PPDSparse: A Parallel Primal and Dual Sparse Method to Extreme Classification
|
|
- Brendan Taylor
- 6 years ago
- Views:
Transcription
1 PPDSparse: A Parallel Primal and Dual Sparse Method to Extreme Classification Ian E.H. Yen 1, Xiangru Huang 2, Wei Dai 1, Pradeep Ravikumar 1, Inderjit S. Dhillon 2 and Eric Xing 1 1 Carnegie Mellon University. 2 University of Texas at Austin KDD, 2017 Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
2 Outline 1 Problem Setting Extreme Classification Related Works 2 Algorithm Separable Loss Algorithm Diagram 3 Theory Analysis of primal and dual sparsity 4 Experimental Results Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
3 Outline 1 Problem Setting Extreme Classification Related Works 2 Algorithm Separable Loss Algorithm Diagram 3 Theory Analysis of primal and dual sparsity 4 Experimental Results Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
4 Extreme Classification Goal: Learn a function h(x) : R D R K from D input features to K output scores that is consistent with labels y {0, 1} K. an E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
5 Extreme Classification Goal: Learn a function h(x) : R D R K from D input features to K output scores that is consistent with labels y {0, 1} K. K is large (e.g ). an E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
6 Extreme Classification Goal: Learn a function h(x) : R D R K from D input features to K output scores that is consistent with labels y {0, 1} K. K is large (e.g ). Average number of positive labels (per sample) k p = 1 N N i=1 ( k y ik). Multiclass: k p = 1; Multilabel: k p K. Average number of positive samples (per class) n p = 1 K K k=1 nk p = 1 K K k=1 ( i y ik). n p = Nk p /K N an E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
7 Extreme Classification We consider Linear Classifier: h(x) := W T x where W R D K. an E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
8 Extreme Classification We consider Linear Classifier: h(x) := W T x where W R D K. Challenge: When K is large, training of simple linear model requires O(NDK) cost. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
9 Outline 1 Problem Setting Extreme Classification Related Works 2 Algorithm Separable Loss Algorithm Diagram 3 Theory Analysis of primal and dual sparsity 4 Experimental Results Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
10 Approaches Approach 1 Structural, i.e. Low-rank or Tree-hierarchy Good accuracy when assumption holds. Lower accuracy when assumptions not hold. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
11 Approaches Approach 1 Structural, i.e. Low-rank or Tree-hierarchy Good accuracy when assumption holds. Lower accuracy when assumptions not hold. Approach 2 Parallelized one-vs-all Good accuracy, slow, parallelizable. Need days on largest dataset with 100 cores. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
12 Approaches Approach 1 Structural, i.e. Low-rank or Tree-hierarchy Good accuracy when assumption holds. Lower accuracy when assumptions not hold. Approach 2 Parallelized one-vs-all Good accuracy, slow, parallelizable. Need days on largest dataset with 100 cores. Approach 3 Primal-Dual Sparse Good accuracy, fast, not parallelizable, memory issue O(DK). Need days on largest dataset. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
13 Approaches Approach 1 Structural, i.e. Low-rank or Tree-hierarchy Good accuracy when assumption holds. Lower accuracy when assumptions not hold. Approach 2 Parallelized one-vs-all Good accuracy, slow, parallelizable. Need days on largest dataset with 100 cores. Approach 3 Primal-Dual Sparse Good accuracy, fast, not parallelizable, memory issue O(DK). Need days on largest dataset. This paper Parallel PD-Sparse Good accuracy, fast, parallelizable. Need only < 30 min on largest dataset with 100 cores. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
14 Outline 1 Problem Setting Extreme Classification Related Works 2 Algorithm Separable Loss Algorithm Diagram 3 Theory Analysis of primal and dual sparsity 4 Experimental Results Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
15 We consider the classwise-separable hinge loss L(z, y) := K l(z k, y k ) = k=1 K max(1 y k z k, 0) k=1 Minimizing a separable loss is equivalent to One-versus-all: ( N K K N ) min l(w T W R D K k x i, y ik ) = l(w T k x i, y ik ) i=1 k=1 k=1 i=1 Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
