International audience

International audience

The use of Artificial Intelligence (AI), notably Machine Learning (ML), is gaining momentum in archaeology, opening up new possibilities such as artifact classification, site location prediction, and remains analysis. One of the major challenges in this regard is the lack of qualified archaeologists who are experts in machine learning. In this study, we introduce IArch, a tool that enables eXplainable Artificial Intelligence (XAI) data analytics for archaeologists without requiring specific programming skills. It specifically allows data analysis performed to either validate existing data-supported hypotheses or generate new ones. The tool covers the entire workflow for applying ML, from data processing to explaining the final results. The tool allows the use of supervised and unsupervised ML algorithms, as well as the SHapley Additive exPlanations (SHAP) technique to provide archaeologists with global and individual explanations for the predictions. We demonstrate its use on data from a Xiongnu cemetery (100 BC/AD 100) in the Mongolian steppes.

International audience

International audience

International audience

The high failure rate is a common issue among online institutions. Early Warning Systems (EWSs) are widely adopted as a solution to deal with this issue. However, these systems do not go beyond the early identification of failing learners. In this paper, we propose a new alert algorithm of an educational EWS for generating risk alerts at the earliest. This algorithm is based on a weekly prediction model that aims to generate early alerts. The regular tracking of prediction results enabled to propose measures for the right prediction earliness and the model’s temporal stability. These measures prepare the last step of the algorithm which is the alerts generation according to a predefined rule. The objective of this rule is to target at-risk learners to improve their learning. For this aim, we used data of k-12 learners enrolled in an online physics-chemistry module.

The exploration and exploitation (E&E) of a search space are fundamental processes in many fields of artificial intelligence. Indeed, it is important to ensure that many regions of the search space are examined, so as not to get trapped in a local optimum, but also that the promising regions are examined more in depth in order to find good local optima. Metaheuristics (MHs) are popular to perform E&E tasks. Explaining the dynamics of both processes is a repeated demand of researchers and practitioners, but remains a challenge. This paper presents an approach giving information about how a MH explores and exploits the search space. This approach is based on a way of representing transitions through presence variation vectors and associated proposed indicators, used to make transitions between E&E explicit and intelligible. Experimentations show the relevance of the proposed indicators.

The rapid and intense development of distance learning in recent years has led to increasingly comprehensive solutions, recording students’ activity in the form of learning traces. Adapted learning systems can exploit this data to analyze students’ behavior and help them be more successful in their learning process. In this paper, we propose an approach that is new in data mining, which consists of representing the behavioral dynamics of a set of students through Behavioral Primitives. A Behavioral Primitive represents the temporal evolution of learning indicators to model the behavior of a group of students. The primitive then corresponds to a dynamic distribution, representing both the mean values and the variance of specific indicators over time. We experiment this method on a known, widely used, and open dataset (OULAD). Our results confirm the relevance of the proposed methodology to characterize students’ behaviors intelligibly and visually while respecting each student's anonymity. Our work provides a powerful and explainable tool for educational actors. It allows them to analyze learners’ behaviors and perform pedagogical actions.

Syphilis is a Sexually Transmitted Infection (STI) which the Brazilian Ministry of Health has acknowledged an epidemic since 2016. To face such a problem, it is essential to develop and implement educational actions enhanced by information and communication technologies to qualify, train and raise awareness nationally. Considering the increasing number of Open Educational Resources developed, existing open health repositories and other digital platforms that allow interaction in the Brazilian Unified Health System (SUS) as well as the vast number of Healthcare Information Systems, it is essential to the develop solutions to efficiently and costly recommend content accordingly to the interest of health professionals and the current needs and priorities of the SUS, such as the epidemic of syphilis. This paper presents a discussion on the information systems and the data available and strategies of recommendation systems based on this scenario, integrating health surveillance, formative needs, georeference of health teams and professionals and epidemiological data to recommend content to health professionals all over the country.

