Search

We propose a quantitative approach for categorising mammalian diets based on the taxonomy of food items and parts consumed (the form of food). Our analysis covers 82% of the mammalian species alive today. The diet information comes from different data sources—textbooks, datasets and peer-reviewed literature and includes a transformation of narrative quantitative data into qualitative data. We link a database on nutrient composition of diet items of the quantitative diet data and analyse the distribution of macronutrients of diets across taxonomic groups and map them to the dental morphology of the eaters. The results show associations between dental complexity and the concentrations of some nutrients. Our analysis highlights omnivory as a multi-faceted concept—there are many kinds of omnivores within the dietary space we report. The developed dataset and the proposed approach relating the chemical composition of diets offers a basis for future comparative studies of living and fossil mammals. With this study, we make the accompanying large-scale dietary data publicly available online (https://www.mammalbase.net).

We know that the fossil record is incomplete. But how incomplete? Here we very coarsely estimate the completeness of the mammalian record in the Miocene, assuming that the duration of a mammalian species is about 1 Myr and the species diversity has stayed constant and is structurally comparable to the taxonomic diversity today. The overall completeness under these assumptions appears to be around 4%, but there are large differences across taxonomic groups. We find that the fossil record of proboscideans and perissodactyls as we know it for the Miocene must be close to complete, while we might know less than 15% of the species of artiodactyl or carnivore fossil species and only about 1% of primate species of the Miocene. The record of small mammals appears much less complete than that of large mammals.

The Red Queen's hypothesis portrays evolution as a never-ending competition for expansive energy, where one species’ gain is another species’ loss. The Red Queen is neutral with respect to body size, implying that neither small nor large species have a universal competitive advantage. Here we ask whether, and if so how, the Red Queen's hypothesis really can accommodate differences in body size. The maximum population growth in ecology clearly depends on body size—the smaller the species, the shorter the generation length, and the faster it can expand given sufficient opportunity. On the other hand, large species are more efficient in energy use due to metabolic scaling and can maintain more biomass with the same energy. The advantage of shorter generation makes a wide range of body sizes competitive, yet large species do not take over. We analytically show that individuals consume energy and reproduce in physiological time, but need to compete for energy in real time. The Red Queen, through adaptive evolution of populations, balances the pressures of real and physiological time. Modeling competition for energy as a proportional prize contest from economics, we further show that Red Queen's zero-sum game can generate unimodal hat-like patterns of species rise and decline that can be neutral in relation to body size.

The Guadix-Baza Basin (GBB) in Andalucía, Spain, comprises palaeontological and archaeological sites dating from the Early Pliocene to the Middle Pleistocene, including some of the earliest sites with evidence for the presence of early humans (Homo sp.) in Europe. Thus, the history of climate and environments in this basin contributes significantly to our understanding of the conditions under which early humans spread into Europe during the Early Pleistocene. Here we present estimates of precipitation and primary productivity in the GBB from the Pliocene to the Middle Pleistocene based on dental ecometrics in fossil communities of large herbivorous mammals, and perform an ecometrics-based distribution modelling to analyse the environmental conditions of Early and Middle Pleistocene human sites in Europe. Our results show that Early Pleistocene humans generally occupied on average relatively diverse habitats with ecotones, such as woodlands and savannas, but avoided very open and harsh (cool or dry) environments. During the Middle Pleistocene in Europe, humans occupied a comparatively much broader range of environments than during the Early Pleistocene, but were on average more concentrated in environments where the dental ecometric of mammals indicate wooded palaeoenvironments. In the earliest human occupation sites of the GBB, Barranco Leon and Fuente Nueva 3, the mean annual precipitation and net primary production estimates indicate climatic conditions close to modern Mediterranean sclerophyllous woodland environments, but with slightly higher primary productivity,indicating some similarity with East African woodlands. On the other hand, the environments did not resemble African grassland savannas. The browse-dominated diets of ungulates from Barranco Leon and Fuente Nueva 3 further suggest palaeoenvironments where grasses were a minor component of the vegetation. In the slightly older site of Venta Micena that has no evidence for the presence of hominins, dental ecometric estimates indicate climate and environments similar to Mediterranean “forest steppe” environments existing in the surroundings of Baza today. Grasses were prevalent in the diet of some taxa, especially equids, in Venta Micena, but most of the species show browse-dominated diets even there.

