3.1 Introduction

The research of the Lacodam project team is organized around four research axes, the first three being based on the long standing research interests of the team members, and the last one (interpretability) being a topic that emerged more recently in the machine learning community. The team rapidly committed itself to this novel research area, as it sits perfectly at the intersection of our traditional research interests, and has strong societal expectations.

• The first research axis (Section 3.2) is dedicated to the design of novel pattern mining and machine learning methods. Pattern mining is one of the most important approaches to discover novel knowledge in data, and one of our strongest areas of expertise. Machine learning (ML), especially classification, is the basis for many decision support systems. The work on this axis will thus serve as foundations for work on the other two axes.
• The second axis (Section 3.3) tackles another aspect of knowledge discovery in data: the interaction between the user and the system in order to co-discover novel knowledge. Our team has plenty of experience collaborating with domain experts, and is therefore aware of the need to improve such interaction.
• The third axis (Section 3.4) concerns decision support. With the help of methods from the two previous axes, our goal here is to design systems that can either assist humans with making decisions, or make relevant decisions in situations where extremely fast reaction is required.
• The fourth axis (Section 3.5) concerns interpretable machine learning and can be considered as a transversal axis that has roots in all the other three axes.

Figure 1 sums up the detailed work presented in the next few pages: we show the four research axes of the team (X-axis) on the left and our main applications areas (Y-axis) below. The colored nodes in the middle represent the precise research topics of the team aligned with their corresponding axis and main application area. These research topics will be described in this section. Lines represent projects that can link several topics and application areas.

3.2 Novel Pattern Mining and Machine Learning Algorithms

Twenty years of research in pattern mining have resulted in efficient approaches to handle the algorithmic complexity of the problem. Existing algorithms are now able to efficiently extract patterns with complex structures (ex: sequences, graphs, co-variations) from large datasets. However, when dealing with large, real-world datasets, these methods still output a large number of patterns, which is impractical for human analysis. This problem is called pattern explosion. The ongoing challenge of pattern mining research is to extract fewer but more meaningful patterns. The LACODAM team is committed to solve the pattern explosion problem by pursuing the following research topics:

• the design of dedicated algorithms for mining temporal patterns
• the design of flexible pattern mining approaches
• the automatic selection of interesting data mining results

The originality of our contributions relies on the exploration of knowledge-based approaches whose principle is to incorporate dedicated domain knowledge (aka application background knowledge) deep into the mining process. While most data mining approaches are based on agnostic approaches designed to cope with pattern explosion, we propose to develop data mining techniques that rely on knowledge-based artificial intelligence techniques. This entails the use of structured knowledge representations, as well as reasoning methods, in combination with mining.

The first topic concerns classical pattern mining in conjunction with expert knowledge in order to define new pattern types (and related algorithms) that can solve applicative issues. In particular, we investigate how to handle temporality in pattern representations which turns out to be important in many real world applications (in particular for decision support) and deserves particular attention.

The next two topics aim at proposing alternative pattern mining methods to let the user incorporate, on their own, knowledge that will help define their domain of interest. Flexible pattern mining approaches enable analysts to easily incorporate extra knowledge, for example domain related constraints, in order to extract only the most relevant patterns. On the other hand, the selection of interesting data mining results aims at devising strategies to filter out the results that are useless to the data analyst. Besides the challenge of algorithmic efficiency, we are interested in formalizing the foundations of interestingness, according to background knowledge modeled with either logical knowledge representation paradigms or with statistical approaches.

The team is also active in machine learning research. Several of our decision support applications rely on supervised classification algorithms. While supervised classification is a well researched area, there are several settings where the state of the art still requires improvements:

• when the classes are extremely unbalanced, it is difficult to learn a correct classifier. We are interested in techniques such as cost-sensitive learning to handle such situations.
• a lot of our applications deal with time series data: we are thus especially involved on proposing novel time series classifiers.

3.3 User/System Interaction

As we pointed out before, there is a strong need to present relevant patterns to the user. This can be done by using more specific constraints, background knowledge and/or tailor-made optimization functions. Due to the difficulty of determining these elements beforehand, one of the most promising solutions is that the system and the user co-construct the definition of relevance, i.e., to have a human in the loop. This requires to have means to present intermediate results to the user, and to get user feedback in order to guide the search space exploration process in the right direction. This is an important research axis for LACODAM, which will be tackled in several complementary ways as detailed next.