For many decades society did need to monitor and assess the standard of living of the population. In the 1950s, the United Nations (UN) saw this need and proposed 12 areas that should be evaluated, the first of which is listed under "Health and Demography", which focuses on what is expressed as the level of a population's health. Decades have passed and great results have been gained from similar initiatives such as reducing mortality from infectious diseases and even eradicating some others. In the age of the digital society, needs have grown. Monitoring demands that once perished from data to become concrete now suffer from the opposite effect, the excess of data from everywhere. Healthcare systems around the world use many different information systems, collecting and generating hundreds of data at unimaginable speed. We are billions of people on the planet and most of us are connected to the virtual world, sharing information, experiences and events with some kind of cloud. In this information age, the ability to aggregate and process this data is a major factor in raising public health to a new level. The development of tools capable of analyzing a large volume of data in seconds and producing knowledge for targeted decision making can help in the fight against specific diseases, in the process of continuing education of professionals, in the formation of new professionals, in the elaboration of new policies. with the specific locoregional look, in the analysis of hidden trends in front of so much information faced in everyday life and other possibilities. The present work proposes an architecture capable of storing and manipulating seeking to standardize the variables in order to allow to correlate this large amount of data in a systematic way, providing to several services and researchers the possibility of consuming health, social, economic and educational data for the promotion of public health.

In this paper, we rely on the Cognitive Load Theory and explore how multimodal data can be used to measure students' effort at the task level. Different from what we expected, the subjective effort ratings have a higher correlation with the students' scores, while the behavioral and physiological data have higher correlations with the scores than with the effort ratings. Moreover, we found that, in the context of our study, ability had a stronger influence on students' success than the prior knowledge, while none of these variables had an influence on the effort ratings. Finally, we propose a new effort model based on students' activity.

Data clustering is an important topic in data science in general, but also in user modeling and recommendation systems. Some clustering algorithms like K-means require the adjustment of many parameters, and force the clustering without considering the clusterability of the dataset. Others, like DBSCAN, are adjusted to a fixed density threshold, so can't detect clusters with different densities. In this paper we propose a new clustering algorithm based on the mutual vote, which adjusts itself automatically to the dataset, demands a minimum of parameterizing, and is able to detect clusters with different densities in the same dataset. We test our algorithm and compare it to other clustering algorithms for clustering users, and predict their purchases in the context of recommendation systems.

Context-aware recommendation became a major topic of interest within the recommender systems community as the context is crucial to provide the right items at the right moment. Many studies aim at developing complex models to include contextual factors in the recommendation process. Despite a real improvement on the recommendations quality, such contextual factors face users' privacy and data collection issues. We support the idea that context could be expressed in term of item attributes rather than contextual factors. To investigate that hypothesis, we designed an online experiment where 174 users were asked to describe the context in which they would listen the proposed songs for which we collected 12 musical attributes. We make available all the material collected during this study for research purposes and non-commercial use.

International audience

In this paper we introduce a new pattern, referred to as “set of contrasting rules”. The main originality of this pattern is that it allows to easily identify trigger factors: factors that can bring some event state changes. In real applications this pattern can thus be used to influence the values of some attributes, what was shown through the experiments conducted on a real dataset of census data.

Data mining, through association rules mining, is one of the best known approaches for patterns identification. However, it results most of the time in a huge set of patterns (rules), so their exploitation is not easy and often requires expert analysis. In this paper we describe a new pattern "set of contrasting rules" which, contrary to most state-of-the-art patterns, has the characteristic of being made up of a set of rules. It has also the advantage of not only identifying a reduced set of rules, but also structuring it into sets. One main originality of this pattern is that it allows to automatically identify trigger factors: factors that can bring some event state changes. In this work we show that the proposed pattern methodologically belongs to the supervised descriptive rules induction paradigm. We also show through the experiments on a real dataset of census data that "set of contrasting rules" can be considered as a way to filter the huge amount of association rules and can be used to identify trigger factors.

Most of recommender systems build their predictions by analysing the preferences of users. However, there are many situations, such as in intelligent tutoring systems, where recommendations of pedagogical resources should rather be based on their memory. So as to infer in real time and with low involvement what has been memorized by users, we highlight in the paper the link between gaze features and visual memory. We designed a user experiment where different subjects had to remember a large set of images. In the meantime, we collected about 19,000 fixation points. Among other metrics, our results show a strong correlation between the relative path angles and the memorized items. It is thus possible to predict the users' memory status by analyzing their gaze data while interacting with the system, so as to provide recommendations that fits their learning curve.

