Section: New Results

Machine learning for model acquisition

Participants : Marie-Odile Cordier, Thomas Guyet, Simon Malinowski, René Quiniou, Sid Ahmed Benabderrahmane.

Model acquisition is an important issue for model-based diagnosis, especially while modeling dynamic systems. We investigate machine learning methods for temporal data recorded by sensors or spatial data resulting from simulation processes. Our main objective is to extract knowledge, especially sequential and temporal patterns or prediction rules, from static or dynamic data (data streams). We are particularly interested in mining temporal patterns with numerical information and in incremental mining from sequences recorded by sensors.

Mining temporal patterns with numerical information

We are interested in mining interval-based temporal patterns from event sequences where each event is associated with a type and time interval. Temporal patterns are sets of constrained interval-based events. This year we have begun to work on multiscale temporal abstraction to represent time series by codewords at different temporal and amplitude scales. We have improved the method of Wang et al. [70] by introducing Dynamic Time Warping to compute better codewords for time series abstraction. The codeword-based time series representation is then used by QTIPrefixSpan [3] to extract temporal patterns. A paper is in preparation. We are also working on a multivariate version of the method for mining multivariate temporal patterns at different resolution levels.

Incremental sequential mining

Sequential pattern mining algorithms operating on data streams generally compile a summary of the data seen so far from which they compute the set of actual sequential patterns. We propose another solution where the set of actual sequential patterns are incrementally updated as soon as new data arrive on the input stream. Our work stands in the framework of mining an infinite unique sequence. Though being of great importance, this problem has not received a similar attention as mining from a transaction database. Our method [13] provides an algorithm that maintains a tree representation (inspired by the PSP algorithm [56] ) of frequent sequential patterns and their minimal occurrences [54] in a window that slides along the input data stream. It makes use of two operations: deletion of the itemset at the beginning of the window (obsolete data) and addition of an itemset at the end of the window (new data). The experiments were conducted on simulated data and on real data of instantaneous power consumption. The results show that our incremental algorithm significantly improves the computation time compared to a non-incremental approach [14] .

Incremental learning of preventive rules

The problem is to learn preventive rules in order to avoid malfunctioning on smartphones. A monitoring module is embedded on the phones and sends reports to a server. Reports are labeled with a normal or abnormal label. From this set of reports new rules are learned. As a lot of smartphones are supervised, it is impossible to store all the reports. Therefore incremental learning has to be used.

Last year, we achieved two main tasks: a report database has been built in order to test the future algorithms, and a new algorithm [20] has been developed for implementing an incremental version of the algorithm AQ21 [72] .

Multiscale segmentation of satellite image time series

Satellite images allow the acquisition of large-scale ground vegetation. Images are available along several years with a high acquisition frequency (1 image every two weeks). Such data are called satellite image time series (SITS). In [12] , we present a method to segment an image through the characterization of the evolution of a vegetation index (NDVI) on two scales: annual and multi-year. We test this method to segment Senegal SITS and compare our method to a direct classification of time series. The results show that our method using two time scales better differentiates regions in the median zone of Senegal and locates fine interesting areas (cities, forests, agricultural areas).

Mining a big unique graph for spatial pattern extraction

Researchers in agro-environment needs a great variety of landscapes to test the agro-ecological models of their scientific hypotheses. As the representation of real landcapes necessitates lots of on-land measures, good big representations are difficult to acquire. Working with landscape simulations is then an alternative to get a sufficient variety of experimental landscapes. We propose to extract spatial patterns from a well described geographic area and to use these patterns to generate realistic landscapes. We have begun the exploration of graph mining techniques to discover the relevant spatial patterns present in a graph expressing the spatial relationships between the agricultural plots as well as the roads, the rivers, the buildings, etc., of a specific geographic area.

This year, we have been working on extending algorithm gSPAN [73] with an adaptive support threshold and with a taxonomy to be able to extract interesting patterns involving agricultural plots with rare features. We plan to submit a paper in 2013.