• Domain Specific Languages: One way to interact with the user is to propose a domain specific language (DSL) tailored to the domain at hand and to the analysis tasks. The challenge is to propose a DSL allowing the users to easily express the required processing workflows, to deploy those workflows for mining on large volumes of data and to offer as much automation as possible.
• What if / What for scenarios: We also investigate the use of scenarios to query results from data mining processes, as well as other complex processes such as complex system simulations or model predictions. Such scenarios are answers to questions of the type “what if [situation]?” or “what [should be done] for [expected outcome]?”.
• User preferences: In exploratory analysis, users often do not have a precise idea of what they want, and are not able to formulate such queries. Hence, in LACODAM we investigate simple ways for users to express their interests and preferences, either during the mining process – to guide the search space exploration –, or afterwards during the filtering and interpretation of the most relevant results.
• Data visualization: All our research axes deal with scenarios where a system displays information or some sort of distilled knowledge (e.g., patterns, predictions, explanations) to users. For example, the output of most pattern mining algorithms is a (usually long) list of patterns. Interpretable machine learning and applications in decision support may require to visualize the output of an algorithm in a comprehensible way. Fueled by our work in interpretable and explainable machine learning, data visualization techniques (referred as DataVis) have become a transversal research topic in LACODAM. We are particularly interested in the effective visualization of patterns and explanations for machine learning algorithms. Since DataVis is beyond the expertise of LACODAM, our strategy is to establish collaborations with prominent data visualization teams for this line of research, with a long term goal to recruit a specialist in data visualization if the opportunity arises.

3.4 Decision support

The increasing volumes of data coming from a range of different systems (ex: sensor data from agro-environmental systems, log data from software systems and ML models, biological data coming from health monitoring systems) can help human and software agents make better decisions. Hence, LACODAM builds upon the idea that decision support systems should take advantage of the available data.

Our works in decision support are thus rooted in applications and conducted in partnership with applicative partners: mainly Inrae researchers for agro-environmental systems, but also companies in various domains (such as cyber-security, retail, transportation) for other applications. Our aim is to help users reach a better decisions by exploiting data mining and/or machine learning techniques over the available data. In some ways, the first axis of the team can be seen as the “engine” leading to the extraction of information from data, while this axis is interested in the complete chain from raw data to the knowledge provided to the user to help them make a better decision.

3.5 Interpretability

The pervasiveness of complex decision support systems, as well as the general consensus about the societal importance of understanding their logic1, has given momentum to the field of interpretable and explainable ML. Being a team specialized in data science, we are fully aware that many problems can be solved by means of complex and accurate ML models. Alas, this accuracy sometimes comes at the expense of interpretability, which can be a major requirement in some contexts (e.g., regression using expertise/rule mining). For this reason, one of the interests of LACODAM is the study of the interpretability-accuracy trade-off. Our studies may be able to answer questions such as “how much accuracy can a model lose (or perhaps gain) by becoming more interpretable?”. Such a goal requires us to define interpretability in a more principled way —an endeavour that has been very recently addressed, still not solved. LACODAM is interested in the two main currents of research in interpretability, namely the development of natively interpretable methods, as well as the construction of interpretable mediators between users and black-box models, known as post-hoc interpretability.

We highlight the link between interpretability and LACODAM's axes of decision support, and user/system interaction. In particular, interpretability is a prerequisite for proper user/system interaction and is a central incentive for the advent of data visualization techniques for ML models. This convergence has motivated our interest in user-oriented post-hoc interpretability, a sub-field of interpretable ML that adds the user into the formula when generating proper explanations of black-box ML algorithms. This rationale is supported by existing work  37 that suggests that interpretability possesses a subjective component known as plausibility. Moreover, our user-oriented vision meets with the notion of semantic interpretability, where an explanation may resort to high level semantic elements (objects in image classification, or verbal phases in natural language processing) instead of surrogate still-machine-friendly features (such as super-pixels). LACODAM tackles these unaddressed aspects of interpretable ML with other Inria teams through the IPL HyAIAI.