The collaborative filtering (CF) approach in recommender systems assumes that users' preferences are consistent among users. Although accurate, this approach fails on some users. We presume that some of these users belong to a small community of users who have unusual preferences, such users are not compliant with the CF underlying assumption. They are grey sheep users. This paper aims at accurately identifying grey sheep users. We introduce a new distribution-based grey sheep users identification technique, that borrows from outlier detection and from information retrieval, while taking into account the specificities of preference data on which CF relies: extreme sparsity, imprecision and users' bias. The experimental evaluation conducted on a state-of-the-art dataset shows that this new distribution-based technique outperforms state-of-the-art grey sheep users identification techniques.

The main goal of recommender systems is to help users to filter all the information available by suggesting items they may like without they had to find them by themselves. Although the rating prediction is a pretty well controlled topic, being able to make a recommendation at the right moment still remain a challenging task. To this end, most researches try to integrate contextual information (weather, mood, location of users, etc.) in the recommendation process. Even if this process increases users satisfaction, using personal information faces with users' privacy issues. In a different way, our approach is only giving credits to the evolution of diversity within the recent history of consultations, allowing us to automatically detect implicit contexts. In this paper, we will discuss the scientific challenges to be overcome to take maximum advantage of those implicit contexts in the recommendation process.

Recommender systems usually rely on users' preferences. Nevertheless, there are many situations (e-learning, e-health) where recommendations should rather be based on their memory. So as to infer in real time and with low involvement what has been memorized by users, we propose in this paper to establish a link between gaze features and visual memory. We designed a user experiment where 24 subjects had to remember 72 images. In the meantime, we collected 18,643 fixation points. Among other metrics, our results show a strong correlation between the relative path angles and the memorized items.

International audience

Predicting the emergence of an event enables to anticipate and make decisions upstream. For instance, in the employment sector, it becomes necessary to anticipate the emergence of competencies requirements to help job seekers, education and training organization to better match the needs of the job market. Several approaches address the competencies mining with ontologies, we adopt a different point of view by using pattern mining. We propose a new methodology to predict emerging patterns and apply it to competencies with a dataset of job offers collected on the Web. Our model allows to identify potential emerging pattern over time and thus enables to take decisions accordingly.

Being able to automatically and quickly understand the user context during a session is a main issue for recommender systems. As a first step toward achieving that goal, we propose a model that observes in real time the diversity brought by each item relatively to a short sequence of consultations, corresponding to the recent user history. Our model has a complexity in constant time, and is generic since it can apply to any type of items within an online service (e.g. profiles, products, music tracks) and any application domain (e-commerce, social network, music streaming), as long as we have partial item descriptions. The observation of the diversity level over time allows us to detect implicit changes. In the long term, we plan to characterize the context, i.e. to find common features among a contiguous sub-sequence of items between two changes of context determined by our model. This will allow us to make context-aware and privacy-preserving recommendations, to explain them to users. As this is an on-going research, the first step consists here in studying the robustness of our model while detecting changes of context. In order to do so, we use a music corpus of 100 users and more than 210,000 consultations (number of songs played in the global history). We validate the relevancy of our detections by finding connections between changes of context and events, such as ends of session. Of course, these events are a subset of the possible changes of context, since there might be several contexts within a session. We altered the quality of our corpus in several manners, so as to test the performances of our model when confronted with sparsity and different types of items. The results show that our model is robust and constitutes a promising approach.

Episode rules are event patterns mined from a single event sequence. They are mainly used to predict the occurrence of events (the consequent of the rule), once the antecedent has occurred. The occurrence of the consequent of a rule may however be disturbed by the occurrence of another event in the sequence (that does not belong to the antecedent). We refer such an event to as an influencer event. To the best of our knowledge, the identification of such events in the context of episode rules has never been studied. However, identifying influencer events is of the highest importance as these events can be viewed as a way to act to impact the occurrence of events, here the consequent of rules. We propose to identify three types of influencer events: distance influencer events, confidence influencer events and disappearance events. To identify these influencer events, we propose to rely on the set of episode rules discovered by mining algorithms. The proposed approach for discovering influencer events is evaluated on an event sequence of social networks messages. Experiments measure the execution time efficiency according to the adopted episode rules mining algorithm. In addition, they show that some events do actually highly influence the consequent of some rules, that influencer events may not only influence several consequents, but also influence several characteristics of rules.