The global distribution of vegetation is largely determined by climatic conditions and feeds back into the climate system. To predict future vegetation changes in response to climate change, it is crucial to identify and understand key patterns and processes that couple vegetation and climate. Dynamic global vegetation models (DGVMs) have been widely applied to describe the distribution of vegetation types and their future dynamics in response to climate change. As a process- based approach, it partly relies on hard- coded climate thresholds to constrain the distribution of vegetation. What thresholds to implement in DGVMs and how to replace them with more processbased descriptions remain among the major challenges. In this study, we employ machine learning using decision trees to extract large- scale relationships between the global distribution of vegetation and climatic characteristics from remotely sensed vegetation and climate data. We analyse how the dominant vegetation types are linked to climate extremes as compared to seasonally or annually averaged climatic conditions. The results show that climate extremes allow us to describe the distribution and eco- climatological space of the vegetation types more accurately than the averaged climate variables, especially those types which occupy small territories in a relatively homogeneous ecological space. Future predicted vegetation changes using both climate extremes and averaged climate variables are less prominent than that predicted by averaged climate variables and are in better agreement with those of DGVMs, further indicating the importance of climate extremes in determining geographic distributions of different vegetation types. We found that the temperature thresholds for vegetation types (e.g. grass and open shrubland) in cold environments vary with moisture conditions. The coldest daily maximum temperature (extreme cold day) is particularly important for separating many different vegetation types. These findings highlight the need for a more explicit representation of the impacts of climate extremes on vegetation in DGVMs. climate extremes , climate thresholds , decision trees , DGVMs , machine learning , vegetation distribution

Machine learning (ML) provides a powerful set of tools that can improve the accuracy of distribution models by automatically representing underlying ecological relationships empirically captured from large sets of data. However, more recent methodological advances in ML that have been less frequently applied, e.g. in the field of deep learning, may yield additional potential for Distribution Modeling. In this project, we use two ML algorithms, Random Forest (RF) and multi-layer feed-forward artificial neural networks (ANN), to predict the occurrences of Vegetation Types (VT) across Norway. Accurate predictions may support environmental management or the validation of earth system models. The VT data (derived from the AR18x18 data set; n=31 classes) covers the entire spatial scope in 0.9 ha plots on a systematically sampled 18 km grid (n = 1,081 plots, n = 22,154 observations). It was obtained through a field-based survey by a group of trained experts between 2004 and 2014. We use the cloud-based platform "Google Earth Engine" to generate a set of remotely sensed predictor variables based on SENTINEL-2 satellite imagery (i.e. surface reflectance from 12 spectral bands and six vegetation indices). These are then combined with ancillary environmental rasters used previously to model Norwegian VT distribution, e.g. representing climate, land cover, or geological properties (n=55, before one-hot encoding). Preliminary results suggest that in both modeling approaches, the generated SENTINEL-2 variables, particularly the Normalized Difference Vegetation Index (NDVI), have the highest predictive power as measured by permutation importance. The mean overall accuracy using 5-fold cross-validation shows only minor differences between the two methods (approx. 0.45 for ANN vs. 0.44 for RF; the respective F1-scores are 0.35 for ANN and 0.34 for RF; most frequent class baseline accuracy = 0.136). The modeling challenges we currently face include a class imbalance in the VT data set, reconciling the different spatial resolutions of the environmental predictors, and discrepancies in the timing of data acquisition. The next steps in the project will be to incorporate spatial cross-validation into the workflow and to analyze the differences between the ML methods in detail (e.g. regarding the ability to model rare VTs or differences in variable importance). Moreover, we will evaluate the possibility to include additional satellite data sources. This work is a contribution to the Strategic Research Initiative ‘Land Atmosphere Interaction in Cold Environments’ (LATICE) of the University of Oslo.

We study how the behavior and wing-beat frequency of hymenopteran flying insects depend on environmental conditions, such as temperature and relative humidity. We use flight data from seven bee species and two wasp species collected in Brazil in a completely non-invasive way under field conditions employing a low-cost optical sensor. With this data, we demonstrate that it is possible to accurately classify each insect species into two groups based on their flight frequency signal. The signals, however, are not the same under distinct environmental conditions. We statistically show that the bee and wasp wing-beat frequencies are significantly different, given distinct temperature and humidity ranges. This study’s findings, along with our discussion of the species’ flight activity, are crucial elements in developing intelligent methods for automatic detection, recognition, and monitoring of flying insects.