3.6 Long-term goals

The following perspectives are at the convergence of the four aforementioned research axes and can be seen as ideal towards our goals:

• A crisp definition for interpretability. The formal definition of what is “interpretable” is still work in progress in the international research community. We would therefore like to find a more principled way to define interpretability. This would allow us to answer questions such as “is an explanation based on decision trees more suitable to a given user than an explanation based on linear attribution?”. This poses multiple challenges such as characterizing the user's background and knowledge, defining the right metrics to quantify the interpretability of a model or explanation, and understanding the cognitive aspects of the existing methods and its visual representations.
• Logic argumentation based on epistemic interest. Having increasingly automated approaches will require better and better ways to handle the interactions with the user. A term goal is to explore the use of logic argumentation, i.e., the formalisation of human strategies for reasoning and arguing, in the interaction between users and data analysis tools. Alongside visualization and interactive data mining tools, logic argumentation can be a way for users to query both the results and the way they are obtained. Such querying can also help the expert to reformulate her query in an interactive analysis setting.
• User-guided MDL model selection. To mitigate the pattern explosion, the research in pattern mining has turnd the attention to methods borrowed from information theory, in particular the Minimum Description Length (MDL) principle. The idea is to select a small number of patterns that together allow to compress best the data. The team now has a solid expertise and reputation in this area. A problem of MDL approaches in pattern mining is that they are challenging to develop. Especially, they require to properly design an encoding for the patterns and the data, an activity that is mostly designed in an ad-hoc way by researchers, and which has some pitfalls. A long term objective would be to formalize general principles of a “good” MDL encoding for pattern mining, in order to be able to automatically produce encodings for different pattern mining problems. It could allow, for example, to incorporate user preferences in the encoding, in order to take them into account in the selection made by MDL, without breaking the good properties of the approach.

4.1 Industry

We present below our industrial collaborations. Some are well established partnerships, while others are more recent collaborations with local industries that wish to reinforce their Data Science R&D with us (e.g. Energiency, Amossys).

• Car Sharing Data Analysis. Peugeot-Citroën (PSA) group’s know-how encompasses all areas of the automotive industry, from production to distribution and services. Among others, its aim is to provide a car sharing service in many large cities. This service consists in providing a fleet of cars and a “free floating” system that allows users to use a vehicle, then drop it off at their convenience in the city. To optimize their fleet and the availability of the cars throughout the city, PSA needs to analyze the trajectory of the cars and understand the mobility needs and behavior of their users. We tackle this subject together through the CIFRE PhD of Gregory Martin.
• Multimodal Data Analysis for the Supervision of Sensitive Sites. ATERMES is an international mid-sized company with a strong expertise in high technology and system integration from the upstream design to the long-life maintenance cycle. It has recently developed a new product, called BARIER TM (“Beacon Autonomous Reconnaissance Identification and Evaluation Response”), which provides operational and tactical solutions for mastering borders and areas. Once in place, the system allows for a continuous night and day surveillance mission with a small crew in the most unexpected rugged terrain. The CIFRE PhD of Heng Zhang aims at developing a deep learning architecture and algorithms to detect anomalies (mainly people) from multimodal data. The data are “multimodal” because information about the same phenomenon can be acquired from different types of detectors, at different conditions, in multiple experiments.
• Root Cause Analysis in Networks. AdvisorSLA is a French company specialized in software solutions for network monitoring. For this purpose, the company relies on techniques of network metrology. By continuously measuring the state of the network, monitoring solutions detect events (e.g., overloaded router) that may degrade the network's operation and the quality of the services running on top of it (e.g., video transmission could become choppy). When a monitoring solution detects a potentially problematic sequence of events, it triggers an alarm so that the network manager can take actions. Those actions can be preventive or corrective. Some statistics show that only 40% of the triggered alarms are conclusive, that is, they manage to signal a well-understood problem that requires an action from the network manager. This means that the remaining 60% are presumably false alarms. While false alarms do not hinder network operation, they do incur an important cost in terms of human resources. Thus, the CIFRE PhD of Erwan Bourrand proposes to characterize conclusive and false alarms. This will be achieved by designing automatic methods to “learn” the conditions that most likely precede each type of alarm, and therefore predict whether the alarm will be conclusive or not. This can help adjust existing monitoring solutions in order to improve their accuracy. Besides, it can help network managers automatically trace the causes of a problem in the network.

4.2 Health

• Care Pathways Analysis for Supporting Pharmaco-Epidemiological Studies. Pharmaco-epidemiology applies the methodologies developed in general epidemiology to answer to questions about the uses and effects of health products, drugs 41, 40 or medical devices 33, on population. In classical pharmaco-epidemiology studies, people who share common characteristics are recruited to build a dedicated prospective cohort. Then, meaningful data (drug exposures, diseases, etc.) are collected from the cohort within a defined period of time. Finally, a statistical analysis highlights the links (or the lack of links) between drug exposures and outcomes (e.g., adverse effects). The main drawback of prospective cohort studies is the time required to collect the data and to integrate them. Indeed, in some cases of health product safety, health authorities have to answer quickly to pharmaco-epidemiology questions.

