Giuseppe A. Paleologo

Subagging for credit scoring models

The logistic regression framework has been for long time the most used statistical method when assessing customer credit risk. Recently, a more pragmatic approach has been adopted, where the first issue is credit risk prediction, instead of explanation. In this context, several classification techniques have been shown to perform well on credit scoring, such as support vector machines among others. While the investigation of better classifiers is an important research topic, the specific methodology chosen in real world applications has to deal with the challenges arising from the real world data collected in the industry. Such data are often highly unbalanced, part of the information can be missing and some common hypotheses, such as the i.i.d. one, can be violated. In this paper we present a case study based on a sample of IBM Italian customers, which presents all the challenges mentioned above. The main objective is to build and validate robust models, able to handle missing information, class unbalancedness and non-iid data points. We define a missing data imputation method and propose the use of an ensemble classification technique, subagging, particularly suitable for highly unbalanced data, such as credit scoring data. Both the imputation and subagging steps are embedded in a customized cross-validation loop, which handles dependencies between different credit requests. The methodology has been applied using several classifiers (kernel support vector machines, nearest neighbors, decision trees, Adaboost) and their subagged versions. The use of subagging improves the performance of the base classifier and we will show that subagging decision trees achieve better performance, still keeping the model simple and reasonably interpretable.

The optimality of the online greedy algorithm in carpool and chairman assignment problems

We study several classes of related scheduling problems including the carpool problem, its generalization to arbitrary inputs and the chairman assignment problem. We derive both lower and upper bounds for online algorithms solving these problems. We show that the greedy algorithm is optimal among online algorithms for the chairman assignment problem and the generalized carpool problem. We also consider geometric versions of these problems and show how the bounds adapt to these cases.

Autonomic Admission Control for Networked Information Servers

While the issue of enabling performance guarantees on the Internet has been the subject of intense research in recent years, the problem of enabling QoS guarantees in edge servers has received relatively little attention. The need for QoS guarantees is already present in today’s internet: while most backbones operate at a low level of utilization, web servers are often congested and are the main cause for the delay experienced by the end user. Here, we present a novel approach to admission control and resource allocation of sessions in edge servers. The model we adopt is quite general and its implementation does not depend on the type of application supported by the server (e.g., http or SSL). We model the system as a single server accessed by N + 1 users. N users have a lower bound on QoS, while one “super—user” aggregates the best—effort traffic to the server. The control rules admit a simple interpretation. Admitted classes “track” a target delay which is slightly smaller than their lower bound. The choice of a conservative target protects them from the performance degradation caused by the arrival of candidate classes into the system. On the other hand, candidate classes follow a “slow start” mechanism, similar to the update rule for TCP Reno. The intuitive rationale for this choice is similar to that of congestion—control algorithms: by increasing their priorities slowly, the candidate classes do not degrade the QoS of the admitted classes below their upper bounds. This resource allocation algorithm enjoys several attractive properties: it is measurement—based, since it only relies on the measurement of each class’ delay during a busy cycle; it is decentralized, since each class updates its priority based on local information; and finally it is closed—loop, while most admission control schemes are open—loop. As a consequence, the algorithm does not require signaling to admit a new class.