Ensembles of example dependent cost-sensitive decision trees slides
3 likes5,330 views
The document discusses the development of cost-sensitive decision tree ensembles aimed at improving classification in applications like credit card fraud detection and churn modeling. It presents a new framework that combines various algorithms and incorporates example-dependent cost metrics, demonstrating superior performance compared to traditional methods. The findings emphasize the significance of recognizing real-world financial implications when designing classification systems.
![We define a cost measure based on the cost matrix [Elkan 2001]
From which we calculate the cost of applying a classifier to a given set
Background - Cost-sensitive evaluation
5](https://image.slidesharecdn.com/ensemblesofexample-dependentcost-sensitivedecisiontreesslides-150526100803-lva1-app6892/85/Ensembles-of-example-dependent-cost-sensitive-decision-trees-slides-5-320.jpg)
![Research in example-dependent cost-sensitive classification has been
narrow, mostly because of the lack of publicly available datasets [Aodha
and Brostow 2013].
Standard approaches consist in re-weighting the training examples based
on their costs:
• Cost-proportionate rejection sampling [Zadrozny et al. 2003]
• Cost-proportionate oversampling [Elkan 2001]
Background - State-of-the-art methods
7](https://image.slidesharecdn.com/ensemblesofexample-dependentcost-sensitivedecisiontreesslides-150526100803-lva1-app6892/85/Ensembles-of-example-dependent-cost-sensitive-decision-trees-slides-7-320.jpg)
![• Proposed cost-sensitive stacking
Using the cost-sensitive logistic regression [Correa et. al, 2014] model:
Then the weights are estimated using
Proposed combination methods
14](https://image.slidesharecdn.com/ensemblesofexample-dependentcost-sensitivedecisiontreesslides-150526100803-lva1-app6892/85/Ensembles-of-example-dependent-cost-sensitive-decision-trees-slides-14-320.jpg)
Ensembles of example dependent cost-sensitive decision trees slides
- 1. Ensembles of example-dependent cost-sensitive decision trees April 28, 2015 Alejandro Correa Bahnsen with Djamila Aouada, SnT Björn Ottersten, SnT
- 2. Motivation 2 • Classification: predicting the class of a set of examples given their features. • Standard classification methods aim at minimizing the errors • Such a traditional framework assumes that all misclassification errors carry the same cost • This is not the case in many real-world applications: Credit card fraud detection, churn modeling, credit scoring, direct marketing.
- 3. • Cost-sensitive classification Background, previous contributions • Cost-sensitive Ensembles Introduction, random inducers, combination methods, propose algorithms • Datasets Credit card fraud detection, churn modeling, credit scoring, direct marketing • Experiments Experimental setup, results • Conclusions Contributions Agenda 3
- 4. predict the class of set of examples given their features Where each element of S is composed by It is usually evaluated using a traditional misclassification measure such as Accuracy, F1Score, AUC, among others. However, these measures assumes that different misclassification errors carry the same cost Background - Binary classification 4
- 5. We define a cost measure based on the cost matrix [Elkan 2001] From which we calculate the cost of applying a classifier to a given set Background - Cost-sensitive evaluation 5
- 6. However, the total cost may not be easy to interpret. Therefore, we propose a savings measure as the cost vs. the cost of using no algorithm at all Where is the cost of predicting the costless class Background - Cost-sensitive evaluation 6
- 7. Research in example-dependent cost-sensitive classification has been narrow, mostly because of the lack of publicly available datasets [Aodha and Brostow 2013]. Standard approaches consist in re-weighting the training examples based on their costs: • Cost-proportionate rejection sampling [Zadrozny et al. 2003] • Cost-proportionate oversampling [Elkan 2001] Background - State-of-the-art methods 7
