• 9/15/06 Boosting Strategies
  Speaker: Yuning Xu
  

  Back to top

• 9/29/06 Semantic Subspace Learning
  Speaker: Jerry Yu
  

  Back to top

• 10/6/06 2DPCA and 2DLDA for Face Representation and Recognition
  Speaker: Yijuan Lu
  Paper: "GPCA: an efficient dimension reduction scheme for image compression and retrieval", Jieping Ye, Ravi Janardan, and Qi Li, KDD '04: Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining: 354-363, New York, NY, 2004.

  Back to top

• 10/13/06 Using Validation Sets to Avoid Overfitting in AdaBoost
  Speaker: Lisa Tate
  Paper: "Using Validation Sets to Avoid Overfitting in AdaBoost", Tom Bylander and Lisa Tate, Proceedings of the Nineteenth International Florida Artificial Intelligence Research Society Conference: 544-549, Melbourne Beach, Florida,2006.
  Abstract:
  AdaBoost is a well known, effective technique for increasing the accuracy of learning algorithms. However, it has the potential to overfit the training set because its objective is to minimize error on the training set, especially in the presence of noisy datasets. We demonstrate that overfitting in AdaBoost can be alleviated in a time-efficient manner using a combination of subset training and validation sets. Subset training creates subsets from the training set, trains on each subset, and aggregates the results for final classification. This paper focuses on applying dagging to AdaBoost. Dagging partitions the training set into disjoint subsets, each subset is trained on AdaBoost, the results are aggregated with weighted majority vote. We modified the original dagging algorithm to use the entire training set for validation. We will show that dagging with validation alleviates overfitting in AdaBoost with noisy datasets.

  Back to top

• 10/20/06 Learning Regulatory Networks from Sparsely Sampled Time Series Expression Data
  Speaker: Jia Meng
  Paper: "Learning Regulatory Networks from Sparsely Sampled Time Series Expression Data", A. Kundaje, O. Antar, T. Jebara and C. Leslie, Technical Report, 2002.
  Abstract: We present a probabilistic modeling approach to learning gene transcriptional regulation networks from time series gene expression data that is appropriate for the sparsely and irregularly sampled time series datasets currently available. We use a clustering algorithm based on statistical splines to estimate continuous probabilistic models for clusters of genes with similar time expression profile and for individual genes. Using the learned models, we present a novel mutual information score for causal edges between pairs of clusters and between pairs of genes corresponding to a given time lag. This score computes dependency between expression values as continuous quantities rather than discretizing them. We present empirical results on times series data for the yeast cell cycle, using randomization trials to determine statistically significant candidate network edges and the Chow-Liu graph learning algorithm to learn the network structure, to obtain a dynamic model of cell cycle regulation. Biological validation of the inferred network suggests that our method can learn a meaningful, higher-level view of regulatory networks from sparse time series data.

  Back to top

• 11/3/06 Incremental Nonlinear Dimensionality Reduction by Manifold Learning
  Speaker: Dragana Veljkovic
  Paper: Law, M. H. C. and A. K. Jain (2006). "Incremental nonlinear dimensionality reduction by manifold learning." IEEE transactions on pattern analysis and machine intelligence 28(3): 377-391
  Abstract: Understanding the structure of multidimensional patterns, especially in unsupervised cases, is of fundamental importance in data mining, pattern recognition, and machine learning. Several algorithms have been proposed to analyze the structure of high-dimensional data based on the notion of manifold learning. These algorithms have been used to extract the intrinsic characteristics of different types of high-dimensional data by performing nonlinear dimensionality reduction. Most of these algorithms operate in a ?batch? mode and cannot be efficiently applied when data are collected sequentially. The authors describe an incremental version of ISOMAP, one of the key manifold learning algorithms. The experiments on synthetic data as well as real world images demonstrate that the modified algorithm can maintain an accurate low-dimensional representation of the data in an efficient manner.

  Back to top

• 11/10/06 Similarity Metrics with Lp and Geodesic Variations
  Speaker: Atlas Cook IV
  Papers: Alt, H.; Behrends, B. & Blömer, J. Approximate matching of polygonal shapes (extended abstract) SCG '91: Proceedings of the seventh annual symposium on Computational geometry, ACM Press, 1991, 186-193

  Alt, H. & Godau, M. Computing the Frechet Distance Between Two Polygonal Curves International Journal of Computational Geometry and Applications, 1995, 5, 75-91

  Ben-Or, M. Lower bounds for algebraic computation trees STOC '83: Proceedings of the fifteenth annual ACM symposium on Theory of computing, ACM Press, 1983, 80-86

  Maheshwari, A. & Yi, J. On Computing Frechet Distance of Two Paths on a Convex Polyhedron EWCG 2005, 2005 , 41-4

  Mitchell, J. Geometric shortest paths and network optimization Handbook of Computational Geometry, 1998

  Abstract: Similarity metrics measure how "similar" two objects are to each other. They are useful for authentication, map matching, and shape recognition. These applications rely on using similarity metrics to find an input object's "closest match" in a known collection of objects. For example, given a thumbprint one can find the closest matching thumbprint in a thumbprint database. Although there are many ways to compare objects, both the Hausdorff and Fréchet similarity metrics compare objects based on distance. This means that "similar" objects are close together while "different" objects are far apart. Given that "distance" is key to calculating similarity, it is interesting to change how "distance" is measured. For example, distance can be measured using Lp metrics or with shortest obstacle-avoiding paths (geodesics).

  Back to top

• 11/17/06 l-Diversity: Privacy Beyond k-Anonymity
  Speaker: Hongwei Tian
  Paper: Ashwin Machanavajjhala, Johannes Gehrke, Daniel Kifer, and Muthuramakrishnan Venkitasubramaniam. "l-Diversity: Privacy Beyond k-Anonymity". In Proceedings of the 22nd IEEE International Conference on Data Engineering (ICDE 2006), Atlanta Georgia, April 2006

  Abstract: Publishing data about individuals without revealing sensitive information about them is an important problem. In recent years, a new definition of privacy called k-anonymity has gained popularity. In a k-anonymized dataset, each record is indistinguishable from at least k-1 other records with respect to certain "identifying" attributes. However, there are two simple attacks that a k-anonymized dataset has some subtle, but severe privacy problems: First, an attacker can discover the values of sensitive attributes when there is little diversity in those sensitive attributes; second, attackers often have background knowledge. Unfortunately k-anonymity does not guarantee privacy against such attackers. Based on the k-anonymity, they propose a novel and powerful privacy definition called l-diversity. In addition to building a formal foundation for l-diversity, they show how to inject utilities to randomized datasets.

  Back to top

• 12/1/06 A New Scheme on Privacy-Preserving Data Classification
  Speaker: Lijie Zhang
  Paper: Nan Zhang, Shengquan Wang, and Wei Zhao, "A New Scheme on Privacy-Preserving Data Classification", KDD'05, Chicago, Illinois, USA.

  Abstract: This paper addresses privacy-preserving classification problem in a distributed system. Randomization has been the approach proposed to preserve privacy in such scenario. However, this approach has been proven to be insecure as it has been discovered that some privacy intrusion techniques can be used to reconstruct private information from the randomized data tuples. So it introduces an algebraic technique- based scheme. Compared to the randomization approach, this scheme can build classifiers more accurately but disclose less private information.

  Back to top

• Fall 2004
• Spring 2005
• Fall 2005
• Spring 2006

Back to top

Please send emails to carola@cs.utsa.edu, or seminar co-organizers: Kay Robbins, Weining Zhang, Yufei Huang, Carola Wenk, and Qi Tian.