A new study by Fraser et al (2018) urges the use of phylogenetic comparative methods, whenever possible, in analyses of mammalian tooth wear. We are concerned about this for two reasons. First, this recommendation may mislead the research community into thinking that phylogenetic signal is an artifact of some sort rather than a fundamental outcome of the evolutionary process. Secondly, this recommendation may set a precedent for editors and reviewers to enforce phylogenetic adjustment where it may unnecessarily weaken or even directionally alter the results, shifting the emphasis of analysis from common patterns manifested by large clades to rare cases.

How food intake and mastication scale to satisfy the metabolic needs of mammals has been the subject of considerable scientific debate. Existing theory suggests that the negative allometric scaling of metabolic rate with body mass is compensated by a matching allometric scaling of the chewing rate. Why empirical studies have found that the scaling coefficients of the chewing rate seem to be systematically smaller than expected from theory remains unknown. Here we explain this imparity by decoupling the functional surface area of teeth from overall surface area. The functional surface area is relatively reduced in forms emphasizing linear edges (e.g., lophodont) compared with forms lacking linear structures (e.g., bunodont). In forms with reduced relative functional surface, the deficit in food processed per chew appears to be compensated for by increased chewing rate, such that the metabolic requirements are met. This compensation accounts for the apparent difference between theoretically predicted and observed scaling of chewing rates. We suggest that this reflects adaptive functional evolution to plant foods with different fracture properties and extend the theory to incorporate differences in functional morphology. (C) 2017 Elsevier Ltd. All rights reserved.

Human society relies on four main livestock animals – sheep, goat, pig and cattle, which all were domesticated at nearly the same time and place. Many arguments have been put forward to explain why these animals, place and time were suitable for domestication, but the question – why only these, but not other animals, still does not have a clear answer. Here we offer a biogeographical perspective: we survey global occurrence of large mammalian herbivore genera around 15 000 – 5 000 years before present and compile a dataset characterising their ecology, habitats and dental traits. Using predictive modelling we extract patterns from this data to highlight ecological differences between domesticated and nondomesticated genera. The most suitable for domestication appear to be generalists adapted to persistence in marginal environments of low productivity, largely corresponding to cold semi-arid climate zones. Our biogeographic analysis shows that even though domestication rates varied across continents, potentially suitable candidate animals were rather uniformly distributed across continents. We interpret this pattern as a result of an interface between cold semi-arid and hot semi-arid climatic zones. We argue that hot Semi-arid climate was most suitable for plant domestication, cold Semi-arid climate selected for animals most suitable for domestication as livestock. We propose that the rates of domestication across biogeographic realms largely reflect how much intersection between hot and cold Semi-arid climatic zones was available at each continent.

Fossils are the remains organisms from earlier geological periods preserved in sedimentary rock. The global fossil record documents and characterizes the evidence about organisms that existed at different times and places during the Earth's history. One of the major directions in computational analysis of such data is to reconstruct environmental conditions and track climate changes over millions of years. Distribution of fossil animals in space and time make informative features for such modeling, yet concept drift presents one of the main computational challenges. As species continuously go extinct and new species originate, animal communities today are different from the communities of the past, and the communities at different times in the past are different from each other. The fossil record is continuously increasing as new fossils and localities are being discovered, but it is not possible to observe or measure their environmental contexts directly, because the time is gone. Labeled data linking organisms to climate is available only for the present day, where climatic conditions can be measured. The approach is to train models on the present day and use them to predict climatic conditions over the past. But since species representation is continuously changing, transfer learning approaches are needed to make models applicable and climate estimates to be comparable across geological times. Here we discuss predictive modeling settings for such paleoclimate reconstruction from the fossil record. We compare and experimentally analyze three baseline approaches for predictive paleoclimate reconstruction: (1) averaging over habitats of species, (2) using presence-absence of species as features, and (3) using functional characteristics of species communities as features. Our experiments on the present day African data and a case study on the fossil data from the Turkana Basin over the last 7 million of years suggest that presence-absence approaches are the most accurate over short time horizons, while species community approaches, also known as ecometrics, are the most informative over longer time horizons when, due to ongoing evolution, taxonomic relations between the present day and fossil species become more and more uncertain.

