Federated Learning (FL) is a decentralized paradigm that enables a client-server architecture to collaboratively train a global Artificial Intelligence model without sharing raw data, thereby preserving privacy. A key challenge in FL is Non-IID data. Quantity Skew (QS) is a particular problem of Non-IID, where clients hold highly heterogeneous data volumes. Clustered Federated Learning (CFL) is an emergent variant of FL that presents a promising solution to Non-IID problem. It improves models' performance by grouping clients with similar data distributions into clusters. CFL methods generally fall into two operating strategies. In the first strategy, clients select the cluster that minimizes the local training loss. In the second strategy, the server groups clients based on local model similarities. However, most CFL methods lack systematic evaluation under QS but present significant challenges because of it.

Clustering is an important data mining task which is widely spread in various domains such as biology, finance, marketing, healthcare, and social sciences. It allows the end user to discover, through built clusters, relationships within data. Many non-expert users perceive clustering as an "easy" task because it always produces a result. However, choosing a clustering algorithm at random, without proper parameter tuning, often leads to poor results. In particular, an important choice when applying a clustering algorithm to a specific dataset is the similarity measure. Since clustering algorithms rely on similarities between data points to build clusters, the chosen similarity measure should fit the data as accurately as possible in order to form the best clusters. Mixed Data are data that are characterized by numerical as well as categorical attributes. When clustering mixed data, the same similarity measure cannot be used for the two attribute types. Commonly a pair of similarity measures is used, one dedicated to numerical attributes and one dedicated to categorical attributes. The choice of these two most appropriate similarity measures is very important in mixed data, as it significantly affects the clustering performance.

In many domains, we face heterogeneous data with both numeric and categorical attributes. Clustering such data is challenging because the notion of similarity is not well defined due to the multiple data types. Existing clustering algorithms for these data are mainly based on two strategies: the homogenization one where all attributes are converted to a single type and the mixed one where similarity measures for the different data types are combined to define a similarity measure for heterogeneous data. We propose a framework in which we evaluate and compare several clustering algorithms using these two strategies on many real-world data sets. Then, motivated by the importance of similarity in clustering and the diversity of similarity measures for each data type, we proposed as a second study, to evaluate how their choice affects the performance of clustering algorithms using the mixed strategy. Our results suggest that the mixed strategy is preferable to the homogenization one since it uses adapted similarity measures for the different data types. Furthermore, the choice of similarity measures is very important for most of used mixed methods and an optimal choice may lead to great improvements compared to classically used similarity measures.

International audience

Abstract Herein, the problem of vehicle re‐identification using distance comparison of images in CNN latent spaces is addressed. First, the impact of the distance metrics, comparing performances obtained with different metrics is studied: the minimal Euclidean distance ( MED ), the minimal cosine distance ( MCD ) and the residue of the sparse coding reconstruction ( RSCR ). These metrics are applied using features extracted from five different CNN architectures, namely ResNet18, AlexNet, VGG16, InceptionV3 and DenseNet201. We use the specific vehicle re‐identification dataset VeRi to fine‐tune these CNNs and evaluate results. Overall, independently of the CNN used, MCD outperforms MED , commonly used in the literature. These results are confirmed on other vehicle retrieval datasets. Second, the state‐of‐the‐art image‐to‐track process (I2TP) is extended to a track‐to‐track process (T2TP). The three distance metrics are extended to measure distance between tracks, enabling T2TP. T2TP and I2TP are compared using the same CNN models. Results show that T2TP outperforms I2TP for MCD and RSCR. T2TP combining DenseNet201 and MCD ‐based metrics exhibits the best performances, outperforming the state‐of‐the‐art I2TP‐based models. Finally, experiments highlight two main results: i) the impact of metric choice in vehicle re‐identification, and ii) T2TP improves the performances compared with I2TP, especially when coupled with MCD ‐based metrics.

This paper presents our work on forensic video analysis that aimed to assist videosurveillance operators by reducing the volume of video to analyze during the search for post-evidence in videos. This work is conducted in collaboration with the French National Police and is based on requirements defined in a project related to videos analysis in the context of investigations. Due to the constant increasing volume of video generated by CCTV cameras, one of the investigators' goals is to reduce video analysis time. For this purpose, we propose a negative filtering approach based on quality and usability/utility metadata, enabling to eliminate video sequences that do not satisfy requirements for their analysis through automatic processing. Our approach involves a data model which is able to integrate different levels of video metadata, and an associated query mechanism. Experiments performed using the developed framework demonstrate the utility of our approach in a real-world case. Results show that our approach helps CCTV operators to significantly reduce video analysis times.

