 Merci et n'hésitez pas à poser des questions lors de ma présentation, si vous voulez, bien sûr. Je vais présenter mes activités de recherche depuis les six ou sept ans, je ne l'ai pas remarqué exactement. Il y a des systèmes de recommandation et je vais discuter un peu sur les analyses de métal. Juste un petit coup de colère, mais je vais expliquer ce qu'il y a à ce moment. Oui, j'ai fait mes thésies sur l'Université de Tour, précisément sur une belle ville, c'est Blois, 60 000 habitants avec beaucoup de casseurs etc. Et après ça, j'ai performé une position postdoc pour l'Université de Nantes, et j'ai obtenu ma position, la position de professeur de l'Université de Toulouse, sur le Lab de l'Hirith. Plus précisément, ma PhD a été focussée sur la recommandation de la recommandation de l'Université de Toulouse où j'ai proposé un système de recommandation basé sur la compétition de l'Université de Toulouse, je vais entrer dans les détails de cette présentation. Et j'ai aussi proposé un outil de soutenir la composition de la session de l'Université de Toulouse, incluant les fonctions de l'Université de Toulouse, que je vais aussi détailler ici, dans cette présentation. Durant ma position postdoc, j'ai été focussé sur la compétition de l'Université de Toulouse, et j'ai appris sur un objectif culturel et sur les recommandations de l'Université de Toulouse. Et maintenant, j'ai travaillé sur les systèmes de recommandation de la recommandation de l'Université de Toulouse et les mesures de qualité dans les recommandations. Donc, ce sont des petits papiers que j'ai évoqués, des journalistes spécifiques, des chefs de boulot, des conférences internationales, des workshops, des conférences nationales, etc. Et j'ai aussi, c'est vraiment un plaisir pour moi, d'être un membre de l'Université de Toulouse, ce qui est un fameux workshop sur les systèmes de l'Université de Toulouse, et d'actuellement plus dans les approches de Big Data. Donc, sur les systèmes de recommandation de l'Université de Toulouse, dans le contexte de l'Université de Toulouse. Donc, supposons un utilisateur qui veut interrompre avec plusieurs applications, et qu'on puisse aussi configurer un nouveau car dans un web site. Donc, bien sûr, c'est un joke ici, c'est un renou, un car. Mais on peut imaginer un web site où on veut configurer un nouveau car sur le web site, ou analyser des cubes de données ou acheter quelque chose dans un web site, ou acheter des données au format de Big Data. Le problème est que, dans ce cas, un utilisateur, ok, après plusieurs heures, ou après plusieurs minutes, ce utilisateur peut être confusé. Pourquoi ? Parce que, pour exemple, dans le cas de configuration de car, je ne sais pas si vous avez déjà utilisé ça, mais généralement, ça peut être difficile si vous n'avez pas en mind la façon de comment le web site fonctionne pour configurer un nouveau car. C'est difficile si vous n'avez pas la possibilité d'avoir des propositions de ce web site. Ça peut être appliqué avec la même chose quand nous voulons analyser des cubes de données ou acheter quelque chose dans les web sites, ou acheter des données de Big Data. C'est pourquoi, afin d'améliorer et d'améliorer l'accès à l'information, nous proposons généralement un système scientifique qui propose la personnalisation ou la recommandation pour ce fameux utilisateur confusé. Personnalisation, généralement, pour modifier l'action que le utilisateur fait, mais juste, ok, modifier, mais, afin d'être plus près à l'expectation de cet utilisateur. La recommandation est plus pour proposer les actions de cet utilisateur. C'est important, ces différences entre personnalisation et recommandation. Et, de la littérature, nous basons le système scientifique sur ce que nous appelons « collaborative filtering approaches » où l'information de l'ancien utilisateur qui a des actions dans le passé, et qui ont été utilisées pour recommandation de nouvelles actions à l'aide d'une action curante de l'utilisateur. Donc, juste pour détenir les détails de ce contexte, parce que le système recommandateur est proposé dans cet contexte, j'ai besoin de proposer un modèle multi-démotionnel pour travailler avec cela. C'est-à-dire, pour exemple, les hierarchies dans les cubes de données, pour exemple, si nous voulons focusser sur ces données, nous avons besoin, pour exemple, 1, 2, 3, 4, 5 dimensions comme résidence, race, temps, sexe, occupation et où les niveaux, pour exemple, toute la ville, région, state et la ville sont organisés dans les hierarchies. Bien sûr, parce que nous voulons analyser un cube. Nous avons besoin de mesures pour analyser, quand nous avons analysé les dimensions, pour exemple, dans une table de course nous avons besoin de mesures, nous avons besoin de valeurs pour analyser. C'est pourquoi nous avons besoin de mesures comme income, gaz, de l'eau, si nous voulons analyser les gens j'ai aussi basé mon travail sur une définition précise qui peut être considérée comme une définition basée mais cette définition qui peut être considérée comme la main qui est possible de faire dans un cube de données donc nous considérons un set