Statistical models for environmental monitoring strongly rely on automatic data acquisition systems that use various physical sensors. Often, sensor readings are missing for extended periods of time, while model outputs need to be continuously available in real time. With a case study in solar-radiation nowcasting, we investigate how to deal with massively missing data (around 50% of the time some data are unavailable) in such situations. Our goal is to analyze characteristics of missing data and recommend a strategy for deploying regression models which would be robust to missing data in situations where data are massively missing. We are after one model that performs well at all times, with and without data gaps. Due to the need to provide instantaneous outputs with minimum energy consumption for computing in the data streaming setting, we dismiss computationally demanding data imputation methods and resort to a mean replacement, accompanied with a robust regression model. We use an established strategy for assessing different regression models and for determining how many missing sensor readings can be tolerated before model outputs become obsolete. We experimentally analyze the accuracies and robustness to missing data of seven linear regression models. We recommend using the regularized PCA regression with our established guideline in training regression models, which themselves are robust to missing data.

International audience; As organisms are adapted to their environments, assemblages of taxa can be used to describe environments in the present and in the past. Here, we use a data mining method, namely redescription mining, to discover and analyze patterns of association between large herbivorous mammals and their environments via their functional traits. We focus on functional properties of animal teeth, characterized using a recently developed dental trait scoring scheme. The teeth of herbivorous mammals serve as an interface to obtain energy from food, and are therefore expected to match the types of plant food available in their environment. Hence, dental traits are expected to carry a signal of environmental conditions. We analyze a global compilation of occurrences of large herbivorous mammals and of bioclimatic conditions. We identify common patterns of association between dental traits distributions and bioclimatic conditions and discuss their implications. Each pattern can be considered as a computational biome. Our analysis distinguishes three global zones, which we refer to as the boreal-temperate moist zone, the tropical moist zone and the tropical-subtropical dry zone. The boreal-temperate moist zone is mainly characterized by seasonal cold temperatures, a lack of hypsodonty and a high share of species with obtuse lophs. The tropical moist zone is mainly characterized by high temperatures, high isothermality, abundant precipitation and a high share of species with acute rather than obtuse lophs. Finally, the tropical dry zone is mainly characterized by a high seasonality of temperatures and precipitation, as well as high hypsodonty and horizodonty. We find that the dental traits signature of African rain forests is quite different from the signature of climatically similar sites in North America and Asia, where hypsodont species and species with obtuse lophs are mostly absent. In terms of climate and dental signatures, the African seasonal tropics share many similarities with Central-South Asian sites. Interestingly, the Tibetan plateau is covered both by redescriptions from the tropical-subtropical dry group and by redescriptions from the boreal-temperate moist group, suggesting a combination of features from both zones in its dental traits and climate.

Predicting mobility patterns requires user‐specific accuracy measures that would take into account changing user behavior over time.

Most suids (Mammalia: Suidae, pigs) worldwide are omnivores living in closed environments, but the African warthog (Phacochoerus) has special adaptations for grazing in open environments. Similar specializations have been recorded from Plio-Pleistocene African suids. Four genera, Nyanzachoerus, Notochoerus, Kolpochoerus, and Metridiochoerus, have been discovered in late Miocene to middle Pleistocene locations around the Turkana Basin. We analyse the relative abundances of these four suid genera compared to other mammals, from approximately 8-0.7 Ma. The dataset includes most of the mammal specimens collected from locations around the Kenyan side of the Turkana Basin. Species of genus Nyanzachoerus were dominant before 4 Ma, but their relative abundance decreases through time thereafter. At the same time, Notochoerus increases in relative abundance, followed by Kolpochoerus, and finally Metridiochoerus. Their peak relative abundances do not overlap: Notochoerus peaks at 3.44-2.53 Ma, Kolpochoerus at 2.53-1.87 Ma, and Metridiochoerus at 1.38-0.7 Ma. We interpret the palaeoecology of these suids based on their relative abundances over time and on published isotope and pollen data. We find that Nyanzachoerus was replaced by its abrasive-diet-specialized successor Notochoerus, possibly in response to a rapid decrease in forest cover. Notochoerus adapted at first to the expanding wood- and grasslands, and then to more arid shrublands. After a period of severe aridity around 2.7-2.5 Ma, more variable environments allowed Kolpochoerus and Metridiochoerus to disperse, while Notochoerus disappeared, perhaps having lost its competitive edge. Further changes in the environment encouraged the expansion of grasslands over shrublands, favouring Metridiochoerus. Kolpochoerus persisted in the more closed environments near water sources.

