Section: New Results

Machine learning for model acquisition

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

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.

Representing and mining time series

Time series are sequences of numerical values, e.g. recorded by sensors. Since these series can be huge and subject to noise, they are often transformed into sequences of symbols. The best known symbolic transformation method is SAX (Symbolic Aggregate approXimation) [68] . SAX is based on a piecewise constant approximation method that does not take into account the slope of the time series values in successive windows. We have extended the SAX method by adding a symbolic slope information to the SAX symbols. We have experimented our new representation, 1d-SAX, on three mining tasks. In most of these experiments 1d-SAX leads to a better accuracy than SAX [19] .

We have also investigated a probabilistic representation of temporal patterns based on the latent Dirichlet allocation model (LDA). Such patterns can approximate the dynamics of a set of similar multivariate time series. We have experimented the method on hydrological flood time series to extract temporal patterns [7] . The extracted patterns were considered relevant and easy to understand by experts of the domain.

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. Our method [60] provides an algorithm that maintains a tree representation (inspired by the PSP algorithm [71] ) of frequent sequential patterns and their minimal occurrences [69] 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 [61] .

Recently, we have worked on the adaptation of our algorithm to closed sequential patterns. A closed pattern is a local maximal pattern such there exists no extension of this pattern having the same support. Closed patterns are known to provide a condensed represention of the solution patterns and lead to more efficient algorithms without losing information or completeness on extracted patterns. The tree of closed-patterns is less deep than the pattern-tree but the transformations of the tree by addition or deletion of items are more complex. The algorithm is under evaluation. We plan to submit a paper in 2014.

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 [58] , 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 first propose to adapt image segmentation algorithm to SITS. Segmented images reduces the number of time series to analyze and the computation time. We secondly applied 1D-SAX to reduce data dimensionality [20] . We evaluated this approach on the supervised classification of large SITS of Senegal and we showed that 1D-SAX approaches the classification results of time series while significantly reducing the required memory storage of the images.

Analysis of landscape based on spatial patterns

Researchers in agro-environment need a great variety of landscapes to test the agro-ecological models of their scientific hypotheses. Real landscapes are difficult to acquire and do not enable the agronomist to test all their hypothesis. 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 realistic landscapes. This approach 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.), of a specific geographic area. We extract spatial patterns from a real geographic area and we use these patterns to generate new realistic landscapes. Using patterns preserves the interface properties between parcels.

We have begun the exploration of graph mining techniques, such as gSPAN [87] , to discover the relevant spatial patterns present in a spatial-graph. But the graph-mining techniques are very time-consuming in comparison to sequence mining.

This year, we would like to test if using a path instead of a graph would be a faithful representation of the spatial organization of the landscape. In [17] , we compare the potential expressivity of graphs and Hilbert-Peano curves [66] to characterize an agricultural landscape. The results show that mining frequent patterns in Hilbert-Peano curves would be as discriminant as mining frequent patterns in graphs.

The perception of the environment is an important dimension of the landscape we live in. One of our objectives is to study the relationships between the landscape patterns and their perception. We cope with this dimension by analysing the textual content of `atlas du paysage" (landscape atlas), that are produce by each french administrative regions. This year we worked on the construction of an ontology of landscape perception [21] .

Subdimensional clustering for fast similarity search over time series data. Application to Information retrieval tasks

Information retrieval and similarity search tasks in time series databases remains a challenge that require to discover relevant pattern-sequences that are recurrent over the overall time series sequences, and to find temporal associations among these frequently occurring patterns. Previous work on information retrieval and similarity search in time series has been performed in different contexts such as diagnosis or failure detection of industrial materials. In whole query matching, a time series given as query is entirely compared to every time series of a database. The series should have same length, and a similarity measure is used to retrieve either a most similar time series or the top-k ranked time series. However, theses methods suffer from a lack of flexibility of the used similarity measures, a lack of scalability of the representation model, and a penalizing runtime to retrieve the information. Moreover, in some real world applications, one can be interested in retrieving specific interesting subsequences that are frequently present at different instants.

Motivated by these observations, we have designed a framework tackling the query by content problem on time series data, ensuring (i) fast response time, (ii) multi-level information representation, and (iii) representing temporal associations between extracted patterns. During the preparation step, all the multi-valued time series present in the database are transformed into a multi-resolution symbolic representation thus ensuring a lower dimensionality. Then, to accelerate and enhance the similarity search and the retrieval over the database, our model creates an index over recurrent patterns in the time series collection. These patterns can be generated by different techniques. Finally, the extracted patterns are grouped by clustering and the resulting clusters are indexed in a table within their centroids. A paper presenting the preliminary results is under submission to an international journal.

Knowledge Extraction from Heterogeneous Data

Recently, mining microarrays data has became a big challenge due to the growing sources of available data. We are using machine learning methods such as clustering, dimensionality reduction, association rules discovery on transcriptomic data, by combining a domain ontology as source of knowledge, in order to supervise the KDD process. Our objectives concern the identification of genes that could participate in the development of tumors. A two-way classification method was proposed, combining genes expression levels, represented as numerical data, and Gene Ontology (GO) annotations as symbolic data. The hopeful results obtained with genes clustering, through GO annotations, are an encouraging track to predict transcriptional regulatory networks, and for refining the existing sets of genes [11] , [12] .

We also introduced a new method for extracting enriched biological functions from transcriptomic databases using an integrative bi-classication approach. The initial gene datasets are firstly represented as a formal context (objects attributes), where objects are genes, and attributes are their expression profiles and complementary information of different knowledge bases. Formal Concept Analysis (FCA) is applied for extracting formal concepts regrouping genes having similar transcriptomic profiles and functional behaviors. An enrichment analysis is then performed in order to identify the relevant formal concepts from the generated Galois lattice, and to extract biological functions that could participate in the proliferation of cancers. Preliminary results seem very promising, and could help experts during the identification of degenerated biological functions [13] .