Since the emergence of learning analytics in North America, researchers and practitioners have worked to develop an international community. The organization of events such as SoLAR Flares and LASI Locals, as well as the move of LAK in 2013 from North America to Europe, has supported this aim. There are now thriving learning analytics groups in North American, Europe and Australia, with smaller pockets of activity emerging on other continents. Nevertheless, much of the work carried out outside these forums, or published in languages other than English, is still inaccessible to most people in the community. This panel, organized by Europe's Learning Analytics Community Exchange (LACE) project, brings together researchers from five European countries to examine the field from European perspectives. In doing so, it will identify the benefits and challenges associated with sharing and developing practice across national boundaries.

International audience

International audience

Social web permits users to acquire information from anonymous people around the world. This leads to a serious question about the trustworthiness of information and sources. During the last decade, numerous models were proposed to model social trust in the service of social web. Trust modeling follows two main axes, local trust (trust between pair of users), and global trust (user's reputation within the community). Subjective logic, is an extension of probabilistic logic that deals with the cases of lack of evidences. An elaborated local trust model based on subjective logic already exists. The aim of this work is to apply this model to the first time on a real data set. Then, we propose another global trust model based also on subjective logic. We apply both models on a real data set of a question answering social network that aims to assist people to find solutions to their technical problems in various domains. Our proposed global trust model ensures a better performance thanks to its precise interpretation of the context of trust, and its ability to satisfy new arrived users.

More and more personalization systems are emerging to reduce the information overload of the Web. As a result, it has become vital to model users' preferences accurately. Our focus lies in the quality of users' expressed preferences, in terms of reliability and stability through time. Today, users are often brought to express their preferences through ratings on a multi-point scale. However, several studies have highlighted problems with ratings. We propose a new preference modality whereby users compare items two-by two ("I prefer x to y").This initial work on comparisons shows that users are in favor of this new preference mechanism and that comparisons are almost 20% more stable over time than those conveyed through ratings, thus more reliable. These encouraging findings let us think that comparisons may lead to a better user modeling and an increase in the quality of personalization services, such as recommender systems.

This paper proposes a new approach of mentor selection in memory-based collaborative filtering when no rating is available. Users are represented under the form of a social network. The selection of mentors is performed through the use of a community detection algorithm used in the frame of social networks. It allows to recommend items to a given user, by applying democratic voting rules within his community.

The development of the Web engendered the emergence of virtual communities. Analyzing information flows and discovering leaders through these communities becomes thus, a major challenge in different application areas. In this paper, we present an algorithm that aims at detecting leaders in the context of behavioral networks. This algorithm considers the high connectivity and the potentiality of propagating accurate appreciations so as to detect reliable leaders through these networks. This approach is evaluated in terms of precision using a real usage dataset. The results of the experimentation show the interest of our approach to detect TopN behavioral leaders that predict accurately the preferences of the other users. Besides, our approach can be harnessed in different application areas caring about the role of leaders.

Collaborative filtering exploits user preferences, generally ratings, to provide them with recommendations. However, the ratings may not be completely trustworthy: the rating scale is usually reduced and the rating values may be influenced by many factors. This paper is a first attempt at studying the expression of preferences under the form of preference relations where users are asked to compare pairs of resources. First experiments show that this new approach compares with, and sometimes improves, the classical one.

Recommender systems are widely used for personalization of information on the web and information retrieval systems. Collaborative Filtering (CF) is the most popular recommendation technique. However, classical CF systems use only direct links and common features to model relationships between users. This paper presents a new Collaborative Filtering approach (BNCF) based on a behavioral network that uses navigational patterns to model relationships between users and exploits social networks techniques, such as transitivity, to explore additional links throughout the behavioral network. The final aim consists in involving these new links in prediction generation, to improve recommendations quality. BNCF is evaluated in terms of accuracy on a real usage dataset. The experimentation shows the benefit of exploiting new links to compute predictions. Indeed, BNCF highly improves the accuracy of predictions, especially in terms of HMAE.