We study evaluation of supervised learning models that adapt to changing data distribution over time (concept drift). The standard testing procedure that simulates online arrival of data (test-then-train) may not be sufficient to generalize about the performance, since that single test concludes how well a model adapts to this fixed configuration of changes, while the ultimate goal is to assess the adaptation to changes that happen unexpectedly. We propose a methodology for obtaining datasets for multiple tests by permuting the order of the original data. A random permutation is not suitable, as it makes the data distribution uniform over time and destroys the adaptive learning task. Therefore, we propose three controlled permutation techniques that make it possible to acquire new datasets by introducing restricted variations in the order of examples. The control mechanisms with theoretical guarantees of preserving distributions ensure that the new sets represent close variations of the original learning task. Complementary tests on such sets allow to analyze sensitivity of the performance to variations in how changes happen and this way enrich the assessment of adaptive supervised learning models.

A major focus in evolutionary biology is to understand how the evolution of organisms relates to changes in their physical environment. In the terrestrial realm, the interrelationships among climate,vegetation, and herbivores lie at the heart of this question. Here we introduce and test a scoring scheme for functional traits present on theworn surfaces of large mammalian herbivore teeth to capture their relationship to environmental conditions. We modeled local precipitation, temperature, primary productivity, and vegetation index as functions of dental traits of large mammal species in 13 national parks in Kenya over the past 60 y. We found that these dental traits can accurately estimate local climate and environment, even at small spatial scales within areas of relatively uniform climate (within two ecoregions), and that they predict limiting conditions better than average conditions. These findings demonstrate that the evolution of key functional properties of organisms may be more reflective of demands during recurring adverse episodes than under average conditions or during isolated severe events. This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1609409113/-/DCSupplemental

Author correction: vol 2, pg, 402, 2018 DOI:10.1038/s41559-018-0468-8 Despite much interest in the ecology and origins of the extensive grassland ecosystems of the modern world, the biogeographic relationships of savannah palaeobiomes of Africa, India and mainland Eurasia have remained unclear. Here we assemble the most recent data from the Neogene mammal fossil record in order to map the biogeographic development of Old World mammalian faunas in relation to palaeoenvironmental conditions. Using genus-level faunal similarity and mean ordinated hypsodonty in combination with palaeoclimate modelling, we show that savannah faunas developed as a spatially and temporally connected entity that we term the Old World savannah palaeobiome. The Old World savannah palaeobiome flourished under the influence of middle and late Miocene global cooling and aridification, which resulted in the spread of open habitats across vast continental areas. This extensive biome fragmented into Eurasian and African branches due to increased aridification in North Africa and Arabia during the late Miocene. Its Eurasian branches had mostly disappeared by the end of the Miocene, but the African branch survived and eventually contributed to the development of Plio-Pleistocene African savannah faunas, including their early hominins. The modern African savannah fauna is thus a continuation of the extensive Old World savannah palaeobiome.

Watson and Szathmary [1] have presented an intriguing idea linking evolution through natural selection to algorithmic learning. They argue that algorithmic learning can provide models for evolutionary processes; specifically, that evolution of evolvability is similar to generalization in supervised learning (neural networks as an example), that evolution of ecological organization is analogous to unsupervised learning (clustering), and that evolutionary transitions in individuality are comparable to specific forms of deep learning, namely learning of multilayered model structures.

Although ecometric methods have been used to analyse fossil mammal faunas and environments of Eurasia and North America, such methods have not yet been applied to the rich fossil mammal record of eastern Africa. Here we report results from analysis of a combined dataset spanning east and west Turkana from Kenya between 7 and 1 million years ago (Ma). We provide temporally and spatially resolved estimates of temperature and precipitation and discuss their relationship to patterns of faunal change, and propose a new hypothesis to explain the lack of a temperature trend. We suggest that the regionally arid Turkana Basin may between 4 and 2 Ma have acted as a ‘species factory’, generating ecological adaptations in advance of the global trend. We show a persistent difference between the eastern and western sides of the Turkana Basin and suggest that the wetlands of the shallow eastern side could have provided additional humidity to the terrestrial ecosystems. Pending further research, a transient episode of faunal change centred at the time of the KBS Member (1.87–1.53 Ma), may be equally plausibly attributed to climate change or to a top-down ecological cascade initiated by the entry of technologically sophisticated humans.This article is part of the themed issue ‘Major transitions in human evolution’.