  New approaches of pharmaco-epidemiology consist in using large EHR (Electronic Health Records) databases to investigate the effects and uses (or misuses) of drugs in real conditions. The objective is to benefit from nationwide available data to answer accurately and in a short time pharmaco-epidemiological queries for national public health institutions. Despite the potential availability of the data, their size and complexity make their analysis long and tremendous. The challenge we tackle is the conception of a generic digital toolbox to support the efficient design of a broad range of pharmaco-epidemiology studies from EHR databases. We propose to use pattern mining algorithms and reasoning techniques to analyse the typical care pathways of specific groups of patients.

  To answer the broad range of pharmaco-epidemiological queries from national public health institutions, the PEPS 2 platform exploits, in secondary use, the French health cross-schemes insurance system, called SNDS. The SNDS covers most of the French population with a sliding period of 3 past years. The main characteristics of this data warehouse are described in 39. Contrary to local hospital EHR or even to other national initiatives, the SNDS data warehouse covers a huge population. It makes possible studies on unfrequent drugs or diseases in real conditions of use. To tackle the volume and the diversity of the SNDS data warehouse, a research program has been established to design an innovative toolbox. This research program is focused first on the modeling of care pathways from the SNDS database and, second, on the design of tools supporting the extraction of insights about massive and complex care pathways by clinicians. In such a database a care pathway is an individual sequence of drugs exposures, medical procedures and hospitalizations.

• Care Sequences for the Exploration of Medico-administrative Data. The difficulty of analyzing medico-administrative data is the semantic gap between the raw data (for example, database record about the delivery at date t of drug with ATC 2 code N 02BE01) and the nature of the events sought by clinicians (“was the patient exposed to a daily dose of paracetamol higher than 3g?”). The solution that is used by epidemiologists consists in enriching the data with new types of events that, on the one side, could be generated from raw data and on the other side, have a medical interpretation. Such new abstract events are defined by clinician using proxies. For example, drugs deliveries can be translated in periods of drug exposure (drug exposure is a time-dependent variable for non-random reasons) or identify patient stages of illness, etc. A proxy can be seen as an abstract description of a care sequence.

  Currently, the clinicians are limited in the expression of these proxies both by the coarse expressivity of their tools and by the need to process efficiently large amount of data. From a semantic point of view, care sequences must fully integrate the temporal and taxonomic dimensions of the data to provide significant expression power. From a computational point of view, the methods employed must make it possible to efficiently handle large amounts of data (several millions care pathways). The aim of the PhD of Johanne Bakalara is to study temporal models of sequences in order 1) to show their abilities to specify complex proxies representing care sequences needed in pharmaco-epidemiological studies and 2) to build an efficient querying tool able to exploit large amount of care pathways.

4.3 Agriculture and Environment

• Dairy Farming. The use and analysis of data acquired in dairy farming is a challenge both for data science and animal science. The goal is to improve farming conditions, i.e., health, welfare and environment, as well as farmers’ income. Nowadays, animals are monitored by multiple sensors giving a wealth of heterogeneous data such as temperature, weight, or milk composition. Current techniques used by animal scientists focus mostly on mono-sensor approaches. The dynamic combination of several sensors could provide new services and information useful for dairy farming. The PhD thesis of Kevin Fauvel (#DigitAg grant), aims to study such combinations of sensors and to investigate the use data mining methods, especially pattern mining algorithms. The challenge is to design new algorithms that take into account data heterogeneity —in terms of nature and time units—, and that produce useful patterns for dairy farming. The outcome of this thesis will be an original and important contribution to the new challenge of the IoT (Internet of Things) and will interest domain actors to find new added value to a global data analysis. The PhD thesis, started on October 2017, takes place in an interdisciplinary setting bringing together computer scientists from INRIA and animal scientists from INRA, both located in Rennes.
• Optimizing the Nutrition of Individual Sow. Another direction for further research is the combination of data flows with prediction models in order to learn nutrition strategies. Raphaël Gauthier started a PhD thesis (#DigitAg Grant) in November 2017 with both INRIA and INRA supervisors. His research addresses the problem of finding the optimal diet to be supplied to individual sows. Given all the information available, e.g., time-series information about previous feeding, environmental data, scientists models, the research goal is to design new algorithms to determine the optimal ration for a given sow in a given day. Efficiency issues of developed algorithms will be considered since the proposed software should work in real-time on the automated feeder. The decision support process should involve the stakeholder to ensure a good level of acceptance, confidence and understanding of the final tool.
• Ecosystem Modeling and Management. Ongoing research on ecosystem management includes modelling of ecosystems and anthroprogenic pressures, with a special concern on the representation of socio-economical factors that impact human decisions. A main research issue is how to represent these factors and how to integrate their impact on the ecosystem simulation model. This work is an ongoing cooperation with ecologists from the Marine Spatial Ecology of Queensland University, Australia and from Agrocampus Ouest.