With the constant evolution of CCTV cameras deployed in major cities to ensure the citizens' security, CCTV operators have to watch a huge amount of video when they are searching for scenes, objects, or target persons. Watching or processing some video sequences can be useless for several reasons: content is unsuitable for operators' needs, unusable shooting conditions, etc. Filtering useless content can be an efficient way for operators to save time. In this paper we propose an approach for CCTV contents filtering based on contextual information in order to provide CCTV operators with video sequences of interest. The proposed approach takes into account many sources of contextual information such as: open data, social media, mobility, geolocation, and crowdsourcing. We provide an analysis of contextual information relevant for this approach. Since interoperability is one of the main problems of context-based approaches, we propose a generic data model of contextual information used in our approach in order to tackle this issue. Based on this data, we propose a framework architecture for relevant video segments recommendation.

In surveillance applications, humans and vehicles are the most important common elements studied. In consequence, detecting and matching a person or a car that appears on several videos is a key problem. Many algorithms have been introduced and nowadays, a major relative problem is to evaluate precisely and to compare these algorithms, in reference to a common ground-truth. In this paper, our goal is to introduce a new dataset for evaluating multi-view based methods. This dataset aims at paving the way for multidisciplinary approaches and applications such as 4D-scene reconstruction, object identification/tracking, audio event detection and multi-source meta-data modeling and querying. Consequently, we provide two sets of 25 synchronized videos with audio tracks, all depicting the same scene from multiple viewpoints, each set of videos following a detailed scenario consisting in comings and goings of people and cars. Every video was annotated by regularly drawing bounding boxes on every moving object with a flag indicating whether the object is fully visible or occluded, specifying its category (human or vehicle), providing visual details (for example clothes types or colors), and timestamps of its apparitions and disappearances. Audio events are also annotated by a category and timestamps.

The use of mobile devices and the development of geo-positioning technologies make applications that use location-based services very attractive and useful. These applications are composed of sensors that generate various and heterogeneous spatio-temporal data. Exploiting this spatio-temporal data to support video surveillance systems remains a relevant purpose for video content filtering. Since the data processed in such a context are heterogeneous (indoor and outdoor environment, various position types and reference systems, various data format), interoperability and management of these data remains a problem to be solved.

Research has shown the effectiveness of inferring user interests from social neighbors, also called "social profiling". However, the evolution in the social profile is not widely taken into consideration. To overcome this drawback, we propose a time-aware social profiling method that considers the temporal factors of the information and the relationships between the user and his/her social neighbors. This method aims at weighting user interests in the social profile, by applying a time decay function. The temporal score of a given interest is computed by combining the temporal score of information used to extract the interests with the temporal score of individuals who share the information in the network. The experiments conducted on a co-authorship network, DBLP showed that the time-aware social profiling process applying our proposed time-aware method outperforms the existing time-agnostic social profiling process. The combinatorial analysis and the parametric study led us to observe that in the context of co-authorship network, the individual temporal score has more influence than the information temporal score. As this kind of network does not exhibit a rapid evolution of information and relationships, to obtain a relevant social profile, the information should be damped slowly.

Incorporating user interests evolution over time is a crucial problem in user profiling. We particularly focus on social profiling process that uses information shared on user social network to extract his/her interests. In this work, we apply our existing time-aware social profiling method on Twitter. The aim of this study is to measure the effectiveness of our approach on this kind of social network platform, which has different characteristics from those of other social networking sites. Although the improvement compared to the time-agnostic baseline method is still low, the experiments using a parametric study showed us the benefit of applying a time-aware social profiling process on Twitter. We also found that our method performs well on sparse networks and that the information dynamic influences more the quality of our proposed time-aware method than the relationships dynamic while building the social profile on Twitter. This observation will lead us to a more complex study to find out meaningful factors to incorporate user interests evolution on social profiling process in such a network.

