The rise of digital music distribution has provided users with unprecedented access to vast song catalogs. In order to help users cope with large collections, music information retrieval systems have been developed to automatically analyze, index, and recommend music based on a user's preferences or search criteria. This dissertation proposes machine learning approaches to content-based, query-by-example search, and investigates applications in music information retrieval. The proposed methods automatically infer and optimize content-based similarity, fuse heterogeneous feature modalities, efficiently index and search under the optimized distance metric, and finally, generate sequential playlists for a specified context or style. Robust evaluation procedures are proposed to counteract issues of subjectivity and lack of explicit ground truth in music similarity and playlist generation

We study metric learning as a problem of information retrieval. We present a general metric learning algorithm, based on the structural SVM framework, to learn a metric such that rankings of data induced by dis- tance from a query can be optimized against various ranking measures, such as AUC, Precision-at-k, MRR, MAP or NDCG. We demonstrate ex- perimental results on standard classication data sets, and a large-scale online dating recommendation problem.

Pre-2018 CSE ID: CS2010-0955

We develop an efficient optimization procedure for learning linear transformations of data to produce structured outputs, e.g., nearest-neighbor rankings or connectivity graphs induced by distance. The proposed method solves a semi-definite programming problem by reducing to a sequence of small quadratic programs, resulting in significant reductions in training complexity. Experiments demonstrate that the proposed method is robust, efficient, and outperforms alternative methods in high-noise settings.

Pre-2018 CSE ID: CS2012-0979

Accurately estimating all link-level properties of a large network has proven to be very difficult. The measurements used for these estimates require significant collaboration from all endpoints on the network, significantly reducing their applicability for large scale Internet measurements. We present a scalable approach using a small number of hosts without collaboration from existing routers and minimal collaboration between the hosts. Our approach is based on adaptive sampling. Initially, each host probes a set of receivers at a low frequency. When packet losses are detected, the sampling rate increases. By detecting correlations between time series and combining them with information about network connectivity, the host identifies a set of suspected lossy routers. Hosts then communicate with each other, combining evidence to identify routers with high packet loss. Our experiments show that using a relatively small set of hosts and receivers, we can gather sufficient evidence to identify a small number of routers that cause most of the packet loss in a geographically diverse sample of the Internet. We deployed our method for one month on 68 PlanetLab nodes. As a result of that deployment, we identified 128 routers of the ~4,500 accounting for 87% of the observed packet loss.

Pre-2018 CSE ID: CS2007-0891