16 We consider the classwise-separable hinge loss L(z, y) := K l(z k, y k ) = k=1 K max(1 y k z k, 0) k=1 Minimizing a separable loss is equivalent to One-versus-all: ( N K K N ) min l(w T W R D K k x i, y ik ) = l(w T k x i, y ik ) i=1 k=1 To obtain sparse iterates, we add l 1 -penalty on W and add bias per class w 0k. The dual problem of the l 1 -l 2 -regularized problem is: k=1 i=1 min α k R N G(α k ) := 1 2 w(α k) 2 s.t. w(α k ) = prox λ ( ˆX T α k ), 0 α ik 1. N i=1 α ik (1) Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
17 Outline 1 Problem Setting Extreme Classification Related Works 2 Algorithm Separable Loss Algorithm Diagram 3 Theory Analysis of primal and dual sparsity 4 Experimental Results Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
18 Primal-Dual-Sparse Active-set Method ( O nnz(w k )nnz(x j ) }{{} Search ) + nnz(α k )nnz(x i ) per iteration. }{{} Update+Maintain Apply Random Sparsification on (already sparse) w k before search. Update α by Coordinate Descent within A k. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
19 Parallel & Primal-Dual Sparse Method Due to the separable loss, the optimization can be embarrassingly parallelized with one-time communication. The input y k and output w k of each sub-problem are sparse. Can be implemented in a distributed, shared-memory, or two-level parallelization setting. Space: O(nnz(X ) + D). Nearly linear speedup even with thousands of cores. an E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
20 Outline 1 Problem Setting Extreme Classification Related Works 2 Algorithm Separable Loss Algorithm Diagram 3 Theory Analysis of primal and dual sparsity 4 Experimental Results Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
21 Theory: Primal and Dual Sparsity Key Insight: The number of positive samples for each class n p = Nk p K is small. The following results hold if class-wise bias w k0 are added. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
22 Theory: Primal and Dual Sparsity Key Insight: The number of positive samples for each class n p = Nk p K is small. The following results hold if class-wise bias w k0 are added. Step-1: bound w 1, and optimal α 1 in terms of n p : w k 1 2nk p λ, α k 1 4n k p. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
23 Theory: Primal and Dual Sparsity Key Insight: The number of positive samples for each class n p = Nk p K is small. The following results hold if class-wise bias w k0 are added. Step-1: bound w 1, and optimal α 1 in terms of n p : w k 1 2nk p λ, α k 1 4n k p. Step-2: bound nnz(w), and nnz(α) in terms of w 1 and α 1 : nnz( w k ) w k 2 1 δ 2, nnz(α t k ) t 4 α k 2 1 ɛ where w is Random-Sparsified version of w with δ-approximation error in G(α), and ɛ is the desired precision of solution. Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
24 Multilabel Classification Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
25 Multiclass Classification Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
26 Thank you Xiangru Huang Ian E.H. Yen, Xiangru Huang, Wei Dai, Pradeep PPDSparse: Ravikumar, A Parallel Inderjit Primal S. Dhillon and and Dual Eric Sparse Xing Method (shortinst) to Extreme KDD, Classification / 17
Progress Report: Predicting Which Recommended Content Users Click Stanley Jacob, Lingjie Kong
Progress Report: Predicting Which Recommended Content Users Click Stanley Jacob, Lingjie Kong Machine learning models can be used to predict which recommended content users will click on a given website.
More informationIntro Logistic Regression Gradient Descent + SGD
Case Study 1: Estimating Click Probabilities Intro Logistic Regression Gradient Descent + SGD Machine Learning for Big Data CSE547/STAT548, University of Washington Sham Kakade March 29, 2016 1 Ad Placement
More informationRandom Forests. Parametrization and Dynamic Induction
Random Forests Parametrization and Dynamic Induction Simon Bernard Document and Learning research team LITIS laboratory University of Rouen, France décembre 2014 Random Forest Classifiers Random Forests
More informationMachine Learning Models for Sales Time Series Forecasting
Article Machine Learning Models for Sales Time Series Forecasting Bohdan M. Pavlyshenko SoftServe, Inc., Ivan Franko National University of Lviv * Correspondence: bpavl@softserveinc.com, b.pavlyshenko@gmail.com
More informationLecture 5: Minimum Cost Flows. Flows in a network may incur a cost, such as time, fuel and operating fee, on each link or node.