de groupes qui est lié entre les niveaux de différentes hierarchies que nous définissons ici, un certain set où nous allons définir plus les valeurs que nous voulons focuser selon le set de groupes et, bien sûr, le set de mesures. Nous allons aussi considérer une définition basée que c'est juste une séquence de clés, ok? Et mon problème lors de mon travail c'était comment réaliser ce problème particulier qui est le contexte de Holap mais aussi considérée comme une façon efficace pour la recommandation. Donc l'originalité de mon système était pour proposer des séquences de clés, pourquoi? Parce que, en général, depuis la littérature, nous proposons, le système recommandé n'est pas un clés ou généralement une partie de clés mais ce n'est pas si sympa d'avoir une vue générale de ce que les réserves ont fait dans le passé. C'est pourquoi je propose des séquences par le système recommandé plutôt que de clés. Et je base mon système sur une expression clés plutôt qu'un résultat clés Pourquoi? Parce que, juste considérant ce que l'utilise est typant est plus efficace pour considérer les résultats, les valeurs donc la compétition de la courbe dans le database que la courbe expression donc pour des consens efficaces. Mon système est proposé dans trois phases successives nous faisons la première phase où nous allons essayer d'aligner la session de courbe la session de courbe ce que l'utilise est faisant la session n'est pas finie et nous allons essayer d'aligner cette session avec toutes les sessions dans le passé, donc produite par d'autres utilisateurs dans le passé une phase de ralenti où nous allons considérer les résultats de la première phase et nous allons essayer de ralentir les sessions intéressantes et après une phase de ralenti où nous allons essayer d'adapter la session obtenue par la meilleure phase au contexte de l'utilise de courbe je veux dire, nous allons essayer de modifier la recommandation que nous avons obtenue ici pour être plus proche de la session de courbe ce n'est pas le commun dans les systèmes de recommandation pour être plus précis avec chaque phase la phase de ralenti c'est comme ça donc nous considérons la session de courbe peut-être que je vais essayer avec le pointeur nous avons ici la session de courbe en gris nous avons la courbe courbe donc l'utilise a fait cette courbe mais il essaie de obtenir une nouvelle courbe en gris en gris vous avez les sessions de courbe et l'idée est d'aligner de trouver un moyen d'aligner la fin de la session de courbe avec le début des sessions de courbe afin d'extraire à l'aliment de trouver un futur potentiel pour le recommandation ok donc cette partie sera le candidat potentiel pour la recommandation selon ce non, non, peut-être pour plus de précision ici pour trouver l'aliment nous avons utilisé un majeur de la majeure de la similarité de la majeure classique d'informations et de retrieval le premier était entre la courbe courbe nous proposons une similarité entre la courbe courbe basée sur l'expression bien sûr c'est l'équivalent si vous êtes à l'aise pour le format d'aliment donc goût sélection et joint, oui, joint et projection each one evaluated by a specific measure so if you remember a courbe is composed with a goodbye set a selection set and a measure set we will apply one measure per type of format of the query for each of the 3 elements of the query but the more interesting similarity measure is between sequences of queries we propose possible possible measure that we have we had in mind for instance the dice coefficient a distance alignment of sub sequences and we ask to real users to validate our approaches according to this different similarity measures and the more interesting similarity measure between sequences and also validated with user test while the alignment of sub sequences alignment of sub sequences implementing by the extension of the algorithm of Smith-Waterman is really known in the biological context where we want to align DNA sequences we finally just we take the same approaches here in order to have a maximal similarity where we have a perfect alignment between 2 sessions a good alignment and so a good similarity when we have for instance a possible gap between one of the 2 sessions to compare a very low similarity where we have a very bad alignment for instance here completing the contrary of the sequence and of course minimal similarity so zero when we have no alignment between 2 sessions in order to have the case where we want to align how can I say a same when you want to align yes to affect a more important weight to the alignment for the end queries of the sessions we applied a gap penalty where we will score finally each queries to obtain this type of things when we want to to measure the sequences finally and we have the representation of the gap penalty where we obtain discontent function so according to the sessions that we obtain with the alignment phase we will have to rank these sessions between them this potential future here between them to do this we will apply weights on each query that are the most representative for instance here in