- 8. • Bayes minimum risk A. Correa Bahnsen, A. Stojanovic, D. Aouada, and B. Ottersten, “Cost Sensitive Credit Card Fraud Detection Using Bayes Minimum Risk,” in 2013 12th International Conference on Machine Learning and Applications. Miami, USA: IEEE, Dec. 2013, pp. 333–338. • Probability calibration for Bayes minimum risk (BMR) A. Correa Bahnsen, A. Stojanovic, D. Aouada, and B. Ottersten, “Improving Credit Card Fraud Detection with Calibrated Probabilities,” in Proceedings of the fourteenth SIAM International Conference on Data Mining, Philadelphia, USA, 2014, pp. 677 – 685. • Cost-sensitive logistic regression (CSLR) A. Correa Bahnsen, D. Aouada, and B. Ottersten, “Example-Dependent Cost-Sensitive Logistic Regression for Credit Scoring,” in 2014 13th International Conference on Machine Learning and Applications. Detroit, USA: IEEE, 2014, pp. 263–269. • Cost-sensitive decision trees (CSDT) A. Correa Bahnsen, D. Aouada, and B. Ottersten, “Example-Dependent Cost-Sensitive Decision Trees,” Expert Systems with Applications, in press, 2015. Previous contributions 8
- 9. • Cost-sensitive classification Background, previous contributions • Cost-sensitive Ensembles Introduction, random inducers, combination methods, propose algorithms • Datasets Credit card fraud detection, churn modeling, credit scoring, direct marketing • Experiments Experimental setup, results • Conclusions Contributions Agenda 9
- 10. The main idea behind the ensemble methodology is to combine several individual base classifiers in order to have a classifier that outperforms everyone of them “The Blind Men and the Elephant”, Godfrey Saxe’s Introduction - Ensemble learning 10 Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Some unknown distribution
- 11. Typical ensemble is made by combining T different base classifiers. Each base classifiers is trained by applying algorithm M in a random subset Introduction - Ensemble learning 11
- 12. Random inducers 12 1 2 3 4 5 6 7 8 8 6 2 5 2 1 3 6 7 1 2 3 8 1 5 8 1 4 4 2 1 9 4 6 1 1 5 8 1 4 4 2 1 1 5 8 1 4 4 2 1 1 5 8 1 4 4 2 1 Bagging Pasting Random forest Random patches Training set
- 13. After the base classifiers are constructed they are typically combined using one of the following methods: • Majority voting • Proposed cost-sensitive weighted voting Proposed combination methods 13
- 14. • Proposed cost-sensitive stacking Using the cost-sensitive logistic regression [Correa et. al, 2014] model: Then the weights are estimated using Proposed combination methods 14
- 15. The subsampling can be done either by: Bagging, pasting, random forest or random patches Proposed algorithms Base classifiers For j in 1..T: 1.Subsample from training set 𝑆𝑗 ← Subsample(𝑆) 2.Train a CSDT on 𝑺𝒋 M𝑗 ← M(𝑆𝑗) 3.Estimate the weight α𝑗 ← 𝑠𝑎𝑣𝑖𝑛𝑔𝑠 M𝑗 𝑆𝑗 𝑜𝑜𝑏 Combination Select combination method: 1.Majority voting 𝐻 ← 𝑓𝑚𝑣 𝑆, 𝑀 2.CS-Weighted voting 𝐻 ← 𝑓𝑚𝑣 𝑆, 𝑀, 𝛼 3.CS-Stacking 𝛽 ← 𝑎𝑟𝑔𝑚𝑖𝑛 𝐽(𝑆, 𝑀, 𝛽) 𝐻 ← 𝑓𝑠 𝑆, 𝑀, 𝛽 15
- 16. • Cost-sensitive classification Background, previous contributions • Cost-sensitive Ensembles Introduction, random inducers, combination methods, propose algorithms • Datasets Credit card fraud detection, churn modeling, credit scoring, direct marketing • Experiments Experimental setup, results • Conclusions Contributions Agenda 16
- 17. Cost matrix Database A. Correa Bahnsen, A. Stojanovic, D. Aouada, and B. Ottersten, “Cost Sensitive Credit Card Fraud Detection Using Bayes Minimum Risk,” in 2013 12th International Conference on Machine Learning and Applications. Miami, USA: IEEE, Dec. 2013, pp. 333–338. Credit card fraud detection 17 # Examples % Positives Cost (Euros) 1,638,772 0.21% 860,448
- 18. Cost matrix Database A. Correa Bahnsen, A. Stojanovic, D. Aouada, and B. Ottersten, “A novel costsensitive framework for customer churn predictive modeling,” Decision Analytics, vol. under review, 2015. Churn modeling 18 # Examples % Positives Cost (Euros) 9,410 4.83% 580,884