Lecture 5: Minimum Cost Flows Flows in a network may incur a cost, such as time, fuel and operating fee, on each link or node. Min Cost Flow Problem Examples Supply chain management deliver goods using
More informationCommunity Level Topic Diffusion
Community Level Topic Diffusion Zhiting Hu 1,3, Junjie Yao 2, Bin Cui 1, Eric Xing 1,3 1 Peking Univ., China 2 East China Normal Univ., China 3 Carnegie Mellon Univ. OUTLINE Background Model: COLD Diffusion
More informationMachine Learning Logistic Regression Hamid R. Rabiee Spring 2015
Machine Learning Logistic Regression Hamid R. Rabiee Spring 2015 http://ce.sharif.edu/courses/93-94/2/ce717-1 / Agenda Probabilistic Classification Introduction to Logistic regression Binary logistic regression
More informationAnalysing the Immune System with Fisher Features
Analysing the Immune System with John Department of Computer Science University College London WITMSE, Helsinki, September 2016 Experiment β chain CDR3 TCR repertoire sequenced from CD4 spleen cells. unimmunised
More informationRandom Forests. Parametrization, Tree Selection and Dynamic Induction
Random Forests Parametrization, Tree Selection and Dynamic Induction Simon Bernard Document and Learning research team LITIS lab. University of Rouen, France décembre 2014 Random Forest Classifiers Random
More informationUsing FPGAs to Accelerate Neural Network Inference
Using FPGAs to Accelerate Neural Network Inference 1 st FPL Workshop on Reconfigurable Computing for Deep Learning (RC4DL) 8. September 2017, Ghent, Belgium Associate Professor Magnus Jahre Department
More informationAdvanced Introduction to Machine Learning
Advanced Introduction to Machine Learning 10715, Fall 2014 Structured Sparsity, with application in Computational Genomics Eric Xing Lecture 3, September 15, 2014 Reading: Eric Xing @ CMU, 2014 1 Structured
More informationStochastic Convex Sparse Principal Component Analysis
Stochastic Convex Sparse Principal Component Analysis Inci M. Baytas 1, Kaixiang Lin 1, Fei Wang 2, Anil K. Jain 1 and Jiayu Zhou 1 1 Michigan State University 2 Cornell University 1 Principal Component
More informationInferring Gene-Gene Interactions and Functional Modules Beyond Standard Models
Inferring Gene-Gene Interactions and Functional Modules Beyond Standard Models Haiyan Huang Department of Statistics, UC Berkeley Feb 7, 2018 Background Background High dimensionality (p >> n) often results
More informationImproving the Accuracy of Base Calls and Error Predictions for GS 20 DNA Sequence Data
Improving the Accuracy of Base Calls and Error Predictions for GS 20 DNA Sequence Data Justin S. Hogg Department of Computational Biology University of Pittsburgh Pittsburgh, PA 15213 jsh32@pitt.edu Abstract
More information+? Mean +? No change -? Mean -? No Change. *? Mean *? Std *? Transformations & Data Cleaning. Transformations
Transformations Transformations & Data Cleaning Linear & non-linear transformations 2-kinds of Z-scores Identifying Outliers & Influential Cases Univariate Outlier Analyses -- trimming vs. Winsorizing
More informationBUSINESS DATA MINING (IDS 572) Please include the names of all team-members in your write up and in the name of the file.
BUSINESS DATA MINING (IDS 572) HOMEWORK 4 DUE DATE: TUESDAY, APRIL 10 AT 3:20 PM Please provide succinct answers to the questions below. You should submit an electronic pdf or word file in blackboard.