Concept drift is a challenge in supervised learning for sequential data. It describes a phenomenon when the data distributions change over time. In such a case accuracy of a classifier benefits from the selective sampling for training. We develop a method for training set selection, particularly relevant when the expected drift is gradual. Training set selection at each time step is based on the distance to the target instance. The distance function combines similarity in space and in time. The method determines an optimal training set size online at every time step using cross validation. It is a wrapper approach, it can be used plugging in different base classifiers. The proposed method shows the best accuracy in the peer group on the real and artificial drifting data. The method complexity is reasonable for the field applications. Keywords: Concept drift, gradual drift, online learning, instance selection

In the version of this Article originally published, each of the five panels in Fig. 5 incorrectly contained a black diagonal line across the plot. This has now been corrected.

Classification in changing environments (commonly known as concept drift) requires adaptation of the classifier to accommodate the\ud changes. One approach is to keep a moving window on the streaming data and constantly update the classifier on it. Here we consider an\ud abrupt change scenario where one set of probability distributions of the classes is instantly replaced with another. For a fixed ‘transition\ud period’ around the change, we derive a generic relationship between the size of the moving window and the classification error rate. We\ud derive expressions for the error in the transition period and for the optimal window size for the case of two Gaussian classes where the\ud concept change is a geometrical displacement of the whole class configuration in the space. A simple window resize strategy based\ud on the derived relationship is proposed and compared with fixed-size windows on a real benchmark data set data set (Electricity Market).

In most challenging data analysis applications, data evolve over time and must be analyzed in near real time. Patterns and relations in such data often evolve over time, thus, models built for analyzing such data quickly become obsolete over time. In machine learning and data mining this phenomenon is referred to as concept drift. The objective is to deploy models that would diagnose themselves and adapt to changing data over time. This chapter provides an application oriented view towards concept drift research, with a focus on supervised learning tasks. First we overview and categorize application tasks for which the problem of concept drift is particularly relevant. Then we construct a reference framework for positioning application tasks within a spectrum of problems related to concept drift. Finally, we discuss some promising research directions from the application perspective, and present recommendations for application driven concept drift research and development.

Concept drift primarily refers to an online supervised learning scenario when the relation between the in- put data and the target variable changes over time. Assuming a general knowledge of supervised learning in this paper we characterize adaptive learning process, categorize existing strategies for handling concept drift, discuss the most representative, distinct and popular techniques and algorithms, discuss evaluation methodology of adaptive algorithms, and present a set of illustrative applications. This introduction to the concept drift adaptation presents the state of the art techniques and a collection of benchmarks for re- searchers, industry analysts and practitioners. The survey aims at covering the different facets of concept drift in an integrated way to reflect on the existing scattered state-of-the-art.

Automatic data acquisition systems provide large amounts of streaming data generated by physical sensors. This data forms an input to computational models (soft sensors) routinely used for monitoring and control of industrial processes, traffic patterns, environment and natural hazards, and many more. The majority of these models assume that the data comes in a cleaned and pre-processed form, ready to be fed directly into a predictive model. In practice, to ensure appropriate data quality, most of the modelling efforts concentrate on preparing data from raw sensor readings to be used as model inputs. This study analyzes the process of data preparation for predictive models with streaming sensor data. We present the challenges of data preparation as a four-step process, identify the key challenges in each step, and provide recommendations for handling these issues. The discussion is focused on the approaches that are less commonly used, while, based on our experience, may contribute particularly well to solving practical soft sensor tasks. Our arguments are illustrated with a case study in the chemical production industry.

International audience; Data labeling is an expensive andt ime-consuming task. Choosing which labels to use is increasingly becoming important. In the active learning setting, a classifier is trained by asking for labels for only a small fraction of all instances. While many works exist that deal with this issue in non-streaming scenarios, few works exist in the data stream setting. In this paper we propose a new active learning approach for evolving data streams based on a pre-clustering step, for selecting the most informative instances for labeling. We consider a batch incremental setting: when a new batch arrives, first we cluster the examples, and then, we select the best instances to train the learner. The clustering approach allows to cover the whole data space avoiding to oversample examples from only few areas. We compare our method w.r.t. state of the art active learning strategies over real datasets. The results highlight the improvement in performance of our proposal. Experiments on parameter sensitivity are also reported.