- 19. Cost matrix Database A. Correa Bahnsen, D. Aouada, and B. Ottersten, “Example-Dependent Cost-Sensitive Logistic Regression for Credit Scoring,” in 2014 13th International Conference on Machine Learning and Applications. Detroit, USA: IEEE, 2014, pp. 263–269. Credit scoring 19 # Examples % Positives Cost (Euros) Kaggle Credit 112,915 6.74% 83,740,181 PAKDD09 Credit 38,969 19.88% 3,117,960
- 20. Cost matrix Database A. Correa Bahnsen, A. Stojanovic, D. Aouada, and B. Ottersten, “Improving Credit Card Fraud Detection with Calibrated Probabilities,” in Proceedings of the fourteenth SIAM International Conference on Data Mining, Philadelphia, USA, 2014, pp. 677 – 685. Direct marketing 20 # Examples % Positives Cost (Euros) 37,931 12.62% 59,507
- 21. • Cost-sensitive classification Background, previous contributions • Cost-sensitive Ensembles Introduction, random inducers, combination methods, propose algorithms • Datasets Credit card fraud detection, churn modeling, credit scoring, direct marketing • Experiments Experimental setup, results • Conclusions Contributions Agenda 21
- 22. • Cost-insensitive (CI): • Decision trees (DT) • Logistic regression (LR) • Random forest (RF) • Under-sampling (u) • Cost-proportionate sampling (CPS): • Cost-proportionate rejection-sampling (r) • Cost-proportionate over-sampling (o) • Bayes minimum risk (BMR) • Cost-sensitive training (CST): • Cost-sensitive logistic regression (CSLR) • Cost-sensitive decision trees (CSDT) Experimental setup - Methods 22
- 23. • Ensemble cost-sensitive decision trees (ECSDT): Random inducers: • Bagging (CSB) • Pasting (CSP) • Random forest (CSRF) • Random patches (CSRP) Combination: • Majority voting (mv) • Cost-sensitive weighted voting (wv) • Cost-sensitive staking (s) Experimental setup - Methods 23
- 24. • Each experiment was carry out 50 times • For the parameters of the algorithms a grid search was made • Results are measured by savings • Then the Friedman ranking is calculated for each method Experimental setup 24
- 25. Results 25 Results of the Friedman rank of the savings (1=best, 28=worst) Family Algorithm Rank ECSDT CSRP-wv-t 2.6 ECSDT CSRP-s-t 3.4 ECSDT CSRP-mv-t 4 ECSDT CSB-wv-t 5.6 ECSDT CSP-wv-t 7.4 ECSDT CSB-mv-t 8.2 ECSDT CSRF-wv-t 9.4 BMR RF-t-BMR 9.4 ECSDT CSP-s-t 9.6 ECSDT CSP-mv-t 10.2 ECSDT CSB-s-t 10.2 BMR LR-t-BMR 11.2 CPS RF-r 11.6 CST CSDT-t 12.6 Family Algorithm Rank CST CSLR-t 14.4 ECSDT CSRF-mv-t 15.2 ECSDT CSRF-s-t 16 CI RF-u 17.2 CPS LR-r 19 BMR DT-t-BMR 19 CPS LR-o 21 CPS DT-r 22.6 CI LR-u 22.8 CPS RF-o 22.8 CI DT-u 24.4 CPS DT-o 25 CI DT-t 26 CI RF-t 26.2
- 26. Results 26 Results of the Friedman rank of the savings organized by family
- 27. Results 27 Percentage of the highest savings Database Algorithm Savings Fraud CSRP-wv-t 0.73 Churn CSRP-s-t 0.17 Credit1 CSRP-mv-t 0.52 Credit2 LR-t-BMR 0.31 Marketing LR-t-BMR 0.5
- 28. Results within the ECSDT family 28 By combination methodBy random inducer
- 29. • New framework for ensembles of example dependent cost-sensitive decision trees • Using five databases, from four real-world applications: credit card fraud detection, churn modeling, credit scoring and direct marketing, we show that the proposed algorithm significantly outperforms the state-of-the- art cost-insensitive and example-dependent cost-sensitive algorithms • Highlight the importance of using the real example-dependent financial costs associated with the real-world applications Conclusions 29
- 30. Costcla - Software CostCla is a Python module for cost-sensitive machine learning built on top of Scikit-Learn, SciPy and distributed under the 3-Clause BSD license. In particular, it provides: • A set of example-dependent cost-sensitive algorithms • Different real-world example-dependent cost-sensitive datasets. Installation pip install costcla Documentation: https://pythonhosted.org/costcla/ Development: https://github.com/albahnsen/CostSensitiveClassification 30
- 33. Thank You!! Alejandro Correa Bahnsen