Application-level network relays possess many desirable properties, including support for communication between disconnected clients, increasing bandwidth between distant clients, and enabling routing around Internet failures. One problem not considered by existing systems is how to assign client load to relay servers in order to maximize throughput of the relay-system. In this paper, we are interested in the particular case where network conditions change frequently so that the ability of clients to adapt flow is restricted and each round of activity is critical. To this end, we present an algorithm, called Aggressive Increase, AAI which improves its competitive ratio in each time round that the network conditions persists. Given a relay network where a client connects to at most N servers, if network conditions persist for log(N) rounds then the algorithm's throughput becomes constant competitive. Our results improve upon the competitive ratio of previous work (of Awerbuch, Hajiaghayi, Kleinberg, and Leighton [2]). In addition we show that the AAI algorithm performs well in simulation studies as compared with the algorithm of [2] and an adaptation of the multiplicative increase algorithm of [8]. On a variety of input graphs, we show that the AAI algorithm typically reaches close to peak bandwidth levels within only a small constant (< 10) number of rounds.
Our collaborative partitioning model posits a bicriteria objective in which we seek the best item clustering that satisfies the most users at the highest level of satisfaction. We consider two basic methods for determining user satisfaction. The first method is based on how well each user's preferences match a given partition, and the second method is based on average correlation scores taken over sufficiently large subpopulations of users. We show these problems are NP-Hard and develop a set of heuristic approaches for solving them. We provide lower bounds on the satisfaction level on random data, and error bounds in the planted partition model, which provide confidence levels for our heuristic methods. Finally, we present experiments on several real examples that demonstrate the effectiveness of our framework.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.