Datasets:
File size: 5,304 Bytes
f6cc031 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 | <Poster Width="1734" Height="1340"> <Panel left="-1" right="195" width="870" height="231"> <Text>Motivation</Text> <Text> Several applications have different costs associated with different classification-errors</Text> <Text>Example: intrusion detection, biometric recognition, etc.</Text> <Text> Most classification systems are geared towards minimizing the error rate and not cost</Text> <Text>True objective function to be minimized is the cost of classification-error and not error-rate itself</Text> <Text> Existing approaches can not handle multi-class problems or dynamically changing costs</Text> <Text>ROC curves (multi-class? [1] ) ; cost-sensitive Adaboost [2] (dynamically changing costs? )</Text> </Panel> <Panel left="868" right="196" width="513" height="230"> <Text>Goal</Text> <Text>Develop a classification-error cost minimization strategy that</Text> <Text> Can deal with multiple classes in a principled manner</Text> <Text> Is a simple post-training step</Text> <Text>Does not require re-training of classifiers for changing costs</Text> <Text> Is classifier type independent</Text> <Text>Exploits statistical properties of the trained classifier</Text> </Panel> <Panel left="1382" right="197" width="347" height="228"> <Text>Contributions</Text> <Text>costs incurred Statistically significant reduction in</Text> <Text> Effective on</Text> <Text> a variety of applications</Text> <Text> data sets of varying dimensionalities</Text> <Text> a variety of classifier types</Text> </Panel> <Panel left="0" right="425" width="1732" height="414"> <Text>Approach</Text> <Figure left="26" right="473" width="227" height="148" no="1" OriWidth="0.183237" OriHeight="0.0808758 " /> <Figure left="254" right="455" width="247" height="169" no="2" OriWidth="0.190751" OriHeight="0.0965147 " /> <Text> Solution for a two-class, one-feature problem, known distributions</Text> <Text> If unknown distribution</Text> <Text>Estimate with a histogram</Text> <Text> If multiple-features</Text> <Text>Classification system: maps multiple-features to a single score/feature</Text> <Text> If multiple-classes</Text> <Text>High dimensional histogram is not feasible … so then?</Text> <Text>Intuition: Convert C-class problem to C 2-class problems</Text> <Text>We have a trained classification system</Text> <Text>Probability of a misclassified instance classified as</Text> <Text>class c actually belonging to class i:</Text> <Text>Expected cost of false positives:</Text> <Figure left="1045" right="489" width="449" height="307" no="3" OriWidth="0" OriHeight="0 " /> <Text>(Iterate to get a new confusion matrix with new thresholds)</Text> <Text>Final classification decision:</Text> <Text>Pick the class corresponding</Text> <Text>to the score furthest away</Text> <Text>from it’s corresponding</Text> <Text>optimum threshold</Text> </Panel> <Panel left="0" right="841" width="1380" height="404"> <Text>Results</Text> <Text> Synthetic data: MLP neural network</Text> <Figure left="32" right="920" width="237" height="144" no="4" OriWidth="0.268208" OriHeight="0.155496 " /> <Figure left="17" right="1065" width="254" height="177" no="5" OriWidth="0.304624" OriHeight="0.182306 " /> <Text> MIT-DARPA intrusion detection [3]</Text> <Text>0.3 million data points</Text> <Text>5 classes: DenialOfService,</Text> <Text>Probe, UserToRoot, RootToLocal,</Text> <Text>Normal</Text> <Text>Ensemble of classifiers based</Text> <Text>classification system: Learn++ [4]</Text> <Text>(can perform data fusion)41 features</Text> <Text>3 feature sets: traffic, content,</Text> <Text>intrinsic features</Text> <Figure left="645" right="885" width="261" height="158" no="6" OriWidth="0.371098" OriHeight="0.0969616 " /> <Figure left="645" right="1061" width="242" height="174" no="7" OriWidth="0.334682" OriHeight="0.184987 " /> <Text> PCA reduced intrusion detection</Text> <Figure left="939" right="887" width="334" height="168" no="8" OriWidth="0.338728" OriHeight="0.0893655 " /> <Text> Other applications [5]</Text> <Figure left="922" right="1099" width="452" height="130" no="9" OriWidth="0.356069" OriHeight="0.0665773 " /> </Panel> <Panel left="1379" right="841" width="350" height="402"> <Text>References:</Text> <Text>[1] N. Lachiche and P. Flach. Improving accuracy and cost</Text> <Text>of two-class and multi-class probabilistic classifiers using</Text> <Text>ROC curves. ICML, 2003.</Text> <Text>[2] Y. Ma and X. Ding. Robust real-time face detection</Text> <Text>based on cost-sensitive AdaBoost method. ICME, 2003</Text> <Text>[3] The UCI KDD Archive, Information and Computer</Text> <Text>Science, University of California, Irvine,</Text> <Text>http://kdd.ics.uci.edu/ databases/kddcup99/kddcup99.html</Text> <Text>[4] D. Parikh and R. Polikar. An Ensemble-Based</Text> <Text>Incremental Learning Approach to Data Fusion. In IEEE</Text> <Text>Transactions on Systems, Man and Cybernetics, 2007.</Text> <Text>[5] C. Blake and C. Merz. UCI Repository of Machine</Text> <Text>Learning Database at Irvine CA, 2005.</Text> <Text>http://mlearn.ics.uci.edu/MLRepository.html</Text> </Panel> </Poster> |