Search

An increasing number of applications require to recognize the class of an incoming time series as quickly as possible without unduly compromising the accuracy of the prediction. In this paper, we put forward a new optimization criterion which takes into account both the cost of misclassification and the cost of delaying the decision. Based on this optimization criterion, we derived a family of non-myopic algorithms which try to anticipate the expected future gain in information in balance with the cost of waiting. In one class of algorithms, unsupervised-based, the expectations use the clustering of time series, while in a second class, supervised-based, time series are grouped according to the confidence level of the classifier used to label them. Extensive experiments carried out on real data sets using a large range of delay cost functions show that the presented algorithms are able to satisfactorily solving the earliness vs. accuracy trade-off, with the supervised-based approaches faring better than the unsupervised-based ones. In addition, all these methods perform better in a wide variety of conditions than a state of the art method based on a myopic strategy which is recognized as very competitive., Accepted for publication in Machine learning journal (MACH)

Early classification of time series assigns each time series to one of a set of pre-defined classes using as few measurements as possible while preserving a high accuracy. This implies solving online the trade-off between the earliness and the prediction accuracy. This has been formalized in previous work where a cost-based framework taking into account both the cost of misclassification and the cost of delaying the decision has been proposed. The best resulting method, called Economy- $\gamma$ , is unfortunately so far limited to binary classification problems. This paper presents a set of six new methods that extend the Economy- $\gamma$ method in order to solve multiclass classification problems. Extensive experiments on 33 datasets allowed us to compare the performance of the six proposed approaches to the state-of-the-art one. The results show that: (i) all proposed methods perform significantly better than the state of the art one; (ii) the best way to extend Economy- $\gamma$ to multiclass problems is to use a confidence score, either the Gini index or the maximum probability.

The field of Weakly Supervised Learning (WSL) has recently seen a surge of popularity, with numerous papers addressing different types of "supervision deficiencies", namely: poor quality, non adaptability, and insufficient quantity of labels. Regarding quality, label noise can be of different types, including completely-at-random, at-random or even not-at-random. All these kinds of label noise are addressed separately in the literature, leading to highly specialized approaches. This paper proposes an original, encompassing, view of Weakly Supervised Learning, which results in the design of generic approaches capable of dealing with any kind of label noise. For this purpose, an alternative setting called "Biquality data" is used. It assumes that a small trusted dataset of correctly labeled examples is available, in addition to an untrusted dataset of noisy examples. In this paper, we propose a new reweigthing scheme capable of identifying noncorrupted examples in the untrusted dataset. This allows one to learn classifiers using both datasets. Extensive experiments that simulate several types of label noise and that vary the quality and quantity of untrusted examples, demonstrate that the proposed approach outperforms baselines and state-of-the-art approaches., 8 pages, 7 figures

We address the problem of event classification for proactive fiber break detection in high-speed optical communication systems. The proposed approach is based on monitoring the State of Polarization (SOP) via digital signal processing in a coherent receiver. We describe in details the design of a classifier providing interpretable decision rules and enabling low-complexity real-time detection embedded in network elements. The proposed method operates on SOP time series, which define trajectories on the 3D sphere; SOP time series are low-pass filtered (to reduce measurement noise), pre-rotated (to provide invariance to the starting point of trajectories) and converted to quaternion domain. Then quaternion sequences are recoded to relational data for automatic variable construction and selection. We show that a naive Bayes classifier using a limited subset of variables can achieve an event classification accuracy of more than 99% for the tested conditions.

International audience; Seasonal behaviours are widely encountered in various applications. For instance, requests on web servers are highly influenced by our daily activities. Seasonal forecasting consists in forecasting the whole next season for a given seasonal time series. It may help a service provider to provision correctly the potentially required resources, avoiding critical situations of over-or under provision. In this article, we propose a generic framework to make seasonal time series forecasting. The framework combines machine learning techniques 1) to identify the typical seasons and 2) to forecast the likelihood of having a season type in one season ahead. We study this framework by comparing the mean squared errors of forecasts for various settings and various datasets. The best setting is then compared to state-of-the-art time series forecasting methods. We show that it is competitive with then.

The incoming smart grid represents a significant break for the European utilities in terms of data volume to be processed. In France, one year of individual consumptions represents more than 600 billion data points. Since real data is not yet available, our objective consists in simulating realistic individual consumptions. A new generative model of time series is proposed which combines the MODL coclustering approach with Markov chains. This approach is evaluated on a real dataset provided by the Irish CER. Our experiments demonstrate the ability of the generative model to efficiently reproduce the dynamic of the original time series.

Classification of time series as early as possible is a valuable goal. Indeed, in many application domains, the earliest the decision, the more rewarding it can be. Yet, often, gathering more information allows one to get a better decision. The optimization of this time vs. accuracy tradeoff must generally be solved online and is a complex problem. This paper presents a formal criterion that expresses this trade-off in all generality together with a generic sequential meta algorithm to solve it. This meta algorithm is interesting in two ways. First, it pinpoints where choices can (have to) be made to obtain a computable algorithm. As a result a wealth of algorithmic solutions can be found. Second, it seeks online the earliest time in the future where a minimization of the criterion can be expected. It thus goes beyond the classical approaches that myopically decide at each time step whether to make a decision or to postpone the call one more time step. After this general setting has been expounded, we study one simple declination of the meta-algorithm, and we show the results obtained on synthetic and real time series data sets chosen for their ability to test the robustness and properties of the technique. The general approach is vindicated by the experimental results, which allows us to point to promising perspectives.

