Final Project

The objective of the final project is to integrate various classification techniques to achieve the best performance. Final project is a group effort. Each group can have 2-3 members. For undergraduate students, you are required to apply all techniques learned in this semester. For graduate student, besides the learned techniques in class, you are required to propose some improvements either in computational speed or classification accuracy.

Schedule

• Milestone 1: Group formation and topic selection (due 04/16). Submit through Canvas. Approval and comments will be returned in one day. The same topic cannot be chosen by more than 1 group. The topic follows the first-come first-served rule. So decide your group formation and pick a topic as soon as possible.
• Milestone 2: Background study (2-page report due 04/25). Submit through Canvas.
• Final presentation (due 05/06). Submit through Canvas.
• Final presentation (05/07, 8:00-10:00)
• Final report (due 05/07). Submit through Canvas.

Potential Topics

• the UCI Machine Learning Repository (471 only)
• KDD-Cup 1997-2006 (471/571)
• Kaggle Competitions (471/571)
• Other topics: You can select a topic yourself from other resources with the approval from the instructor.

Requirement

General steps involved in a pattern recognition problem include

• Data collection (raw data)
• Feature derivation (how to derive features from the raw data)
• Feature selection (dimensionality reduction, Fisher's linear discriminant or PCA)
• Classification
  • classification based on supervised learning or unsupervised learning
  • classification based on parametric or non-parametric density estimation in supervised learning
  • classifier fusion
• Performance evaluation
• Feedback system

• the effect of assuming the data is Gaussian-distributed
• the effect of assuming parametric pdf vs. non-parametric pdf
• the effect of using different prior probability ratio
• the effect of using different orders of Minkowski distance
• the effect of knowing the class label
• the effect of dimension of the feature space (e.g., changed through dimensionality reduction)
• the effect of classifier fusion

To be more specific, you need to at least go through the following steps:

• Data normalization
• Dimensionality reduction
• Classification with the following
  • MPP (case 1, 2, and 3)
  • kNN with different k's
  • BPNN (can use open-source software package or the toolbox comes with MATLAB)
  • Decision tree (can use open-source software package or the toolbox comes with MATLAB)
  • Graduate students only: SVM (can use open-source software package or the toolbox comes with MATLAB)
  • clustering (kmeans, wta, kohonen, or mean-shift)
• Classifier fusion
• Evaluation (use n-fold cross validation to generate confusion matrix and ROC curve if applicable).