Data stream classification plays an important role in modern data analysis, where data arrives in a stream and needs to be mined in real time. In the data stream setting the underlying distribution from which this data comes may be changing and evolving, and so classifiers that can update themselves during operation are becoming the state-of-the-art. In this paper we show that data streams may have an important temporal component, which currently is not considered in the evaluation and benchmarking of data stream classifiers. We demonstrate how a naive classifier considering the temporal component only outperforms a lot of current state-of-the-art classifiers on real data streams that have temporal dependence, i.e. data is autocorrelated. We propose to evaluate data stream classifiers taking into account temporal dependence, and introduce a new evaluation measure, which provides a more accurate gauge of data stream classifier performance. In response to the temporal dependence issue we propose a generic wrapper for data stream classifiers, which incorporates the temporal component into the attribute space.

Nowadays more and more decisions in lending, recruitment, grant or study applications are partially being automated based on computational models (classifiers) premised on historical data. If the historical data was discriminating towards socially and legally protected groups, a model learnt over this data will make discriminatory decisions in the future. As a solution, most of the discrimination free modeling techniques force the treatment of the sensitive groups to be equal and do not take into account that some differences may be explained by other factors and thus justified. For example, disproportional recruitment rates for males and females may be explainable by the fact that more males have higher education; treating males and females equally will introduce reverse discrimination, which may be undesirable as well. Given that the law or domain experts specify which factors are discriminatory (e.g. gender, marital status) and which can be used for explanation (e.g. education), this chapter presents a methodology how to quantify the tolerable difference in treatment of the sensitive groups. We instruct how to measure, which part of the difference is explainable and present the local learning techniques that remove exactly the illegal discrimination, allowing the differences in decisions to be present as long as they are explainable.

Many systems rely on predictive models using sensor data, with sensors being prone to occasional failures. From the operational point of view predictions need to be tolerant to sensor failures such that the loss in accuracy due to temporary missing sensor readings would be minimal. In this paper, we theoretically and empirically analyze robustness of linear predictive models to temporary missing data. We demonstrate that if the input sensors are correlated the mean imputation of missing values may lead to a very rapid deterioration of the prediction accuracy. Based on the theoretical results we introduce a quantitative measure that allows to assess how robust is a given linear regression model to sensor failures. We propose a practical strategy for building and operating robust linear models in situations when temporal sensor failures are expected. Experiments on six sensory datasets and a case study in environmental monitoring with streaming data validate the theoretical results and confirm the effectiveness of the proposed strategy.

In learning to classify streaming data, obtaining the true labels may require major effort and may incur excessive cost. Active learning focuses on learning an accurate model with as few labels as possible. Streaming data poses additional challenges for active learning, since the data distribution may change over time (concept drift) and classifiers need to adapt. Conventional active learning strategies concentrate on querying the most uncertain instances, which are typically concentrated around the decision boundary. If changes do not occur close to the boundary, they will be missed and classifiers will fail to adapt. In this paper we develop two active learning strategies for streaming data that explicitly handle concept drift. They are based on uncertainty, dynamic allocation of labeling efforts over time and randomization of the search space. We empirically demonstrate that these strategies react well to changes that can occur anywhere in the instance space and unexpectedly.

We present concept drift responsive method for classifier training for sequential data. Relevant instance selection for training is based on similarity to the target observation. Similarity in space and in time is combined. The algorithm determines an optimal training set size. It can be used plugging in different base classifiers. The proposed algorithm shows the best accuracy in the peer group. The algorithm complexity is reasonable for the field applications.

Fuel feeding and inhomogeneity of fuel typically cause fluctuations in the circulating fluidized bed (CFB) process. If control systems fail to compensate the fluctuations, the whole plant will suffer from dynamics that is reinforced by the closed-loop controls. This phenomenon causes reducing efficiency and the lifetime of process components. In this paper we address the problem of online mass flow prediction, which is a part of control. Particularly, we consider the problem of learning an accurate predictor with explicit detection of abrupt concept drift and noise handling mechanisms. We emphasize the importance of having domain knowledge concerning the considered case and constructing the ground truth for facilitating the quantitative evaluation of different approaches. We demonstrate the performance of change detection methods and show their effect on the accuracy of the online mass flow prediction with real datasets collected from the experimental laboratory-scale CFB boiler.