More informationInfluence diffusion dynamics and influence maximization in complex social networks. Wei Chen 陈卫 Microsoft Research Asia
Influence diffusion dynamics and influence maximization in complex social networks Wei Chen 陈卫 Microsoft Research Asia 1 (Social) networks are natural phenomena 2 Booming of online social networks 3 Opportunities
More informationData Mining. Chapter 7: Score Functions for Data Mining Algorithms. Fall Ming Li
Data Mining Chapter 7: Score Functions for Data Mining Algorithms Fall 2011 Ming Li Department of Computer Science and Technology Nanjing University The merit of score function Score function indicates
More informationby Xindong Wu, Kui Yu, Hao Wang, Wei Ding
Online Streaming Feature Selection by Xindong Wu, Kui Yu, Hao Wang, Wei Ding 1 Outline 1. Background and Motivation 2. Related Work 3. Notations and Definitions 4. Our Framework for Streaming Feature Selection
More informationNear-Balanced Incomplete Block Designs with An Application to Poster Competitions
Near-Balanced Incomplete Block Designs with An Application to Poster Competitions arxiv:1806.00034v1 [stat.ap] 31 May 2018 Xiaoyue Niu and James L. Rosenberger Department of Statistics, The Pennsylvania
More informationUnravelling Airbnb Predicting Price for New Listing
Unravelling Airbnb Predicting Price for New Listing Paridhi Choudhary H John Heinz III College Carnegie Mellon University Pittsburgh, PA 15213 paridhic@andrew.cmu.edu Aniket Jain H John Heinz III College
More informationExploitation and Exploration in a Performance based Contextual Advertising System
Exploitation and Exploration in a Performance based Contextual Advertising System Wei Li, Xuerui Wang, Ruofei Zhang, Ying Cui Jianchang Mao and Rong Jin Performance-based online advertising Performance
More informationSparsity-Constrained Transportation Problem
MSOM 204 Conference June 20, 204 Sparsity-Constrained Transportation Problem Annie I. Chen (anniecia@mit.edu) Stephen C. Graves (sgraves@mit.edu) Massachusetts Institute of Technology Inventory Positioning
More informationMetaGO: Predicting Gene Ontology of non-homologous proteins through low-resolution protein structure prediction and protein-protein network mapping
MetaGO: Predicting Gene Ontology of non-homologous proteins through low-resolution protein structure prediction and protein-protein network mapping Chengxin Zhang, Wei Zheng, Peter L Freddolino, and Yang
More informationNew Clustering-based Forecasting Method for Disaggregated End-consumer Electricity Load Using Smart Grid Data
New Clustering-based Forecasting Method for Disaggregated End-consumer Electricity Load Using Smart Grid Data Peter Laurinec, and Mária Lucká 4..7 Slovak University of Technology in Bratislava Motivation
More informationDallas J. Elgin, Ph.D. IMPAQ International Randi Walters, Ph.D. Casey Family Programs APPAM Fall Research Conference
Utilizing Predictive Modeling to Improve Policy through Improved Targeting of Agency Resources: A Case Study on Placement Instability among Foster Children Dallas J. Elgin, Ph.D. IMPAQ International Randi
More informationMulti armed Bandits on the Web
Multi armed Bandits on the Web Successes, Lessons and Challenges Lihong Li Microsoft Research 08/24/2014 2 nd Workshop on User Engagement Optimization (KDD 14) BIG DATA correlation Statistics, ML, DM,
More informationFrom Ordinal Ranking to Binary Classification
From Ordinal Ranking to Binary Classification Hsuan-Tien Lin Learning Systems Group, California Institute of Technology Talk at CS Department, National Tsing-Hua University March 27, 2008 Benefited from
More informationUniversity Question Paper Two Marks
University Question Paper Two Marks 1. List the application of Operations Research in functional areas of management. Answer: Finance, Budgeting and Investment Marketing Physical distribution Purchasing,
More informationUncovering the Small Community Structure in Large Networks: A Local Spectral Approach
Uncovering the Small Community Structure in Large Networks: A Local Spectral Approach Yixuan Li 1, Kun He 2, David Bindel 1 and John E. Hopcroft 1 1 Cornell University, USA 2 Huazhong University of Science
More informationScalable Mining of Social Data using Stochastic Gradient Fisher Scoring. Jeon-Hyung Kang and Kristina Lerman USC Information Sciences Institute
