Nicolas Labroche

Fast Unsupervised Clustering with Artificial Ants

AntClust is a clustering algorithm that is inspired by the chemical recognition system of real ants. It associates the genome of each artificial ant to an object of the initial data set and simulates meetings between ants to create nests of individuals that share a similar genome. Thus, the nests realize a partition of the original data set with no hypothesis concerning the output clusters (number, shape, size ...) and with minimum input parameters. Due to an internal mechanism of nest selection and finalization, AntClust runs in the worst case in quadratic time complexity with the number of ants. In this paper, we evaluate new heuristics for nest selection and finalization that allows AntClust to run on linear time complexity with the number of ants.

Visual Clustering with Artificial Ants Colonies

In this paper, we propose a new model of the chemical recognition system of ants to solve the unsupervised clustering problem. The colonial closure mechanism allows ants to discriminate between nestmates and intruders by the mean of a colonial odour that is shared by every nestmate. In our model we associate each object of the data set to the odour of an artificial ant. Each odour is defined as a real vector with two components, that can be represented in a 2D-space of odours. Our method simulates meetings between ants according to pre-established behavioural rules, to ensure the convergence of similar odours (i.e. similar objects) in the same portion of the 2D-space. This provides the expected partition of the objects. We test our method against other well-known clustering method and show that it can perform well. Furthermore, our approach can handle every type of data (from numerical to symbolic attibutes, since there exists an adapted similarity measure) and allows one to visualize the dynamic creation of the nests. We plan to use this algorithm as a basis for a more sophisticated interactive clustering tool.

Materializing Baseline Views for Deviation Detection Exploratory OLAP

Alert-raising and deviation detection in OLAP and explora-tory search concerns calling the user’s attention to variations and non-uniform data distributions, or directing the user to the most interesting exploration of the data. In this paper, we are interested in the ability of a data warehouse to monitor continuously new data, and to update accordingly a particular type of materialized views recording statistics, called baselines. It should be possible to detect deviations at various levels of aggregation, and baselines should be fully integrated into the database. We propose Multi-level Baseline Materialized Views (BMV), including the mechanisms to build, refresh and detect deviations. We also propose an incremental approach and formula for refreshing baselines efficiently. An experimental setup proves the concept and shows its efficiency.

User Interests Clustering in Business Intelligence Interactions

It is quite common these days for experts, casual analysts, executives or data enthusiasts, to analyze large datasets using user-friendly interfaces on top of Business Intelligence (BI) systems. However, current BI systems do not adequately detect and characterize user interests, which may lead to tedious and unproductive interactions. In this paper, we propose to identify such user interests by characterizing the intent of the interaction with the BI system. With an eye on user modeling for proactive search systems, we identify a set of features for an adequate description of intents, and a similarity measure for grouping intents into coherent interests. We validate experimentally our approach with a user study, where we analyze traces of BI navigation. We show that our similarity measure outperforms a state-of-the-art query similarity measure and yields a very good precision with respect to expressed user interests.

Detecting User Focus in OLAP Analyses

In this paper, we propose an approach to automatically detect focused portions of data cube explorations by using different features of OLAP queries. While such a concept of focused interaction is relevant to many domains besides OLAP explorations, like web search or interactive database exploration, there is currently no precise formal, commonly agreed definition. This concept of focus phase is drawn from Exploratory Search, which is a paradigm that theorized search as a complex interaction between a user and a system. The interaction consists of two different phases: an exploratory phase where the user is progressively defining her information need, and a focused phase where she investigates in details precise facts, and learn from this investigation. Following this model, our work is, to the best of our knowledge, the first to propose a formal feature-based description of a focused query in the context of interactive data exploration. Our experiments show that we manage to identify focused queries in real navigations, and that our model is sufficiently robust and general to be applied to different OLAP navigations datasets.

Benchmarking Exploratory OLAP

Supporting interactive database exploration (IDE) is a problem that attracts lots of attention these days. Exploratory OLAP (On-Line Analytical Processing) is an important use case where tools support navigation and analysis of the most interesting data, using the best possible perspectives. While many approaches were proposed (like query recommendation, reuse, steering, personalization or unexpected data recommendation), a recurrent problem is how to assess the effectiveness of an exploratory OLAP approach. In this paper we propose a benchmark framework to do so, that relies on an extensible set of user-centric metrics that relate to the main dimensions of exploratory analysis. Namely, we describe how to model and simulate user activity, how to formalize our metrics and how to build exploratory tasks to properly evaluate an IDE system under test (SUT). To the best of our knowledge, this is the first proposal of such a benchmark. Experiments are two-fold: first we evaluate the benchmark protocol and metrics based on synthetic SUTs whose behavior is well known. Second, we concentrate on two different recent SUTs from IDE literature that are evaluated and compared with our benchmark. Finally, potential extensions to produce an industry-strength benchmark are listed in the conclusion.

Semi-supervised Fuzzy c-Means Variants: A Study on Noisy Label Supervision.

Semi-supervised clustering algorithms aim at discovering the hidden structure of data sets with the help of expert knowledge, generally expressed as constraints on the data such as class labels or pairwise relations. Most of the time, the expert is considered as an oracle that only provides correct constraints. This paper focuses on the case where some label constraints are erroneous and proposes to investigate into more detail three semi-supervised fuzzy c-means clustering approaches as they have been tailored to naturally handle uncertainty in the expert labeling. In order to run a fair comparison between existing algorithms, formal improvements have been proposed to guarantee and fasten their convergence. Experiments conducted on real and synthetical datasets under uncertain labels and noise in the constraints show the effectiveness of using fuzzy clustering algorithm for noisy semi-supervised clustering.

On Evidential Clustering with Partial Supervision

This paper introduces a new semi-supervised evidential clustering algorithm. It considers label constraints and exploits the evidence theory to create a credal partition coherent with the background knowledge. The main characteristics of the new method is its ability to express the uncertainties of partial prior information by assigning each constrained object to a set of labels. It enriches previous existing algorithm that allows the preservation of the uncertainty in the constraint by adding the possibility to favor crisp decision following the inherent structure of the dataset. The advantages of the proposed approach are illustrated using both a synthetic dataset and a real genomics dataset.