Multi-view observations potentially offer a more comprehensive understanding of real-world phenomena compared to observations acquired from a single viewpoint. Existing models that utilize multi-view data often consider that all views are available during inference, but this assumption may not hold in practical scenarios. To address this limitation, we introduce MVLD, a novel method that, by employing a deterministic autoencoder and a score-based diffusion model, is capable of imputing missing views. We finally envision MVLD being used in a communication system for image transmission.

The automatic supervision of IT systems is a current challenge at Orange. Given the size and complexity reached by its IT operations, the number of sensors needed to obtain measurements over time, used to infer normal and abnormal behaviors, has increased dramatically making traditional expert-based supervision methods slow or prone to errors. In this paper, we propose a fast and stable method called UnSupervised Anomaly Detection for multivariate time series (USAD) based on adversely trained autoencoders. Its autoencoder architecture makes it capable of learning in an unsupervised way. The use of adversarial training and its architecture allows it to isolate anomalies while providing fast training. We study the properties of our methods through experiments on five public datasets, thus demonstrating its robustness, training speed and high anomaly detection performance. Through a feasibility study using Orange's proprietary data we have been able to validate Orange's requirements on scalability, stability, robustness, training speed and high performance.

Extracting value from data stored in object stores,such as OpenStack Swift and Amazon S3, can be problematicin common scenarios where analytics frameworks and objectstores run in physically disaggregated clusters. One of the mainproblems is that analytics frameworks must ingest large amountsof data from the object store prior to the actual computation;this incurs a significant resources and performance overhead. Toovercome this problem, we present Scoop. Scoop enables analyticsframeworks to benefit from the computational resources of objectstores to optimize the execution of analytics jobs. Scoop achievesthis by enabling the addition of ETL-type actions to the dataupload path and by offloading querying functions to the objectstore through a rich and extensible active object storage layer. Asa proof-of-concept, Scoop enables Apache Spark SQL selectionsand projections to be executed close to the data in OpenStackSwift for accelerating analytics workloads of a smart energy gridcompany (GridPocket). Our experiments in a 63-machine clusterwith real IoT data and SQL queries from GridPocket show thatScoop exhibits query execution times up to 30x faster than thetraditional “ingest-then-compute” approach.

In this paper, we present an approximate algorithm that is able to quickly modify a large distributed fc-nn graph by adding or removing nodes. The algorithm produces an approximate graph that is highly similar to the graph computed using a naïve approach, although it requires the computation of far fewer similarities. To achieve this goal, it relies on a novel, distributed graph based search procedure. All these algorithms are also experimentally evaluated, using both euclidean and non-euclidean datasets.

Data summarization is essential for users to interact with data. Current state of the art algorithms to optimize its most general form, the multiple Group By queries, have limitations in scalability. In this paper, we propose a novel algorithm, Top-Down Splitting, that scales to hundreds or even thousands of attributes and queries, and that quickly and efficiently produces optimized query execution plans. We analyze the complexity of our algorithm, and evaluate, empirically, its scalability and effectiveness through an experimental campaign. Results show that our algorithm is remarkably faster than alternatives in prior works, while generally producing better solutions. Ultimately, our algorithm reduces up to 34% the query execution time, when compared to un-optimized plans.

We present NG-DBSCAN, an approximate density-based clustering algorithm that operates on arbitrary data and any symmetric distance measure. The distributed design of our algorithm makes it scalable to very large datasets; its approximate nature makes it fast, yet capable of producing high quality clustering results. We provide a detailed overview of the steps of NG-DBSCAN, together with their analysis. Our results, obtained through an extensive experimental campaign with real and synthetic data, substantiate our claims about NG-DBSCAN's performance and scalability.

Clustering items using textual features is an important problem with many applications, such as root-cause analysis of spam campaigns, as well as identifying common topics in social media. Due to the sheer size of such data, algorithmic scalability becomes a major concern. In this work, we present our approach for text clustering that builds an approximate k-NN graph, which is then used to compute connected components representing clusters. Our focus is to understand the scalability / accuracy tradeoff that underlies our method: we do so through an extensive experimental campaign, where we use real-life datasets, and show that even rough approximations of k-NN graphs are sufficient to identify valid clusters. Our method is scalable and can be easily tuned to meet requirements stemming from different application domains.

In this paper we present our new design of NNCTPH, a scalable algorithm to build an approximate k-NN graph from large text datasets. The algorithm uses a modified version of Context Triggered Piecewise Hashing to bin the input data into buckets, and uses NN-Descent, a versatile graph-building algorithm, inside each bucket. We use datasets consisting of the subject of spam emails to experimentally test the influence of the different parameters of the algorithm on the number of computed similarities, on processing time, and on the quality of the final graph. We also compare the algorithm with a sequential and a MapReduce implementation of NN-Descent. For our datasets, the algorithm proved to be up to ten times faster than NN-Descent, for the same quality of produced graph. Moreover, the speedup increased with the size of the dataset, making NNCTPH a sensible choice for very large text datasets.

In this paper, we analyze a public dataset of electricity consumption collected over 3,800 households for one year and half. We show that some socio-economic factors are critical indicators to forecast households' daily peak (and total) load. By using a random forests model, we show that the daily load can be predicted accurately at a fine temporal granularity. Differently from many state-of-the-art techniques based on support vector machines, our model allows to derive a set of heuristic rules that are highly interpretable and easy to fuse with human experts domain knowledge. Lastly, we quantify the different importance of each socio-economic feature in the prediction task.

Size-based scheduling with aging has, for long, been recognized as an effective approach to guarantee fairness and near-optimal system response times. We present HFSP, a scheduler introducing this technique to a real, multi-server, complex and widely used system such as Hadoop. Size-based scheduling requires a priori job size information, which is not available in Hadoop: HFSP builds such knowledge by estimating it on-line during job execution. Our experiments, which are based on realistic workloads generated via a standard benchmarking suite, pinpoint at a significant decrease in system response times with respect to the widely used Hadoop Fair scheduler, and show that HFSP is largely tolerant to job size estimation errors.

Social media sites have used recommender systems to suggest items users might like but are not already familiar with. These items are typically movies, books, pictures, or songs. Here we consider an alternative class of items - pictures posted by design-conscious individuals. We do so in the context of a mobile application in which users find "cool" items in the real world, take pictures of them, and share those pictures online. In this context, temporal dynamics matter, and users would greatly profit from ways of identifying the latest design trends. We propose a new way of recommending trending pictures to users, which unfolds in three steps. First, two types of users are identified - those who are good at uploading trends (trend makers) and those who are experienced in discovering trends (trend spotters). Second, based on what those "special few" have uploaded and rated, trends are identified early on. Third, trends are recommended using existing algorithms. Upon the complete longitudinal dataset of the mobile application, we compare our approach's performance to a traditional recommender system's.