Scalable Mining of Social ata using Stochastic Gradient Fisher Scoring Jeon-Hyung Kang and Kristina Lerman USC Information Sciences Institute Information Overload in Social Media 2,500,000,000,000,000,000
More informationFrom Ordinal Ranking to Binary Classification
From Ordinal Ranking to Binary Classification Hsuan-Tien Lin Learning Systems Group, California Institute of Technology Talk at CS Department, National Chiao Tung University March 26, 2008 Benefited from
More informationAdaptive Design for Clinical Trials
Adaptive Design for Clinical Trials Mark Chang Millennium Pharmaceuticals, Inc., Cambridge, MA 02139,USA (e-mail: Mark.Chang@Statisticians.org) Abstract. Adaptive design is a trial design that allows modifications
More informationGene Prediction: Similarity-Based Approaches Spliced Alignment
Gene Prediction: Similarity-Based Approaches Spliced Alignment Gene Prediction introns S exon exon exon he predicted gene November 15 2 Outline of Agenda he idea of similarity-based approach to gene prediction
More informationSegmenting Customer Bases in Personalization Applications Using Direct Grouping and Micro-Targeting Approaches
Segmenting Customer Bases in Personalization Applications Using Direct Grouping and Micro-Targeting Approaches Alexander Tuzhilin Stern School of Business New York University (joint work with Tianyi Jiang)
More informationUtilizing Optimization Techniques to Enhance Cost and Schedule Risk Analysis
1 Utilizing Optimization Techniques to Enhance Cost and Schedule Risk Analysis Colin Smith, Brandon Herzog SCEA 2012 2 Table of Contents Introduction to Optimization Optimization and Uncertainty Analysis
More informationCovariance Matrices & All-pairs Similarity
Matrices & Similarity April 2015, Stanford DAO (Stanford) April 2014 1 / 34 Notation for matrix A Given m n matrix A, with m n. a 1,1 a 1,2 a 1,n a 2,1 a 2,2 a 2,n A =...... a m,1 a m,2 a m,n A is tall
More informationPricing for Customers with Probabilistic Valuations as a Continuous Knapsack Problem Michael Benisch James Andrews Norman Sadeh
Pricing for Customers with Probabilistic Valuations as a Continuous Knapsack Problem Michael Benisch James Andrews Norman Sadeh December 2005 CMU-ISRI-05-137 Institute for Software Research International
More informationRandom Walk Planning: Theory, Practice, and Application
Random Walk Planning: Theory, Practice, and Application Hootan Nakhost University of Alberta, Canada Google Canada since May 2013 May 9, 2012 Outline RW Planning Design Why does it work? Application Inefficient
More informationCombinational Collaborative Filtering: An Approach For Personalised, Contextually Relevant Product Recommendation Baskets
Combinational Collaborative Filtering: An Approach For Personalised, Contextually Relevant Product Recommendation Baskets Research Project - Jai Chopra (338852) Dr Wei Wang (Supervisor) Dr Yifang Sun (Assessor)
More informationMovie Success Prediction PROJECT REPORT. Rakesh Parappa U CS660
Movie Success Prediction PROJECT REPORT Rakesh Parappa U01382090 CS660 Abstract The report entails analyzing different variables like movie budget, actor s Facebook likes, director s Facebook likes and
More informationInfluence Maximization on Social Graphs. Yu-Ting Wen
Influence Maximization on Social Graphs Yu-Ting Wen 05-25-2018 Outline Background Models of influence Linear Threshold Independent Cascade Influence maximization problem Proof of performance bound Compute
More information2. Genetic Algorithms - An Overview
2. Genetic Algorithms - An Overview 2.1 GA Terminology Genetic Algorithms (GAs), which are adaptive methods used to solve search and optimization problems, are based on the genetic processes of biological
More informationCopyright 2013, SAS Institute Inc. All rights reserved.
IMPROVING PREDICTION OF CYBER ATTACKS USING ENSEMBLE MODELING June 17, 2014 82 nd MORSS Alexandria, VA Tom Donnelly, PhD Systems Engineer & Co-insurrectionist JMP Federal Government Team ABSTRACT Improving
More informationCS 769 Course Project Efficiently Maintaining Conditional Random Fields
CS 769 Course Project Efficiently Maintaining Conditional Random Fields Xixuan (Aaron) Feng, and Guangde Chen Summary. Real-world systems, such as iswiki, Alibaba, Citeseer, Kylin and YAGO, have a growing
More informationOur MCMC algorithm is based on approach adopted by Rutz and Trusov (2011) and Rutz et al. (2012).
