intervals [3] . It has been applied in various application (care-pathways, customer relationship management databases [35] , etc.).

This year we explored chronicle mining algorithms for mining care-pathways (see section 9.1.1 , for an applicative context). Chronicles are alternative patterns for representing temporal behaviors [58] . A chronicle can be briefly defined as a set of events linked by constraints indicating the minimum and maximum time elapsed between two events. A care-pathway contains point-based events (e.g. surgery) and interval-based events (e.g. drug exposures). A chronicle can express such a complex temporal behaviour, for instance: The patient was exposed to a drug X between 1 and 2 years, he met his doctor between 400 to 600 days after the beginning of the exposure and, finally, he was hospitalized.

The first algorithm we worked on [23] is an adaptation of existing chronicle mining algorithms [55] , [63] to mine the complete set of frequent chronicles from a collection of care-pathways. This algorithm uses the search-space browsing strategy of HDCA [55] and the support evaluation of CCP-Miner [63] . As the complete set of chronicle is huge, we also proposed an incomplete algorithms based on the original simplifications of [58] . These algorithms were implemented and evaluated on real and simulated datasets.

We also investigated discriminant chronicles mining which consists in extracting the chronicles that are $\alpha$ times more frequent in a database ${𝒟}_{+}$ than in a database ${𝒟}_{-}$. Mining discriminant chronicles is very useful to discover the features of care-pathways that are related, for instance, to a specific disease. Our approch has been implemented and is under evaluation.

Multiscale segmentation of satellite image time series

Satellite images enable the acquisition of large-scale ground vegetation information. Images have been recorded for several years with a high acquisition frequency (one image every two weeks). Such data are called satellite image time series (SITS). Several articles were published this year and they correspond to past work on algorithms and method to analyse SITS.

In [11] , we presented a method to segment an image through the characterization of the evolution of a vegetation index (NDVI) on two scales: annual and multi-year. The main issue of this approach was the required computation resources (time and memory).

We also explored the supervised classification of SITS using classification trees for time-series [27] by implementing a parallelized version of this algorithm. Next, we explored the adaptation of the object-oriented segmentation to SITS. The object-oriented segmentation is able to segment images based on segment uniformity. We proposed a measure for time-series uniformity and applied the adapted algorithm on large multivariate SITS of Senegal [10] .

Third, we presented an supervised approach to extract features from classified satellite images to analyse urban sprawl [28] . In this work, we have satellite images at only two dates, and the objective is to identify characteristics that can foster or prevent changes.

Our satellite images analysis approaches are used in two applicative contexts: understanding urban sprawl and analyzing drought in Senegal. Analysis of urban sprawl was a collaborative work with collegues in remote sensing, in landscapes analysis and in economical modelling. Our collective contribution was published in a book of the PDD2 (PDD2: Paysage Developpement Durable/Landscape Sustainable Develompent) program [38] . Analysis of drought in Senegal is a long term collaboration with H. Nicolas (INRA/SAS) that we would like to continue in a collaboration with A. Fall (Université of Dakar) to confront our results with ground observations.

Analysis and simulation of landscape based on spatial patterns

Researchers in agro-environment need a great variety of landscapes to test their scientific hypotheses using agro-ecological models. Real landscapes are difficult to acquire and do not enable the agronomists to test all their hypotheses. Working with simulated landscapes is then an alternative to get a sufficient variety of experimental data. Our objective is to develop an original scheme to generate landscapes that reproduce realistic interface properties between parcels. This approach consists of the extraction of spatial patterns from a real geographic area and the use of these patterns to generate new "realistic" landscapes. It is based on a spatial representation of landscapes by a graph expressing the spatial relationships between the agricultural parcels (as well as the roads, the rivers, the buildings, etc.), in a specific geographic area.

In past years, we worked on the exploration of graph mining techniques, such as gSPAN [85] , to discover the relevant spatial patterns present in a spatial-graph. We assume that the set of the frequent graph patterns are the characterisation of the landscape. Our remaining challenge was to simulate new realistic landscapes that reproduce the same patterns.

We have formalized the simulation process as a graph packing problem [66] . The process is illustrated by Figure 1 . Solving instances of the general graph packing problem has a high combinatorics and no efficient algorithm can solve it. We proposed an ASP program to tackle the combinatorics of the graph packing and to assign the land use considering some expert knowledge. Our approach combines the efficiency of ASP to solve the packing issue and the simplicity of the declarative programming to take into account expert contraints on the land use. Contraints about the minimum surface of crops or about the impossibility of some crops colocation can be easily defined. This work have been presented at the conference RFIA and an extended version has been published in the Revue d'Intelligence Artificielle (RIA) [13] .

In addition, we are collaborating with J. Nicolas (EPI Dyliss) to improve the efficiency of our first programs. The improvements are based on symmetry breaking of ASP programs. To this end, we proposed a simplified encoding of the graph patterns using spanning trees and used automorphism detection in graph patterns to automatically encodes symmetry breakings. Intensive evaluation of our encoding shown that this improvement enable to tackle significantly larger graphs than early programs did. This work will be soon submitted to a high ranking conference.

