Presentation is loading. Please wait.

Presentation is loading. Please wait.

Patch Based Prediction Techniques University of Houston By: Paul AMALAMAN From: UH-DMML Lab Director: Dr. Eick.

Similar presentations


Presentation on theme: "Patch Based Prediction Techniques University of Houston By: Paul AMALAMAN From: UH-DMML Lab Director: Dr. Eick."— Presentation transcript:

1 Patch Based Prediction Techniques University of Houston By: Paul AMALAMAN From: UH-DMML Lab Director: Dr. Eick

2 Introduction 1. Research Goals 2. Problem Setting 3. Solutions: TPRTI-A & TPRTI-B 4. Results 5. Conclusion Future Work UH-DMML Lab 1

3 Research Goals To improve Machine Learning techniques for inducing predictive models based on efficient subdivisions of the input space (patches) Areas of Focus:  Linear Regression Tree Induction  Classification Tree Induction UH-DMML Lab 2

4 Linear regression is a global model, where there is a single predictive formula holding over the entire data-space. Y = β 0 + β T X + Linear Regression Tree When the data has lots of input attributes which interact in complicated, nonlinear ways, assembling a single global model can be very difficult. An alternative approach to nonlinear regression is to split, or partition, the space into smaller regions, where the interactions are more manageable. We then partition the sub-divisions again - this is called recursive partitioning - until finally we get to chunks of the space which can fit simple models to them. Splitting Method  selecting the pair {split variable, split value}  minimizing some error/objective function Background (Research Goals Continued) UH-DMML Lab 3

5 Popular approaches: 1-Variance-based  {split variable, split value} selection: try each mean value of each input attribute  objective function: variance minimization scalable, complex trees, often less accurate 2-RSS-based  {split variable, split value} selection: try each value for each input attribute ( Exhaustive search)  objective function: RSS minimization (Residual Sum of Squared Errors) Less scalable, smaller trees, better accuracy Our Research Goals: To induce smaller trees with better accuracy while improving on scalability by designing better splitting methods (patches), and objective functions UH-DMML Lab 4 Background (Research Goals Continued)

6 UH-DMML Lab 5 y y (A) (D) (G)(B)(C) (E) (F) A A B B G G Problem Setting x x Exhaustive search Variance-based

7 y (A) (B)(C) A A B B C C x Problem Setting Our Research Goals: To induce smaller trees with better accuracy while improving on scalability by designing better splitting methods (patches), and objective functions UH-DMML Lab 6

8 1-Variance –based approaches like M5 will miss the optimum split point 2-Exhaustive search approaches like RETIS will find optimum split point but at cost of expensive search (not scalable) Example (Problem Setting Continued) UH-DMML Lab 7

9 Current Proposed Solution o Detect areas in the dataset where the general trend makes sharp turns (Turning Points) o Use Turning Points as potential split points in a Linear Regression Tree induction algorithm Challenges:  Determining the turning points  Balancing accuracy, model complexity, and runtime complexity UH-DMML Lab 8 Solutions

10 UH-DMML Lab 9 Determining Turning Points (Solutions continued)

11 Two algorithms: TPRTI-A and TPRTI-B Both rely on 1.detecting potential split points in the dataset (turning points) 2.then feed a tree induction algorithm with the split points TPRTI-A and TPRTI-B differ by their objective functions o TPRTI-A RSS based node evaluation approach o TPRTI-B uses a two steps node evaluation function  Select split point based on distance  Use RSS computation to select the pair {split variable/split value} UH-DMML Lab 10 Two New Algorithms (Solutions continued)

12 TPRTI-A RSS based node evaluation approach. Does a look-ahead split for each turning point and select the split that best minimizes RSS UH-DMML Lab 11 TPRTI-B uses a two steps node evaluation function  Select split point based on distance  Use RSS computation to select the pair {split variable/split value} Two New Algorithms (Solutions continued)

13 M5TPRTI-ATPRTI-BRETISGUIDESECRET TPRTI-A(6/5/1)-(4/6/2)(4/6/0)(5/1/2)(4/2/2) TPRTI-B(4/6/2)(2/6/4)-(3/7/0)(5/1/2)(1/4/3) Table1. Comparison between TPRTI-A, TPRTI-B and state-of-the-art approaches with respect to accuracy (wins/ties/loses) Results On Accuracy UH-DMML Lab 12

14 Results On Complexity Table2. Number of times an approach obtained the combination (Best accuracy, fewest leaf-nodes) M5TPRTI-ATPRTI-BRETISGUIDESECRET 05 35N.A. UH-DMML Lab 13

15 Results On Scalability UH-DMML Lab 14

16 We propose a new approach for Linear Regression Tree construction called Turning Point Regression Tree Induction (TPRTI) that infuses turning points into a regression tree induction algorithm to achieve  improved scalability while maintaining high accuracy and low model complexity.  Two novel linear regression tree induction algorithms called TPRTI-A and TPRTI-B which incorporate turning points into the node evaluation were introduced and experimental results indicate that TPRTI is a scalable algorithm that is capable of obtaining a high predictive accuracy using smaller decision trees than other approaches. Conclusion UH-DMML Lab 15

17 FUTURE WORK We are investigating how turning point detection can also be used to induce better classification trees. UH-DMML Lab 16

18 Thank You


Download ppt "Patch Based Prediction Techniques University of Houston By: Paul AMALAMAN From: UH-DMML Lab Director: Dr. Eick."

Similar presentations


Ads by Google