ASME14_Ningrong
Download as PPTX, PDF
1 like275 views
This document proposes a decision support system to help with data-driven management. It involves both an offline and online system. The offline system determines the best algorithms for dimension reduction, clustering, and classification using market and product data. The online system then applies these algorithms to segment markets and position products. A case study on automobile data with 639 models and 31 properties evaluates different dimension reduction, clustering, and classification techniques. Support vector machines achieved the highest classification accuracy on average. The proposed system provides a blueprint for objectively analyzing data to support management decisions.
ASME14_Ningrong
- 1. Decision Support Systems Design for Data-Driven Management PhD Candidate: Ningrong LEI Supervisor: Dr. Seung Ki MOON School of Mechanical & Aerospace Engineering Division of Systems & Engineering Management DETC2014-34871 ASME 2014 International Design & Engineering Technical Conferences, Buffalo, New York, USA
- 2. Outline Data-driven dilemma. Construct the decision support systems. Case study and discussion. Conclusion and future work. 1 DETC2014-34871
- 3. Data alone will not improve management decisions Source: : www.v1shal.com 2
- 4. Market Segmentation Source: . :Cameron Industrials Platform Strategy Advisory 3
- 5. Decision support methods for market segmentation Non data-driven Ground on common sense rather than on a solid empirical base Data-driven Generally lack of reliability assessments Restrict to a few comparison 4
- 6. When data-driven management goes bad Source: : www.v1shal.com 5
- 7. The success of the data-driven management relies on: Quality of the gathered data Reliable model Effectiveness of data analysis Objectiveness of results interpretation 6
- 8. Objectives Develop a reliable decision support system to: Identify market segmentation based on market data Determine product positioning 7
- 9. Online system Block diagram of the proposed DSS Intrinsic dimension estimation Dimension reduction Performance evaluation Clustering Dimension reduction Automated classification Clustering Market Data Offline system Best dimension Reduction algorithm Best clustering algorithm Best training model Product Data 8
- 10. Case study: automobile market 34 brands,639 models,31 model properties 9
- 11. Online system Dimension reduction steps Intrinsic dimension estimation Dimension reduction Performance evaluation Clustering Dimension reduction Automated classification Clustering Market Data Offline system Best dimension Reduction algorithm Best clustering algorithm Best training model Product Data 10
- 12. Dimension reduction techniques Intrinsic dimensionality estimation: Correlation Dimension Estimator Eigenvalue-Based Estimator Maximum Likelihood Estimator Geodesic Minimum Spanning Tree Dimension reduction: Principle Component Analysis Multidimensional Scaling Local Linear Embedding 11
- 13. Online system Block diagram of the proposed DSS Intrinsic dimension estimation Dimension reduction Performance evaluation Clustering Dimension reduction Product Positioning Market Segmentation Market Data Offline system Best dimension Reduction algorithm Best clustering algorithm Best training model Product Data 12
- 14. Clustering K-means Fuzzy C-Means Hierarchical Clustering 13
- 15. Performance Evaluation Statistical tests: median and standard deviation Silhouette Means: 14
- 16. Performance Evaluation How accurately a predictive model will perform in practice? Stratified 10-Fold Cross-Validation Classification: Gentle AdaBoost (GA) Nearest Neighbour (NN) Support Vector Machine (SVM) 15
- 17. Online decision support system 16
- 18. Clustering and performance evaluation results Clustering Performance evaluation Cluster- ing Dimension reduction Dimension estimation Number of clusters Silhouette mean GA in % NN in % SVM in % K- means PCA 2 2 0.67 25.24 98.25 99.68 3 2 0.61 25.08 96.83 99.52 10 11 0.40 95.50 95.50 96.50 MDS 2 2 0.67 74.13 97.78 99.68 3 2 0.61 25.08 96.03 99.52 10 2 0.48 25.71 94.13 100.00 LLE 2 2 0.89 97.38 99.76 100.00 3 3 0.81 99.02 98.05 99.76 10 2 0.85 0.95 99.76 100.00 17
- 19. Classification accuracy over all clusters On average, SVM outperforms NN and GA 18
- 20. Conclusion Deliver a blueprint on how to construct a decision support system. Offline system: find the most suitable algorithms structure; Online system: deliver objective and reliable decision support. Data-driven management: right data, proven statistics, logical explanations. 19