4.4 Education

• Data-oriented Academic Counseling. Course selection and recommendation are important aspects of any academic counseling system. The Learning Analytics community has long supported these activities via automatic, data-based tools for recommendation and prediction. LACODAM, in collaboration with the Ecuadorian research center CTI 3 has contributed to this body of research with the design of a tool that allows students to select multiple courses and predict their academic performance based on historical academic data. The tool resorts to visualization and interpretable machine learning techniques, and is intended to be used by the students before the counseling sessions to plan their upcoming semester at the Ecuadorian university ESPOL. In our ongoing collaboration with CTI we have now studied the impact of academic predictions, explanations in the behavior and decision of the students. Moreover, we have studied whether those effects change when the predictions are conveyed using a different visual representation. The results of this study have been published in a paper that will appear in the proceeding of the CHI conference in 2021.

4.5 Others

• RDF Archiving. The dynamicity of the Semantic Web has motivated the development of solutions for RDF archiving, i.e., the task of storing and querying all previous versions of an RDF dataset. Notwithstanding the value of RDF archiving for data maintainers and consumers, this field of research remains under-developed for multiple reasons. These include notably (i) the lack of usability and scalability of the existing systems, (ii) no archiving support for multi-graph RDF datasets, (iii) the absence of a standard SPARQL extension for RDF archives, and (iv) a disregard of the evolution patterns of RDF datasets. The PhD thesis of Olivier Pelgrin aims at developing techniques towards a scalable and full-fledged archiving solution for the Semantic Web. This PhD thesis is a collaboration between LACODAM and the DAISY team at Aalborg University.

5.1 Footprint of research activities

There are two main axes that characterize the bulk of LACODAM's environmental impact: work trips, and computing resources utilisation.

5.2 Impact of research results

We estimate that the research work can have actual impact in three different ways:

• In the short/medium term, a significant part of our research work is conduced in collaboration with companies, through CIFRE PhDs. Hence, the research problems addressed concern an important challenge for the company, and the solutions proposed are evaluated on their relevance to tackle this challenge. For example, the CIFRE PhD of Colin Leverger (defended this year), funded by Orange, was concerned with forecasting seasonal time series evolution, with application to the server load of the company. Its results could help better provision server resources during the year. Colin was hired by Orange after his PhD.
• In the medium/long term, we also have potential impactful research work with scientists from other domains, especially in environment and agriculture. Some earlier work of the team, conduced with INRAE SAS team, helped better understand nitrate pollution in Brittany, and important envirnomental issue. And recent work from the PhD of Kevin Fauvel, conduced with the INRAE Pegase team, allowed for the design of the current best method to detect oestrus in dairy cows from cheap sensors. This is important as reproduction of dairy cows is directly related to milk production, and also to cow culling in case of repetitive failures to inseminate at the right time.
• Last, in the longer term, the team has a fundamental line of work on machine learning and interpretability. This is a critical topic nowadays due to the emergence of the GDPR. Given the increasing use of machine learning solutions in most areas of human activity, work on interpretability is of utmost societal importance, as it will help in designing more useful and also more acceptable machine learning approaches. This will require a sustained effort from the community: LACODAM is taking part in this effort, both on its own, as the coordinator of the Inria HyAIAI project, and last by having several of its members in the large European Project TAILOR dedicated to this topic.

6.1 Awards

The paper of Kevin Fauvel A Distributed Multi-Sensor Machine Learning Approach to Earthquake Early Warning got an “Oustanding paper award” in the special session “AI for social good” of the prestigious AAAI 2020 conference. We are especially happy of this work as it is a collaboration with the KerData team and the Rutgers University (US).

7.1 New software

8.1 Novel Pattern Mining and Machine Learning Algorithms

We frame our research achievements primarily with respect to fundamental research problems and questions they tackle and then make the link with the practical applications they do or may serve.