red the queries are really similar each query is really similar to what we can find in the logs so finally we will apply a more weight for these queries we will also apply a sort of average of score between all these scores of queries and we will just take the best session that we obtain according to this phase and so in the third phase we apply so the fitting phase where we will try to be more close to what the user is doing so for instance here we have the best session that we obtain in the previous phase and we will try just to be close to what the user is doing the idea here is really simple we will just try to modify part of fragments of the recommendation here so for instance and we will apply two types of rules each rules is not so nice because we just name these rules of type 1 so we have here the current session here the session of the log and here the part of the recommendation that we will try to fit the idea is to extract association rules I don't know if you are aware with this but the idea is to compose transaction between one query of the session and one query of the log session according of course on the alignment that we found here if you remember with the similarity measure and Smith-Wetterman algorithm and the idea here is just to find some frequencies of fragments that we find between the current session and the log session in order to have in the body of the rule a fragment of the log session and on the head of the rule a fragment of the current session why? because thanks to this type of rules we will able to modify the the fragment that we have here by the fragment that we have in the head of the rules for instance here 2005 is just replaced by 2002 we see in the current session we are clearly in the recommendation part closer to what the user is doing ok? we are here just focus on the year 2002 before we was focused on 2005 and we apply also the pretty same types of things with rules of type 2 but just here considering 1 transaction 1 query per transaction of the current session it means that we want here just to identify some regularities that are identified clearly in the current session for instance here and yes yes for instance we will so using this transaction we will extract all the types of association the rules that we can and if the rule matches with the recommendation for instance we have the context of year 2002 it matches with this context we will add the fragments that we have in the head of the rules for instance here average income and here also we clearly identify that this recommendation is closer to this current session and we have produced something that seems more natural for the user considering the fragment 2002 and average income for the user whereas before maybe not so not so evident for the user to use this type of recommendation of course it's nice to propose a recommender system it's nice to propose sequence of queries but it is also nice if we are able to propose some quality measures between the recommendation and I tried to recommend some quality measures for recommendations where really the tradeoff is between something that is from a first pass provides some novelty to the user because the user is analyzing a data cube so he has to be he wants to analyze so the recommender the recommender system has to provide some new new fragments new queries to this user so he has to have some noble things to do but he wants also to be well adapted to what the user is doing that is why I also propose a measure that can estimate the adaptation of the queries we commanded or the sequence of queries we commanded to the current user so here a nobility whose recommendation provides new elements informative from the logs and adaptation preserving consistency with the current session yes so it is well to propose a recommender system but it is well if we can really see how the recommender system can be used by users and that is why I propose a tool to assist a user for all up query and session composition this tool is composed of three types of functions holistic functions such as of course the recommender system suggesting queries to the user a system also of log summary I will enter in the details just after in order to overcome the cold start problem that is well known in recommender system just the cold start problem is when we have not past sessions where the system is begin it is impossible to recommend some things so that is why we propose a log summary approach and also a graphical user interface for the query composition so facilitating query addition so about query composition the idea is to reuse the DFM approach proposed by Matteo Golfarelli and Stefano Rizzi I don't remember the year but it is a famous schema for defining data cube and we just reuse this type of thing to construct a query the idea is to when we want to obtain a group by set just to link the levels between them when we want a selection we have just to click on a level to specify the value and we have a selection and when we want to specify a measure we have just to click here in the measure part and add the measure that you