Fatemeh Sheikholeslami

Kernel-based low-rank feature extraction on a budget for big data streams

With nowadays big data torrent, identifying low-dimensional latent structures and extracting features from massive datasets are tasks of paramount importance. To this end, as real data generally lie on (or close to) nonlinear manifolds, kernel-based approaches are well motivated. Being nonparametric, unfortunately kernel-based feature extraction incurs complexity that grows prohibitively with the number of data. In response to this formidable challenge, the present work puts forward a low-rank, kernel-based feature extraction method, where the number of kernel functions is confined to an affordable budget. The resultant algorithm is particularly tailored for online operation, where data streams need not even be stored in memory. Tests on synthetic and real datasets demonstrate and benchmark the efficiency of the proposed method on linear classification applied to the extracted features.

Optimal and Scalable Caching for 5G Using Reinforcement Learning of Space-Time Popularities

Small basestations (SBs) equipped with caching units have potential to handle the unprecedented demand growth in heterogeneous networks. Through low-rate, backhaul connections with the backbone, SBs can prefetch popular files during off-peak traffic hours, and service them to the edge at peak periods. To intelligently prefetch, each SB must learn what and when to cache, while taking into account SB memory limitations, the massive number of available contents, the unknown popularity profiles, as well as the space-time popularity dynamics of user file requests. In this paper, local and global Markov processes model user requests, and a reinforcement learning (RL) framework is put forth for finding the optimal caching policy when the transition probabilities involved are unknown. Joint consideration of global and local popularity demands along with cache-refreshing costs allow for a simple, yet practical asynchronous caching approach. The novel RL-based caching relies on a Q-learning algorithm to implement the optimal policy in an online fashion, thus, enabling the cache control unit at the SB to learn, track, and possibly adapt to the underlying dynamics. To endow the algorithm with scalability, a linear function approximation of the proposed Q-learning scheme is introduced, offering faster convergence as well as reduced complexity and memory requirements. Numerical tests corroborate the merits of the proposed approach in various realistic settings.

Egonet tensor decomposition for community identification

Real-world networks are known to exhibit community structure, characterized by presence of dense node clusters with loose edge connections among them. Although identification of communities is a well-studied subject, most approaches only focus on edge-based criteria which may not incorporate important grouping information captured by higher-order structures e.g., cliques and cycles, to name a few. In order to overcome this limitation, the present paper advocates a novel three-way tensor network representation that captures spatial dependencies among node neighborhoods. Each tensor slice captures a connectivity matrix pertaining to a unique egonet, defined as the subgraph induced by a node and its single-hop neighbors. Constrained tensor factorization is pursued to reveal the hidden and possibly overlapping community structure. Numerical tests on synthetic and real world networks corroborate the efficacy of the novel approach.

Overlapping Community Detection via Constrained PARAFAC: A Divide and Conquer Approach

The task of community detection over complex networks is of paramount importance in a multitude of applications. The present work puts forward a top-to-bottom community identification approach, termed DC-EgoTen, in which an egonet-tensor (EgoTen) based algorithm is developed in a divide-and-conquer (DC) fashion for breaking the network into smaller subgraphs, out of which the underlying communities progressively emerge. In particular, each step of DC-EgoTen forms a multi-dimensional egonet-based representation of the graph, whose induced structure enables casting the task of overlapping community identification as a constrained PARAFAC decomposition. Thanks to the higher representational capacity of tensors, the novel egonet-based representation improves the quality of detected communities by capturing multi-hop connectivity patterns of the network. In addition, the top-to-bottom approach ensures successive refinement of identified communities, so that the desired resolution is achieved. Synthetic as well as real-world tests corroborate the effectiveness of DC-EgoTen.