8.2 User/System interaction

We report three major contributions to the axis of user-system interaction:

• An approach based on model checking that can detect faulty patterns in systems modeled by Petri nets. The proposed approach 4 can tell users if a scenario containing an event sequence of interest can lead to a fault in a systes. The approach has been evaluated and validated in the context of product transportation systems.
• A declarative approach for mining negative sequential patterns 5. Continuing our research line on negative sequential patterns, we propose to encode their different semantics 11 using a formal language, namely Answer Set Programming 38. These encodings enable to compare the outputs and computing performances of the different semantics for negative sequential patterns.
• A data-oriented counseling system for term selection that allows students to select multiple courses and predict their academic performance based on historical academic data  32. The tool is intended to be used prior to counseling sessions in which students plan their upcoming semester at the Ecuadorian university ESPOL. We have conducted a series of user studies that shed light on the impact of the system's advise on the behavior and decisions of the students.

8.3 Decision Support

Differently from the previous section we describe the axis of decision support in terms of the application domains of our research contributions.

8.4 Interpretability

LACODAM's achievements in this domain include the proposal of an explainable by design classifier for time series, a study of techniques to improve the fidelity of a post-hoc explainable method, and an approach to benchmark machine learning methods taking into account the dimension of interpretability.

9.1 Bilateral contracts with industry

• ENEDIS - Univ Rennes 1

  Participants: E. Fromont

  Contract amount: 9k€

  This is a short pre-PhD contract with Enedis, that allowed to prepare the CIFRE PhD of Anthonin Voyez.

• Hyptser: Hybrid Prediction of Time Series

  Participants: T. Guyet, S. Malinowski (LinkMedia), V. Lemaire (Orange)

  Contract amount: 25k€

  Hyptser is a collaborative project between Orange Labs and LACODAM funded by the Fondation Mathématique Jacques Hadamard (PGMO program). It aims at developping new hybrid time series prediction methods in order to improve capacity planning for server farms. Capacity planning is the process of determining the infrastructure needed to meet future customer demands for online services. A well-made capacity planning helps to reduce operational costs, and improves the quality of the provided services. Capacity planning requires accurate forecasts of the differences between the customer demands and the infrastructure theoretical capabilities. The Hyptser project makes the assumption that this information is captured by key performance indicators (KPI), that are measured continuously in the service infrastructure. Thus, we expect to improve capacity planning capabilities by making accurate forecasts of KPI time series. Recent methods about time series forecasting make use of ensemble models. In this project, we are interested in developing hybrid models for time series forecasting. In the next steps of this project, we will analyze the performance of this two strategies on KPI time series provided by Orange and compare them to classical ensemble methods.

  Additional remarks. This project has financed the PhD of Colin Leverger who defended his thesis in November 2020.

• AdvisorSLA 2018 - Inria

  Participants: E. Bourrand, L. Galárraga, E. Fromont, A. Termier

  Contract amount: 25k€

  Context. AdvisorSLA is a French company headquartered in Cesson-Sévigné, a city located in the outskirts of Rennes in Brittany. The company is specialized in software solutions for network monitoring. For this purpose, the company relies on techniques of network metrology. AdvisorSLA's customers are carriers and telecommunications/data service providers that require to monitor the performance of their communication infrastructure as well as their QoE (quality of service). Network monitoring is of tremendous value for service providers because it is their primary tool for proper network maintenance. By continuously measuring the state of the network, monitoring solutions detect events (e.g., an overloaded router) that may degrade the network's operation and the quality of the services running on top of it (e.g., video transmission could become choppy). When a monitoring solution detects a potentially problematic sequence of events, it triggers an alarm so that the network manager can take actions. Those actions can be preventive or corrective. Some statistics gathered by the company show that only 40% of the triggered alarms are conclusive, that is, they manage to signal a well-understood problem that requires an action from the network manager. This means that the remaining 60% are presumably false alarms. While false alarms do not hinder network operation, they do incur an important cost in terms of human resources.

  Objective. We propose to characterize conclusive and false alarms. This will be achieved by designing automatic methods to “learn” the conditions that most likely precede the fire of each type of alarm, and therefore predict whether the alarm will be conclusive or not. This can help adjust existing monitoring solutions in order to improve their accuracy. Besides, it can help network managers automatically trace the causes of a problem in the network. The aforementioned problem has an inherent temporal nature: we need to learn which events occur before an alarm and in which order. Moreover, metrology models take into account the measurements of different components and variables of the network such as latency and packet loss. For these two reasons, we resort to the field of multivariate time sequences and time series. The fact that we know the “symptoms” of an alarm and whether it is conclusive or not, allows for the application of supervised machine learning and pattern mining methods.