This paper presents our work about assisting video-surveillance agents in the search for particular video scenes of interest in transit network. This work has been developed based on requirements defined within different projects with the French National Police in a forensic goal. The video-surveillance agent inputs a query in the form of a hybrid trajectory (date, time, locations expressed with regards to different reference systems) and potentially some visual descriptions of the scene. The query processing starts with the interpretation of the hybrid trajectory and continues with a selection of a set of cameras likely to have filmed the spatial trajectory. The main contributions of this paper are: (1) a definition of the hybrid trajectory query concept, trajectory that is constituted of geometrical and symbolic segments represented with regards to different reference systems (e.g., Geodesic system, road network), (2) a spatio-temporal filtering framework based on a spatio-temporal modeling of the transit network and associated cameras.

Improving the egocentric network-based user's profile building process by taking into account the dynamic characteristics of social networks can be relevant in many applications. To achieve this aim, we propose to apply a time-aware method into an existing egocentric-based user profiling process, based on previous contributions of our team. The aim of this strategy is to weight user's interests according to their relevance and freshness. The time awareness weight of an interest is computed by combining the relevance of individuals in the user's egocentric network (computed by taking into account the freshness of their ties) with the information relevance (computed by taking into account its freshness). The experiments on scientific publications networks (DBLP/Mendeley) allow us to demonstrate the effectiveness of our proposition compared to the existing time-agnostic egocentric network-based user profiling process.

In a social context, the user is more and more an active contributor for producing social information. Then, he needs a tailored information reflecting his current needs and interests in every period of time. This aims to provide a better adaptation while accessing the information space by integrating users' interests dynamic. Indeed, users' interests may change and become “outdated” through time. So, an interest judged as relevant in a period of time may fluctuate in the next period of time. Moreover, analysing the classic user behaviour to deduce his current interests is a difficult task. In fact, his behaviour isn't always reflecting his real interests. In this paper, we propose a new approach for enriching the user profile in an evolutionary environment such as a social network. The enrichment takes into account: i) the social behaviour and more precisely the tagging behaviour (that reflects user's interests) and ii) the temporal information (that reflects the dynamic evolution of users' interests). Our approach focus on the concept of temperature that reflects the importance of a resource in each period of time. This concept is used to infer common interests of users tagging the same “important” resource. The originality of our approach relies on combining information tags, users and resources in a way that guarantees a better enrichment for the social user profile. Our approach has been tested and evaluated with the Delicious social database and shows interesting precision values.

Nowadays, social networks are more and more widely used as a solution for enriching users' profiles in systems such as recommender systems or personalized systems. For an unknown user's interest, the user's social network can be a meaningful data source for deriving that interest. However, in the literature very few techniques are designed to meet this solution. Existing techniques usually focus on people individually selected in the user's social network, and strongly depend on each author's objective. To improve these techniques, we propose to use a community based algorithm that is applied to a part of the user's social network (egocentric network) and that can be reused for any purpose (e.g. personalization, recommendation). We compute weighted user's interests from these communities by considering their semantics (interests related to communities) and their structural measures (e.g. centrality measures) in the egocentric network graph. A first experiment conducted in Facebook demonstrates the usefulness of this technique compared to individuals based techniques, and the influence of structural measures (related to communities) on the quality of derived profiles. The results also raise the problem of users' privacy in platforms such as online social networks. To enable users to better protect their privacy, these platforms should provide their users with a way to also make their friendlist private.

Existing works about user profile modeling always model the user as an independent entity. However, in social sciences, many works show the user's behavior as strongly influenced by his social ties and/or social interactions. These results were diffi

In this paper, we are interested in the context of the French government's new efforts to put into practice the concept of the "personal health record (PHR)". We propose an ontology-based solution to improve the access to PHR and to extend it with the annotations that the patient can add and that will constitute the Annotated PHR. We consider that each user can manage the annotations access policies. We also develop an ontology-based solution for granting medical agents with useful, authorized and relevant information, which adopts XACML query rewriting mechanisms. The paper illustrates how the proposed framework could assist patients during their travels, when an unexpected health disorder takes place and the suitable specialist is not available.

Nowadays, online social networks host more and more applications in order to provide their users with the possibility of finding everything they need on a single platform. The number and diversity of interactions that take place over time between users and applications within these platforms make these environments very good candidates for learning various types of information about users’ interests. We are particularly interested in the determination of users’ short-term and long-term interests which are essential for adaptative systems that take into account the evolution of user’s needs. While studies in adaptative systems focus on computing interests’ weight value and time periods to determine user’s short-term and long-term profile, we focus instead on temporal graphs’ visualization of users’ interests. From a case study on Facebook, we use dynamic graphs in order to view the influence of social ties on the user’s interests.