Recommender systems provide users with pertinent resources according to their context and their profiles, by applying statistical and knowledge discovery techniques. This paper describes a new approach of generating suitable recommendations based on the active user's navigation stream, by considering long and short-distance resources in the history with a tractable model.

The Internet is constantly growing, proposing more and more services and sources of information. Modeling personal preferences enables recommender systems to identify relevant subsets of items. These systems often rely on filtering techniques based on symbolic or numerical approaches in a stochastic context. In this paper, we focus on item-based collaborative filtering (CF) techniques. We propose a new approach combining a classic CF algorithm with a reinforcement model to get a better accuracy. We deal with this issue by exploiting probabilistic skewnesses in triplets of items.

Recommender systems filter resources for a given user by predicting the most pertinent resource given a specific context. This paper describes a new approach of generating suitable recommendations based on the active user's navigation stream. The underlying hypothesis is that the resources order in the stream results from the intrinsic logic of the user's behavior. The sequence based recommender we propose is inspired from language modeling and integrates skipping techniques. It has been tested on a browsing dataset extracted from Intranet logs provided by a French bank. Results show that the use of exponential decay weighting schemes when taking into account non contiguous sequences to compute recommendations enhances the accuracy. Moreover, we propose a skipping variant that provides a high accuracy while being less complex.

Technologies involved in Web information distribution services are evolving to adapt themselves to new user requirements. Usually, these new technologies are used separately. ELIN (electronic newspaper initiative) project is a European Commission funded project that tries to integrate the newest standards and technologies involved in multimedia delivery applied to web newspapers. It has as objective the delivery of any type of media format to any kind of user terminal in an efficient way. In order to do that, it takes the approach of using MPEG standards: MPEG-4 for video delivery, MPEG-7 for data classification and MPEG-21 for data management and adaptation. So it integrates the totality of solutions provided for the MPEG group.

Making heterogeneous agents cooperate is still an open problem. We have studied the interaction between a human agent and an information service agent. Our approach is to introduce a mediator agent to formalize the requests of the users, according to their profile and then to give the relevant answers. The mediator must find the best mediation strategy (a sequence of interactions) with a Markov decision process (MDP). The states are built on an attribute based referential and the capacity of the source to answer the request under formalization. The actions allow to ask questions to the user or to probe the information source. The rewards reflect the satisfaction of the user, the length of the mediation and the quantity of results. Our prototype uses reinforcement learning (Q-learning) for an on-line adaptation without requiring an a priori model. We describe our experiments on a flight information service with a simulated behaviour.

Guaranteed response time is one of the most important issues encountered in designing real time systems. The AI community has developed various approaches to solve this problem, e.g. anytime algorithms, approximate processing, design to time scheduling and progressive reasoning. All these approaches rely on a trade-off between run time and quality of results. In the framework of the ESPRIT projects Nos. 5145 and 7805 REAKT (Real time Knowledge Tool), we have developed a similar approach called PROGRESS (PROGressive REasoning System). PROGRESS manages AI tasks with hard and soft deadlines, provided that competing methods are available for the tasks to be solved. A comparison was made between our model and a conventional algorithm (Earliest Deadline Algorithm) and showed the higher robustness and efficiency of PROGRESS.

The hand-written word recognition problem is formulated in two steps. In the first step, the plausibility of observing each character is computed as a function of sample index of a line. In the second step, word recognition is achieved by finding the word (sequence of characters) which maximizes the sum of plausibilities of individual characters which make up the word. The authors propose an efficient algorithm for the second step which makes use of peaks of plausibility functions and solves the maximization process by two embedded search processes: finding the best path connecting peaks of the plausibility functions of two successive characters, and finding the best transition sample index for two given peaks. In a preliminary experimentation, using template matching for character plausibility estimation, a 96% recognition rate was obtained for a cursive-writing-like font.>

Publie dans : Proceedings EUROSPEECH 89 (European conference on speech communication and technology), Paris, September 1989

Publie dans : Proceedings EUROSPEECH 89 (European conference on speech communication and technology), Paris, September 1989