1 ONLINE APPENDIX A MCMC Algorithm Our MCMC algorithm is based on approach adopted by Rutz and Trusov (2011) and Rutz et al. (2012). The model can be written in the hierarchical form: β X,ω,Δ,V,ε ε β,x,ω
More informationA logistic regression model for Semantic Web service matchmaking
. BRIEF REPORT. SCIENCE CHINA Information Sciences July 2012 Vol. 55 No. 7: 1715 1720 doi: 10.1007/s11432-012-4591-x A logistic regression model for Semantic Web service matchmaking WEI DengPing 1*, WANG
More informationGetting Started with Predictive Analytics
Getting Started with Predictive Analytics What is Predictive Analytics? Predictive analytics uses many techniques from data mining, statistics, modeling, machine learning, and artificial intelligence to
More informationSalford Predictive Modeler. Powerful machine learning software for developing predictive, descriptive, and analytical models.
Powerful machine learning software for developing predictive, descriptive, and analytical models. The Company Minitab helps companies and institutions to spot trends, solve problems and discover valuable
More informationCS3211 Project 2 OthelloX
CS3211 Project 2 OthelloX Contents SECTION I. TERMINOLOGY 2 SECTION II. EXPERIMENTAL METHODOLOGY 3 SECTION III. DISTRIBUTION METHOD 4 SECTION IV. GRANULARITY 6 SECTION V. JOB POOLING 8 SECTION VI. SPEEDUP
More informationOn of the major merits of the Flag Model is its potential for representation. There are three approaches to such a task: a qualitative, a
Regime Analysis Regime Analysis is a discrete multi-assessment method suitable to assess projects as well as policies. The strength of the Regime Analysis is that it is able to cope with binary, ordinal,
More informationModel Selection, Evaluation, Diagnosis
Model Selection, Evaluation, Diagnosis INFO-4604, Applied Machine Learning University of Colorado Boulder October 31 November 2, 2017 Prof. Michael Paul Today How do you estimate how well your classifier
More informationApplying Regression Techniques For Predictive Analytics Paviya George Chemparathy
Applying Regression Techniques For Predictive Analytics Paviya George Chemparathy AGENDA 1. Introduction 2. Use Cases 3. Popular Algorithms 4. Typical Approach 5. Case Study 2016 SAPIENT GLOBAL MARKETS
More informationDynamic Programming Algorithms
Dynamic Programming Algorithms Sequence alignments, scores, and significance Lucy Skrabanek ICB, WMC February 7, 212 Sequence alignment Compare two (or more) sequences to: Find regions of conservation
More informationCONNECTING CORPORATE GOVERNANCE TO COMPANIES PERFORMANCE BY ARTIFICIAL NEURAL NETWORKS
CONNECTING CORPORATE GOVERNANCE TO COMPANIES PERFORMANCE BY ARTIFICIAL NEURAL NETWORKS Darie MOLDOVAN, PhD * Mircea RUSU, PhD student ** Abstract The objective of this paper is to demonstrate the utility
More informationINDIAN INSTITUTE OF MATERIALS MANAGEMENT Post Graduate Diploma in Materials Management PAPER 18 C OPERATIONS RESEARCH.
INDIAN INSTITUTE OF MATERIALS MANAGEMENT Post Graduate Diploma in Materials Management PAPER 18 C OPERATIONS RESEARCH. Dec 2014 DATE: 20.12.2014 Max. Marks: 100 TIME: 2.00 p.m to 5.00 p.m. Duration: 03
More informationBetter Marketing Analytics Using Genetic Algorithms. Doug Newell Genalytics Inc. June 2005
Better Marketing Analytics Using Genetic Algorithms Doug Newell Genalytics Inc. June 2005 Today s Session Direct Marketing Today Predictive Modeling Techniques Building Models with Genetic Algorithms 2
More informationRanking Potential Customers based on GroupEnsemble method
Ranking Potential Customers based on GroupEnsemble method The ExceedTech Team South China University Of Technology 1. Background understanding Both of the products have been on the market for many years,
More informationThe Art of Lemon s solution
The Art of Lemon s solution KDD Cup 2011 Track 2 Siwei Lai/ Rui Diao Liang Xiang Outline Problem Introduction Data Analytics Algorithms Content 11.2175% Item CF 3.8222% BSVD+ 3.5362% NBSVD+ 3.8146% Main
More informationDatabase Searching and BLAST Dannie Durand
Computational Genomics and Molecular Biology, Fall 2013 1 Database Searching and BLAST Dannie Durand Tuesday, October 8th Review: Karlin-Altschul Statistics Recall that a Maximal Segment Pair (MSP) is
More informationAutomated Empirical Selection of Rule Induction Methods Based on Recursive Iteration of Resampling Methods
Automated Empirical Selection of Rule Induction Methods Based on Recursive Iteration of Resampling Methods Shusaku Tsumoto, Shoji Hirano, and Hidenao Abe Department of Medical Informatics, Faculty of Medicine,
More informationIncluding prior knowledge in shrinkage classifiers for genomic data
Including prior knowledge in shrinkage classifiers for genomic data Jean-Philippe Vert Jean-Philippe.Vert@mines-paristech.fr Mines ParisTech / Curie Institute / Inserm Statistical Genomics in Biomedical
More informationWorker Skill Estimation from Crowdsourced Mutual Assessments
Worker Skill Estimation from Crowdsourced Mutual Assessments Shuwei Qiang The George Washington University Amrinder Arora BizMerlin Current approaches for estimating skill levels of workforce either do
More informationPredicting prokaryotic incubation times from genomic features Maeva Fincker - Final report
Predicting prokaryotic incubation times from genomic features Maeva Fincker - mfincker@stanford.edu Final report Introduction We have barely scratched the surface when it comes to microbial diversity.
