Les systèmes de recommandation actuels ont besoin de recommander des objets pertinents aux utilisateurs (exploitation), mais pour cela ils doivent pouvoir également obtenir continuellement de nouvelles informations sur les objets et les utilisateurs encore peu connus (exploration). Il s'agit du dilemme exploration/exploitation. Un tel environnement s'inscrit dans le cadre de ce que l'on appelle " apprentissage par renforcement ". Dans la littérature statistique, les stratégies de bandit sont connues pour offrir des solutions à ce dilemme. Les contributions de cette thèse multidisciplinaire adaptent ces stratégies pour appréhender certaines problématiques des systèmes de recommandation, telles que la recommandation de plusieurs objets simultanément, la prise en compte du vieillissement de la popularité d'un objet ou encore la recommandation en temps réel. / Current recommender systems need to recommend items that are relevant to users (exploitation), but they must also be able to continuously obtain new information about items and users (exploration). This is the exploration / exploitation dilemma. Such an environment is part of what is called "reinforcement learning". In the statistical literature, bandit strategies are known to provide solutions to this dilemma. The contributions of this multidisciplinary thesis the adaptation of these strategies to deal with some problems of the recommendation systems, such as the recommendation of several items simultaneously, taking into account the aging of the popularity of an items or the recommendation in real time.

No description available.

Intruders attempt to penetrate commercial systems daily and cause considerable financial losses for individuals and organizations. Intrusion detection systems monitor network events to detect computer security threats. An extensive amount of network data is devoted to detecting malicious activities. Storing, processing, and analyzing the massive volume of data is costly and indicate the need to find efficient methods to perform network data reduction that does not require the data to be first captured and stored. A better approach allows the extraction of useful variables from data streams in real time and in a single pass. The removal of irrelevant attributes reduces the data to be fed to the intrusion detection system (IDS) and shortens the analysis time while improving the classification accuracy. This dissertation introduces an online, real time, data processing method for knowledge extraction. This incremental feature extraction is based on two approaches. First, Chunk Incremental Principal Component Analysis (CIPCA) detects intrusion in data streams. Then, two novel incremental feature extraction methods, Incremental Structured Sparse PCA (ISSPCA) and Incremental Generalized Power Method Sparse PCA (IGSPCA), find malicious elements. Metrics helped compare the performance of all methods. The IGSPCA was found to perform as well as or better than CIPCA overall in term of dimensionality reduction, classification accuracy, and learning time. ISSPCA yielded better results for higher chunk values and greater accumulation ratio thresholds. CIPCA and IGSPCA reduced the IDS dataset to 10 principal components as opposed to 14 eigenvectors for ISSPCA. ISSPCA is more expensive in terms of learning time in comparison to the other techniques. This dissertation presents new methods that perform feature extraction from continuous data streams to find the small number of features necessary to express the most data variance. Data subsets derived from a few important variables render their interpretation easier. Another goal of this dissertation was to propose incremental sparse PCA algorithms capable to process data with concept drift and concept shift. Experiments using WaveForm and WaveFormNoise datasets confirmed this ability. Similar to CIPCA, the ISSPCA and IGSPCA updated eigen-axes as a function of the accumulation ratio value, forming informative eigenspace with few eigenvectors.

As more devices with different service requirements become connected to networked systems, such as Internet of Things (IoT) devices, maintaining quality of service becomes increasingly difficult. Large data sets can be obtained ahead of time in networks to train prediction models offline, however, resulting in high computational costs. Online learning is an alternative approach where a smaller cache of fixed size is maintained for training using sample selection algorithms, allowing for lower computational costs and real-time model re-computation. This project has resulted in two newly designed sample selection algorithms, Binned Relevance and Redundancy Sample Selection (BRR-SS) and Autoregressive First, In First Out-buffer (AR-FIFO). The algorithms are evaluated on data traces retrieved from a Key Value store and a Video on Demand service. Prediction accuracy of the resulting model while using the sample selection algorithms and the time to process a received sample is evaluated and compared to the pre-existing Reservoir Sampling (RS) and Relevance and Redundancy Sample Selection (RR-SS) with and without model re-computation. The results show that, while RS maintains the lowest computational overhead, BRR-SS outperforms both RS and RR-SS in prediction accuracy on the investigated traces. AR-FIFO, with its low computational cost, outperforms offline learning for larger cache sizes on the Key Value data set but shows inconsistencies on the Video on Demand trace. Model re-computation results in reduced error rates and significantly lowered variance on the investigated data traces, where periodic model re-computation overall outperforms change detection in practicality, prediction accuracy, and computational overhead. / Allteftersom fler enheter med olika servicekrav ansluts till nätverkssystem, såsom Internet of Things (IoT) enheter, ökar svårigheten att erhålla nödvändig servicekvalitet. Nätverk kan ge upphov till stora datamängder för träning av prediktionsmodeller offline, dock till en hög beräkningskostnad. Ett alternativt tillvägagångssätt är onlineinlärning där en mindre cache av fast storlek upprätthålls för träning med hjälp av datapunkturvalsalgoritmer. Detta möjliggör lägre beräkningskostnader samt realtidsmodellomräkningar. Detta projekt har resulterat i två nydesignade datapunkturvalsalgoritmer, Binned Relevance and Redundancy Sample Selection (BRR-SS) och Autoregressive First In, First Out-buffer (AR-FIFO). Algoritmerna utvärderas på dataspår som hämtats från ett Key Value-lager och en Video on Demand-tjänst. Förutsägelseförmåga för den resulterande modellen när datapunkturvalsalgoritmerna används och tid för bearbetning av mottagen datapunkt utvärderas och jämförs med dem redan existerande Reservoir Sampling (RS) och Relevance and Redundancy Sample Selection (RR-SS), med och utan modellomräkning. RS resulterar i lägst beräkningskostnad medan BRR-SS överträffar både RS och RR-SS i förutsägelseförmåga på dem undersökta spåren. AR-FIFO, med sin låga beräkningskostnad, överträffar offlineinlärning för större cachestorlekar på Key Value-spåret, men visar inkonsekvent beteende på Video on Demand-spåret. Modellomräkning resulterar i mindre fel och avsevärt sänkt varians på dem undersökta spåren, där periodisk modellomräkning totalt sett överträffar förändringsdetektering i praktikalitet, förutsägelseförmåga och beräkningskostnad. / Kandidatexjobb i elektroteknik 2022, KTH, Stockholm