Mining with ASP

In pattern mining, a pattern is considered interesting if it occurs frequently in the data, i.e. the number of its occurrences is greater than a fixed given threshold. As non informed mining methods tend to generate massive results, there is more and more interest in pattern mining algorithms able to mine data considering some expert knowledge. Though a generic pattern mining tool that could be tailored to the specific task of a data-scientist is still a holy grail for pattern mining software designers, some recent attempts have proposed generic pattern mining tools [61] for itemset mining tasks. In collaboration with Torsten Schaub, we explore the ability of a declarative language, such as Answer Set Programming (ASP), to solve pattern mining tasks efficiently. In 2011, Jarvisälo proposed a first attempt devoted to itemset mining [64] . In Dream, we are working on sequential pattern mining, which is known to be more challenging than itemset mining and which has been also recently considered by constraint programming approaches [76] .

We have worked on encoding in ASP most of sequential pattern mining tasks: sequences with constraints (gaps, maximum length, etc.), closed/maximal patterns, emergent sequences. Our first result is to show that ASP is suitable for encoding such complex pattern mining tasks. The experimental results show that our purely declarative approach is less efficient than constraint programming approaches [36] . Nonetheless, it is suitable to be blend with intensive knowledge. The challenge is now to show that our ASP framework can extract the meaningful patterns that other approaches loose in the overwhelming amount of sequential patterns.

A first attempt has been done in this direction in collaboration with J. Romero from the University of Potsdam. We used the system ASPRIN to define preferences on patterns. Defining preferences on patterns is also a classical approach to select the most interesting patterns. Some classical preferences on sequential patterns have been defined and the ASPRIN system is used to extract the preferred patterns according to one preference or a combination of preferences (skypatterns [81] )

This work will be soon submitted to a high ranking international conference.

Mining time series

Monitoring cattle. Following the lines of a previous work [79] , we are working on a method for detecting Bovine Respiratory Diseases (BRD) from behavioral (walking, lying, feeding and drinking activity) and physiological (rumen temperature) data recorded on feedlot cattle being fattened up in big farms in Alberta (Canada). This year, we have especially worked on multivariate sensor data analysis, especially on the evaluation of different combinations of sensors for determining the best configuration and parameter setting. This work was part of Afra Verena Mang's master thesis defended in september 2015 [73] . Two papers are in preparation.

SIFT-based time-series symbolisation Time series classification is an application of particular interest with the increase of such data. Computing the distance between time-series is time consuming. An abstract representation of time-series that accurately approximates distances between time-series and makes easier their comparison is highly expected. In [17] , we proposed a time series classification scheme grounded on the SIFT framework [70] adapted to time series. The SIFTs feed a Bag-of-Words representation of time-series. We have shown that this framework efficiently and accurately classifies time series, despite the fact that BoW representation ignores temporal order.

Mining sequential patterns from multimedia data Analyzing multimedia data to extract knowledge is a challenging problem due to the quantity and complexity of such data. Finding recurrent patterns is one method to structure and segment the data. In a collaboration with the EPI LinkMedia, we have proposed audio data symbolization and sequential pattern mining methods to extract patterns from audio streams. Experiments show this the task is hard and that the symbolization is a critical step for extracting relevant audio patterns [29] .

Mining customer data for predicting and explaining attrition

Predicting customer defection in a retail context is difficult because, in most situations, the customer does not leave the store totally (there is no contract break as with banks or phone operators). We have proposed a new pattern model for representing the evolution of an individual customer purchase behavior that enables to early detect and to explain customer attrition. In particular, this model enables the analyst to determine which important kinds of product receives less and less attention from the customer. Thus, this model provides actionable knowledge at an individual scale that lets the retailer trigger targeted marketing actions to counter attrition. A poster has been submitted to the EBDT conference. This work has been performed during Clément Gautrais's master [59] and will be further investigated and extended during his PhD.

Mining energy consumption data

Machine tools in companies consume a lot of energy (before, during and after producing worked pieces). This year, we are beginning to work, with the start-up Energiency, on mining machine tool energy consumption data in order to propose energy savings to the companies. Firstly, we try to determine, according to the analyzed company, which data-mining algorithm should be used and which is the best configuration and parameter setting. Then, we aim to extract actions rules from patterns to help companies to consume less energy.

Trace reduction

One problem of execution trace of applications on embedded systems is that they can grow very large, typically several Gigabytes for 5 minutes of audio/video playback. Some endurance tests require continuous playback for 96 hours, which would lead to hundreds of Gigabytes of traces, that current techniques cannot analyze. We have proposed TraceSquiz, an online approach to monitor the trace output during endurance test, in order to record only suspicious portions of the trace and discard regular ones. This approach is based on anomaly detection techniques. Our detailed experiments have shown that our approach has a good anomaly detection performance, and can reduce the size of an output trace by an order of magnitude [24] . Serge Emteu successfully defended his PhD about this work on the 15/12/2015 [5] .