More informationA simulation approach for evaluating hedonic wage models ability to recover marginal values for risk reductions
A simulation approach for evaluating hedonic wage models ability to recover marginal values for risk reductions Xingyi S. Puckett PhD. Candidate Center for Environmental and Resource Economic Policy North
More informationUsing Decision Tree to predict repeat customers
Using Decision Tree to predict repeat customers Jia En Nicholette Li Jing Rong Lim Abstract We focus on using feature engineering and decision trees to perform classification and feature selection on the
More informationMeta-regression Analysis of Agricultural Productivity Studies
Meta-regression Analysis of Agricultural Productivity Studies Giannis Karagiannis Selected Paper prepared for presentation at the International Agricultural Trade Research Consortium s (IATRC s) 2013 Symposium:
More informationBOOTSTRAPPING AND CONFIDENCE INTERVALS
BOOTSTRAPPING AND CONFIDENCE INTERVALS Fill out your reading report on Learning Suite MPA 630: Data Science for Public Management November 8, 2018 PLAN FOR TODAY Why are we even doing this? Confidence
More informationIterated Conditional Modes for Cross-Hybridization Compensation in DNA Microarray Data
http://www.psi.toronto.edu Iterated Conditional Modes for Cross-Hybridization Compensation in DNA Microarray Data Jim C. Huang, Quaid D. Morris, Brendan J. Frey October 06, 2004 PSI TR 2004 031 Iterated
More informationBasic Linear Programming Concepts. Lecture 2 (3/29/2017)
Basic Linear Programming Concepts Lecture 2 (3/29/2017) Definition Linear Programming (LP) is a mathematical method to allocate scarce resources to competing activities in an optimal manner when the problem
More informationContributed to AIChE 2006 Annual Meeting, San Francisco, CA. With the rapid growth of biotechnology and the PAT (Process Analytical Technology)
Bio-reactor monitoring with multiway-pca and Model Based-PCA Yang Zhang and Thomas F. Edgar Department of Chemical Engineering University of Texas at Austin, TX 78712 Contributed to AIChE 2006 Annual Meeting,
More informationCSE 255 Lecture 3. Data Mining and Predictive Analytics. Supervised learning Classification
CSE 255 Lecture 3 Data Mining and Predictive Analytics Supervised learning Classification Last week Last week we started looking at supervised learning problems Last week We studied linear regression,
More informationFast Consensus Decoding over Translation Forests. John DeNero, David Chiang, and Kevin Knight {chiang,
Fast Consensus Decoding over Translation Forests John DeNero, David Chiang, and Kevin Knight denero@berkeley.edu, {chiang, knight}@isi.edu Overview Minimum Bayes risk (MBR) decoding tends to improve over
More informationApplication of an Improved Neural Network Algorithm in E- commerce Customer Satisfaction Evaluation
Application of an Improved Neural Network Algorithm in E- commerce Customer Satisfaction Evaluation Lei Yang 1,2 *, Ying Ma 3 1 Science and Technology Key Laboratory of Hebei Province, Baoding 071051,
More informationAccelerating Gene Set Enrichment Analysis on CUDA-Enabled GPUs. Bertil Schmidt Christian Hundt
Accelerating Gene Set Enrichment Analysis on CUDA-Enabled GPUs Bertil Schmidt Christian Hundt Contents Gene Set Enrichment Analysis (GSEA) Background Algorithmic details cudagsea Performance evaluation
More informationNew restaurants fail at a surprisingly
Predicting New Restaurant Success and Rating with Yelp Aileen Wang, William Zeng, Jessica Zhang Stanford University aileen15@stanford.edu, wizeng@stanford.edu, jzhang4@stanford.edu December 16, 2016 Abstract