The emergence of Smart Grids is posing a wide range of challenges for electric utility companies and network operators: Integration of non-dispatchable power from renewable energy sources (e.g., photovoltaics, hydro and wind), fundamental changes in the way energy is consumed (e.g., due to dynamic pricing, demand response and novel electric appliances), and more active operations of the networks to increase efficiency and reliability. A key in managing these challenges is the ability to forecast network loads at low levels of locality, e.g., counties, cities, or neighbourhoods. Accurate load forecasts improve the efficiency of supply as they help utilities to reduce operating reserves, act more efficiently in the electricity markets, and provide more effective demand-response measures. In order to prepare for the Smart Grid era, there is a need for a scalable simulation environment which allows utilities to develop and validate their forecasting methodology under various what-if-scenarios. This paper presents a massive-scale simulation platform which emulates electrical load in an entire electrical network, from Smart Meters at individual households, over low- to medium-voltage network assets, up to the national level. The platform supports the simulation of changes in the customer portfolio and the consumers’ behavior, installment of new distributed generation capacity at any network level, and dynamic reconfigurations of the network. The paper explains the underlying statistical modeling approach based on Generalized Additive Models, outlines the system architecture, and presents a number of realistic use cases that were generated using this platform.

The last ten years were prolific in the statistical learning and data mining field and it is now easy to find learning algorithms which are fast and automatic. Historically a strong hypothesis was that all examples were available or can be loaded into memory so that learning algorithms can use them straight away. But recently new use cases generating lots of data came up as for example: monitoring of telecommunication network, user modeling in dynamic social network, web mining, etc. The volume of data increases rapidly and it is now necessary to use incremental learning algorithms on data streams. This article presents the main approaches of incremental supervised classification available in the literature. It aims to give basic knowledge to a reader novice in this subject.

Early classification approaches deal with the problem of reliably labeling incomplete time series as soon as possible given a level of confidence. While developing new approaches for this problem has been getting increasing attention recently, their evaluation are still not thoroughly considered. In this article, we propose a new evaluation protocol for early classifiers. This protocol is generic and does not depend on the criteria used to evaluate the classifiers. Our protocol is successfully applied to 23 publicly available data sets.

In France, the currently emerging “smart grid” and more particularly the 35 millions of “smart meters” will produce a large amount of daily updated metering data. The main french provider of electricity (EDF) is interested by compact and generic representations of time series which allow to accelerate the processing of data. This article proposes a new data-driven symbolic representation of time series named SAXO, where each symbol represents a typical distribution of data points. Furthermore, the time dimension is optimally discretized into intervals by using a parameter free Bayesian coclustering approach (MODL). SAXO is favorably compared with the SAX representation by evaluating a classifier trained from recoded datasets. Our experiments highlight a significant gap in performance between both approaches.

In recent years, the amount of data to process has increased in many application areas such as network monitoring, web click and sensor data analysis. Data stream mining answers to the challenge of massive data processing, this paradigm allows for treating pieces of data on the fly and overcomes exhaustive data storage. The detection of changes in a data stream distribution is an important issue which application area is wide. In this article, change detection problem is turned into a supervised learning task. We chose to exploit the supervised discretization method “MODL” given its interesting properties. Our approach is favorably compared with an alternative method on artificial data streams, and is applied on real data streams.

The labeling of training examples could be a costly task in numerous cases of supervised learning. Active learning strategies address this problem and select unlabeled examples which are considered as the most useful for the training of a predictive model. The choice of examples to be labeled can be considered as a dilemma between the exploration and the exploitation of the input data space. In this article, a new active learning strategy that manages this compromise is proposed. This strategy is based on a Bayesian formalism that minimizes assumptions on data. An experimental validation is conducted on a unidimensional dataset, the objective is to assess the position of a step function from noisy examples. Our approach is favorably compared to an ad hoc strategy : the probabilistic dichotomy.

Date du colloque: 12/2008; International audience; Semi-supervised classification methods aim to exploit labelled and unlabelled examples to train a predictive model. Most of these approaches make assumptions on the distribution of classes. This article first proposes a new semi-supervised discretization method which adopts very low informative prior on data. This method discretizes the numerical domain of a continuous input variable, while keeping the information relative to the prediction of classes. Then, an in-depth comparison of this semi-supervised method with the original supervised MODL approach is presented. We demonstrate that the semi-supervised approach is asymptotically equivalent to the supervised approach, improved with a post-optimization of the intervals bounds location.

Machine learning indicates methods and algorithms which allow a model to learn a behavior thanks to examples. Active learning gathers methods which select examples used to build a training set for the predictive model. All the strategies aim to use the less examples as possible and to select the most informative examples. After having formalized the active learning problem and after having located it in the literature, this article synthesizes in the first part the main approaches of active learning. Taking into account emotions in Human-machine interactions can be helpful for intelligent systems designing. The main difficulty, for the conception of calls center's automatic shunting system, is the cost of data labeling. The last section of this paper propose to reduce this cost thanks to two active learning strategies. The study is based on real data resulting from the use of a vocal stock exchange server.