want for instance I can obtain here the query average income for each region between 2000 and 2001 for women it is this type of query is possible if we want also focused on census data about log summary and log visualization so in order to overcome the problem of call start in recombinative system the logs, the past actions are presented in a summarized view the idea to consider all the sessions of the log in this part the idea is to summarize the sessions between them in order to have consist representation of what we have in the logs just to facilitate the navigation and the view of what it exists in the log and for instance in order to reuse sessions or queries that are presented in the logs the hierarchical clustering of the log for instance here we try to represent all these four sessions using three clusters composed of for instance the first two sessions of the log, four sessions and ten sessions and because we want also to navigate between the sessions we proposed splitting operation for instance we can click on a cluster and when we click on the cluster we have the subdivision in two other clusters composed of seven sessions and three sessions so the sum is ten in order to navigate in the hierarchy and in order also to obtain a potential interesting session or query of course yes because and we know that a log can be composed with many queries or sessions several maybe hundred or thousands queries of more sessions we offer also the possibility to the user to filter the log according to one query for instance I want all the sessions on the log that includes the query sex female year 2001 and the major average income where the system will compute another view of what it exists in the logs considering only this query for instance we have just here three sessions that are represented we offer also the possibility to filter the log not just only using a similar query but a similar session so the user is doing a session and we will try to find just the sessions on the log that are similar to the session that the user is doing and of course we also propose in this system the recommender system that I where I detail the things just before so for instance we have an example of recommended session according to what the user is doing here the current session being here a common problem with recommender system is to test the recommender system and even more the tool that I propose here with real users it's really a problem it's really common it's really difficult that's why we call to master students to answer to different questionnaires different questions in order to obtain some logs that we will use to test our system if the system works and if the system is able to recommend real event recommendations but the problem is that master students may be not so expert with all app systems and that is why we just ask to the students to use the query composition tool so without using recommender system and without using log browsing why? because we decided to compare the efficiency of master students with it's a joke monkeys I mean using naive algorithm the idea being if we are able to recommend to propose recommendations that are more efficient than naive algorithm could be if the students are able to do more efficient sessions than our naive algorithms could be algorithms just use the holistic functions so using the recommender system and the log browsing answering to exactly the same question that we ask to the students and we also finally obtain some logs and we just after that compare the logs between them by applying recall and precision measures between the same set of sessions answering to the same question and of course the good thing is that our system seems work well in terms of precision so our recommender system is really precise compared to naive algorithms are able to do the problem is about recall but recall is really simple to when we want to obtain a high recall it's really simple in this case because we have just to to how can I say to execute our naive algorithms in an infinite time in order to obtain a maximal recall here so now the part 2 it's okay for the first part now maybe because you are from Beichler or master degree sorry both yes so it's really we are really in research with recommender system so it can be difficult but I can retake this example because we have a current session where the user is doing a drill down between the levels of cities so for instance you want to have in a first query a general view of all cities considering of course data considering census data and we will answer more for Q's view on the region and state he once analyzed on 2002 and he tries different types of measures and according to log session that we found with alignment phase we we found a nice log session aligned with the end of the current sessions with the beginning of the log session so these 4 queries are aligned between them and after that we extract this part that will be the recommendations and we will try to modify this recommendation according to the 2 types of rules that I explain where we will extract association rules you