Churn prediction aims to identify subscribers who are about to transfer their business to a competitor. Since the cost associated with customer acquisition is much greater than the cost of customer retention, churn prediction has emerged as a crucial Business Intelligence (BI) application for modern telecommunication operators. The dominant approach to churn prediction is to model individual customers and derive their likelihood of churn using a predictive model. Recent work has shown that analyzing customers' interactions by assessing the social vicinity of recent churners can improve the accuracy of churn prediction. We propose a novel framework, termed Group-First Churn Prediction, which eliminates the a priori requirement of knowing who recently churned. Specifically, our approach exploits the structure of customer interactions to predict which groups of subscribers are most prone to churn, before even a single member in the group has churned. Our method works by identifying closely-knit groups of subscribers using second order social metrics derived from information theoretic principles. The interactions within each group are then analyzed to identify social leaders. Based on Key Performance Indicators that are derived from these groups, a novel statistical model is used to predict the churn of the groups and their members. Our experimental results, which are based on data from a telecommunication operator with approximately 16 million subscribers, demonstrate the unique advantages of the proposed method. We further provide empirical evidence that our method captures social phenomena in a highly significant manner.

In constrained clustering it is common to model the pairwise constraints as edges on the graph of observations. Using results from graph theory, we analyze such constraint graphs in two contexts, both of immediate value to practitioners. First, we explore the issue of constraint noise under several intuitive noise models. We apply results from random graph theory, which facilitate the analysis of finite-sized graphs and realistic data partitions and noise levels, to obtain a quantification of the effect noisy edges may have on any constrained clustering algorithm under a set of commonly-used assumptions. We also demonstrate the dangers in the common practice of connected-component constraint set augmentation, when used in the presence of noise. Second, we describe two practical randomized algorithms that estimate the number of induced clusters using only a small number of constraints. We conclude with an experimental evaluation that shows the effect of noise on common UCI data sets, as well as some aspects of the behavior of our algorithms.

The problem of maximizing the weighted throughput in various switching settings has been intensively studied recently through competitive analysis. To date, the most general model that has been investigated is the standard CIOQ (Combined Input and Output Queued) switch architecture with internal fabric speedup S ≥ 1. CIOQ switches, that comprise the backbone of packet routing networks, are N × N switches controlled by a switching policy that incorporates two components: Admission control and scheduling. An admission control strategy is essential to determine the packets stored in the FIFO queues in input and output ports, while the scheduling policy conducts the transfer of packets through the internal fabric, from input ports to output ports. The online problem of maximizing the total weighted throughput of CIOQ switches was recently investigated by Kesselman and Rosén [2003]. They presented two different online algorithms for the general problem that achieve non-constant competitive ratios (linear in either the speedup or the number of distinct values, or logarithmic in the value range). We introduce the first constant-competitive algorithm for the general case of the problem, with arbitrary speedup and packet values. Specifically, our algorithm is 8-competitive, and is also simple and easy to implement.

Recently, approximation analysis has been extensively used to study algorithms for routing weighted packets in various network settings. Although different techniques were applied in the analysis of diverse models, one common property was evident: the analysis of input sequences composed solely of two different values is always substantially easier, and many results are known only for restricted value sequences. Motivated by this, we introduce our zero-one principle for switching networks which characterizes a wide range of algorithms for which achieving c-approximation (as well as c-competitiveness) with respect to sequences composed of 0's and 1's implies achieving c-approximation. The zero-one principle proves to be very efficient in the design of switching algorithms, and substantially facilitates their analysis. We present three applications. First, we consider the Multi-Queue QoS Switching model and design a 3-competitive algorithm, improving the result from [6]. Second, we study the Weighted Dynamic Routing problem on a line topology of length k and present a (k+1)-competitive algorithm, which improves and generalizes the results from [1,12]. As a third application, we consider the work of [11], that compares the performance of local algorithms to the global optimum in various network topologies, and generalize their results from 2-value sequences to arbitrary value sequences.

The concept of Quality of Service (QoS) networks has gained growing attention recently, as the traffic volume in the Internet constantly increases, and QoS guarantees are essential to ensure proper operation of most communication based applications. A QoS switch serves m incoming queues by transmitting packets arriving at these queues through one output port, one packet per time unit. Each packet is marked with a value indicating its guaranteed quality of service. Since the queues have bounded capacity and the rate of arriving packets can be much higher than the transmission rate, packets can be lost due to insufficient queue space. The goal is to maximize the total value of transmitted packets. This problem encapsulates two dependent questions: admission control, namely which packets to discard in case of queue overflow, and scheduling, i.e. which queue to use for transmission in each time unit. We use competitive analysis to study online switch performance in QoS based networks. Specifically, we provide a novel generic technique that decouples the admission control and scheduling problems. Our technique transforms any single queue admission control strategy (preemptive or nonpreemptive) to a scheduling and admission control algorithm for our general m queues model, whose competitive ratio is at most twice the competitive ratio of the given admission control strategy. We use our technique to derive concrete algorithms for the general preemptive and nonpreemptive cases, as well as for the interesting special cases of the 2-value model and the unit value model. To the best of our knowledge this is the first result combining both scheduling and admission control decisions for arbitrary packets sequences in multi-queue switches. We also provide a 1.58-competitive randomized algorithm for the unit value case. This case is interesting by itself since most current networks (e.g. IP networks) only support a best-effort service in which all packets streams are treated equally.