  Additional remarks. This contract financed the PhD thesis of Erwan Bourrand that was interrupted in September 2020.

• ATERMES 2018-2021 - Univ Rennes 1

  Participants: H. Zhang, E. Fromont

  Contract amount: 45k€

  Context. ATERMES is an international mid-sized company, based in Montigny-le-Bretonneux with a strong expertise in high technology and system integration from the upstream design to the long-life maintenance cycle. It has recently developed a new product, called BARIERTM (“Beacon Autonomous Reconnaissance Identification and Evaluation Response”), which provides operational and tactical solutions for mastering borders and areas. Once in place, the system allows for a continuous night and day surveillance mission with a small crew in the most unexpected rugged terrain. BARIER™ is expected to find ready application for temporary strategic site protection or ill-defined border regions in mountainous or remote terrain where fixed surveillance modes are impracticable or overly expensive to deploy.

  Objective. The project aims at providing a deep learning architecture and algorithms able to detect anomalies (mainly the presence of people or animals) from multimodal data. The data are considered “multimodal” because information about the same phenomenon can be acquired from different types of detectors, at different conditions, in multiple experiments, etc. Among possible sources of data available, ATERMES provides Doppler Radar, active-pixel sensor data (CMOS), different kind of infra-red data, the border context etc. The problem can be either supervised (if label of objects to detect are provided) or unsupervised (if only times series coming from the different sensors are available). Both the multimodal aspect and the anomaly detection one are difficult but interesting topics for which there exist few available works (that take both into account) in deep learning.

• PSA - Inria

  Participants: E. Fromont, A. Termier, L. Rozé, G. Martin

  Contract amount: 75k€

  Context. Peugeot-Citroën (PSA) group aims at improving the management of its car sharing service. To optimize its fleet and the availability of the cars throughout the city, PSA needs to analyze the trajectory of its cars.

  Objective. The aim of the internship is (1) to survey the existing methods to tackle the aforementioned need faced by PSA and (2) to also investigate how the techniques developed in LACODAM (e.g., emerging pattern mining) could be serve this purpose. A framework, consisting of three main modules, has been developped. We describe the modules in the following.

  • A town modelisation module with clustering. Similar towns are clustered in order to reuse information from one town in other towns.
  • A travel prediction module with basic statistics.
  • A reallocation strategy module (choices on how to relocate cars so that the most requested areas are always served). The aim of this module is to be able to test different strategies.

  Additional remarks. This is the doctoral contract for the PhD of Gregory Martin (Thèse CIFRE).

10.1 International initiatives

10.2 European initiatives

10.3 National initiatives

• HyAIAI: Hybrid Approaches for Interpretable AI

  Participants: E. Fromont (leader), A. Termier, L. Galárraga

  The Inria Project Lab HyAIAI is a consortium of Inria teams (Sequel, Magnet, Tau, Orpailleur, Multispeech, and LACODAM) that work together towards the development of novel methods for machine learning, that combine numerical and symbolic approaches. The goal is to develop new machine learning algorithms such that (i) they are as efficient as current best approaches, (ii) they can be guided by means of human-understandable constraints, and (iii) their decisions can be better understood.

• #DigitAg: Digital Agriculture

  Participants: A. Termier, V. Masson, C. Largouët, A.I. Graux

  #DigitAg is a “Convergence Institute” dedicated to the increasing importance of digital techniques in agriculture. Its goal is twofold: First, making innovative research on the use of digital techniques in agriculture in order to improve competitiveness, preserving the environment, and offer correct living conditions to farmers. Second, preparing future farmers and agricultural policy makers to successfully exploit such technologies. While #DigitAg is based on Montpellier, Rennes is a satellite of the institute focused on cattle farming.

  LACODAM is involved in the “data mining” challenge of the institute, which A. Termier co-leads. He is also the representative of Inria in the steering comittee of the institute. The interest for the team is to design novel methods to analyze and represent agricultural data, which are challenging because they are both heterogeneous and multi-scale (both spatial and temporal).

• DRIAS: Accountability of Artificial Intelligence in Health

  Participants: T. Guyet, T. Allard (DRUID)

  DRIAS is an inter-disciplinary project funded by MITI-CNRS. It gathers researches in computer science, public law, and health informatics to investigate the current limits of the legal frameworks (mainly French and European laws) facing the changes induced by the use of artificial intelligence in healthcare systems. This project is mainly focused on the notion of accountability, that is a motivating concept for addressing issues related to machine learning interpretability.