know what these association rules association rules or frequent patterns are well known in databases where we want to extract frequencies of associations between data that we have, that you have in a database I don't remember the paper but I think it is in the 90's during the 90's years and according to the association rules we extract we are able to modify the recommendations to be closer to what the user is doing you see for the evaluation no because the best way is just to to find real users and experts that we will use this type of system but the problem is that I finish my PhD and I have to to work with postdoc position in another domain so actually it is not I do not continue this type of thing but I will deep learning what do you mean for every commander system yes why not it could be applied of course we have for instance no renal network but the renal network has to be of course defined but yes of course it's possible so it's okay maybe it will be more 10 minutes oh ok ah ok so maybe I will try to to speak about the third part that are my current works the idea is pretty the same that in my thesis here we consider for instance to an analyst that is Dr Flo that we want she wants to analyze flowers and Dr Cactus that we want to analyze cactus a lot of cactus and different types of cactus and every guys have a lot of information on flowers and cactus but the problem is that the Dr Flo so she is an expert on flowers but she has not the way to analyze correctly the data but Dr Cactus has this and the idea is how to to use the information computer sciences that this guy has to be used by the Dr Flo the idea is is to propose a recommender system in three phases where we will try according to a current data set that a user wants to analyze and where he has expectation of results to find some similar data sets on the past so for instance realized by Dr Cactus and here we have Dr Flo and so when we identify similar data sets we will try to identify workflows done in the past that could match with the data sets according to the expectations that the user wants and according to this we will recommend the best workflows so in terms of that matches with the best past data sets and the best indicators that maximizing what the user wants I will not enter in the details of workflow definitions but here it is what I did it is what I so we have said the system recommends workflows from past analysis according to their relevance related to performance indicators we have three steps the system has ratified the past data sets which are the most similar to the current data set the related workflows are selected and executed on the current data set and the workflow maximizing the indicators are recommended to the current user so for instance the Dr Flo yes maybe no just to know that we have also of course our system is the best to compare to other approaches here but just to have a focus of what I plan for the next years maybe 3 years or maybe more is to define criteria for analysis quality it's really a problem in research is how to estimate the relevance of analysis that can be recommended for instance how to appropriate a recommendation by user it's also a real problem in research and how to indicate to the user the relevance of a recommended analysis in the simplest possible way that is why maybe proposed during my thesis a tool to facilitate the life of the expert for instance also adapt the recommendations to the context of the user so how to handle the workflow maybe by modifying it and in order to respect at the best the user wishes so we are also in the user interest that we call the recommender system as a user-centric system and also exploring past analysis processing so offering the possibility to a user to navigate between the workflows for instance but the questions are how to represent them to the user for instance by a graphical interface and also to offer the possibility to query the workflows so just by filtering them according to what the user here wants so maybe I will finish here so merci beaucoup and thank you but maybe a bit difficult yeah I understand recommender system it's a bit hard to do this in a simple way but what you have in mind is just the three things that you have the recommender propose a sequence of queries in order to have a general view of what it exists in the past so not just one query propose to the user but really a view by sequence and also so yes so it's the first thing the other is to use similarity measure when we want to recommend something according to what the user is doing and what it exists in the log and the third is to able to be closer to what the user wants and what the user is doing when he wants to fit recommendation and so by modifying some parts of recommendations to the context of the user the context is really important in recommendation the profile of the user is really important it's the recommender system has really to be user centric it's really important Is it possible to say some example later? Yes of course I will try Yes Yes Of course Yes Yes Yes Yes Yes Thank you Thank you Thank you