11.1 Promoting scientific activities

• Board member of the French Association of Artificial Intelligence (AFIA) (T. Guyet)

11.2 Teaching - Supervision - Juries

11.3 Popularization

12.1 Major publications

• 1 inproceedingsKévinK. Fauvel, VéroniqueV. Masson, ElisaE. Fromont, PhilippeP. Faverdin and AlexandreA. Termier. Towards Sustainable Dairy Management - A Machine Learning Enhanced Method for Estrus DetectionKDD 2019 - ACM SIGKDD International Conference on Knowledge Discovery & Data Mining25th SIGKDD Conference on Knowledge Discovery and Data Mining proceedingsAnchorage, United StatesAugust 2019, 1-9
  HALDOI
• 2 inproceedings EstherE. Galbrun, PeggyP. Cellier, NikolajN. Tatti, AlexandreA. Termier and BrunoB. Crémilleux. Mining Periodic Patterns with a MDL Criterion European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD) Dublin, Ireland 2018
  HAL

12.2 Publications of the year

12.3 Cited publications

• 32 inproceedings JaimeJ. Castells, Poul-DoustP.-D. Mohammad, LuisL. Galárraga, GonzaloG. Méndez, MargaritaM. Ortiz-Rojas and AlbertoA. Jiménez. A Student-oriented Tool to Support Course Selection in Academic Counseling Sessions Proceedings of the Workshop on Adoption, Adaptation and Pilots of Learning Analytics in Under-represented Regions co-located with the 15th European Conference on Technology Enhanced Learning 2020 Virtual Event, Germany September 2020
  HAL back to text
• 33 articleSS. Colas, CC. Collin, PP. Piriou and MM. Zureik. Association between total hip replacement characteristics and 3-year prosthetic survivorship: A population-based studyJAMA Surgery150102015, 979--988
  back to text
• 34 incollectionMatthiasM. Feurer, AaronA. Klein, KatharinaK. Eggensperger, JostJ. Springenberg, ManuelM. Blum and FrankF. Hutter. Efficient and Robust Automated Machine LearningAdvances in Neural Information Processing Systems 28Curran Associates, Inc.2015, 2962--2970URL: http://papers.nips.cc/paper/5872-efficient-and-robust-automated-machine-learning.pdf
  back to text
• 35 inproceedings LuisL. Galárraga, ChristinaC. Teflioudi, KatjaK. Hose and FabianF. Suchanek. AMIE: Association Rule Mining Under Incomplete Evidence in Ontological Knowledge Bases WWW 2013
  back to text
• 36 article LuisL. Galárraga, ChristinaC. Teflioudi, KatjaK. Hose and Fabian M.F. Suchanek. Fast Rule Mining in Ontological Knowledge Bases with AMIE+ VLDB Journal 24 6 2015
  back to text
• 37 articleToddT. Kulesza, SimoneS. Stumpf, Margaret M.M. Burnett, SherryS. Yang, IrwinI. Kwan and Weng-KeenW.-K. Wong. Too much, too little, or just right? Ways explanations impact end users' mental models2013 IEEE Symposium on Visual Languages and Human Centric Computing2013, 3-10
  back to text
• 38 inproceedingsVladimirV. Lifschitz. What Is Answer Set Programming?Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence, AAAI 2008, Chicago, Illinois, USA, July 13-17, 20082008, 1594--1597
  back to text
• 39 articleG. Moulis, M. Lapeyre-Mestre, A. Palmaro, G. Pugnet, J.-L. Montastruc and L. Sailler. French health insurance databases: What interest for medical research?La Revue de Médecine Interne3662015, 411 - 417
  back to text
• 40 articleEmmanuelE. Nowak, AndréA. Happe, JacquesJ. Bouget, FrancoisF. Paillard, CécileC. Vigneau, Pierre-YvesP.-Y. Scarabin and EmmanuelE. Oger. Safety of Fixed Dose of Antihypertensive Drug Combinations Compared to (Single Pill) Free-Combinations: A Nested Matched Case--Control AnalysisMedicine94492015, e2229
  back to text
• 41 articleElisabethE. Polard, EmmanuelE. Nowak, AndréA. Happe, ArnaudA. Biraben and EmmanuelE. Oger. Brand name to generic substitution of antiepileptic drugs does not lead to seizure-related hospitalization: a population-based case-crossover studyPharmacoepidemiology and drug safety24112015, 1161--1169
  back to text