More informationAirbnb Capstone: Super Host Analysis
Airbnb Capstone: Super Host Analysis Justin Malunay September 21, 2016 Abstract This report discusses the significance of Airbnb s Super Host Program. Based on Airbnb s open data, I was able to predict
More informationA Personalized Company Recommender System for Job Seekers Yixin Cai, Ruixi Lin, Yue Kang
A Personalized Company Recommender System for Job Seekers Yixin Cai, Ruixi Lin, Yue Kang Abstract Our team intends to develop a recommendation system for job seekers based on the information of current
More informationAdvanced Quantitative Research Methodology, Lecture Notes: Text Analysis: Supervised Learning
Advanced Quantitative Research Methodology, Lecture Notes: Text Analysis: Supervised Learning Gary King Institute for Quantitative Social Science Harvard University April 22, 2012 Gary King (Harvard, IQSS)
More informationAdvanced Metaheuristics. Daniele Vigo D.E.I. - Università di Bologna
Advanced Metaheuristics Daniele Vigo D.E.I. - Università di Bologna daniele.vigo@unibo.it Main families of Metaheuristics Single-solution methods Basic: Tabu Search, Simulated Annealing Advanced: Iterated
More informationThe Combined Model of Gray Theory and Neural Network which is based Matlab Software for Forecasting of Oil Product Demand
The Combined Model of Gray Theory and Neural Network which is based Matlab Software for Forecasting of Oil Product Demand Song Zhaozheng 1,Jiang Yanjun 2, Jiang Qingzhe 1 1State Key Laboratory of Heavy
More informationMulti-objective optimization
Multi-objective optimization Kevin Duh Bayes Reading Group Aug 5, 2011 The Problem Optimization of K objectives simultaneously: min x [F 1 (x), F 2 (x),..., F K (x)], s.t. x X (1) X = {x R n g j (x) 0,
More informationConvex and Non-Convex Classification of S&P 500 Stocks
Georgia Institute of Technology 4133 Advanced Optimization Convex and Non-Convex Classification of S&P 500 Stocks Matt Faulkner Chris Fu James Moriarty Masud Parvez Mario Wijaya coached by Dr. Guanghui
More informationAutomatic detection of plasmonic nanoparticles in tissue sections
Automatic detection of plasmonic nanoparticles in tissue sections Dor Shaviv and Orly Liba 1. Introduction Plasmonic nanoparticles, such as gold nanorods, are gaining popularity in both medical imaging
More informationIs Confirmatory PK/PD Modeling Possible?
Is Confirmatory PK/PD Modeling Possible? Chuanpu Hu, Ph.D. Director, Pharmacometrics Johnson & Johnson May 19, 2009 MBSW 2009 Outline PK/PD modeling why, how Exploratory vs. confirmatory: when to use which?
More informationRestaurant Recommendation for Facebook Users
Restaurant Recommendation for Facebook Users Qiaosha Han Vivian Lin Wenqing Dai Computer Science Computer Science Computer Science Stanford University Stanford University Stanford University qiaoshah@stanford.edu
More informationMonte-Carlo Tree Search
Introduction Selection and expansion References An introduction Jérémie DECOCK May 2012 Introduction Selection and expansion References Introduction 2 Introduction Selection and expansion References Introduction
More informationRobust Decentralized Task Assignment for Cooperative UAVs
AIAA Guidance, Navigation, and Control Conference and Exhibit 21-24 August 26, Keystone, Colorado AIAA 26-6454 Robust Decentralized Task Assignment for Cooperative UAVs Mehdi Alighanbari and Jonathan P.
More informationStock Price Prediction with Daily News
Stock Price Prediction with Daily News GU Jinshan MA Mingyu Derek MA Zhenyuan ZHOU Huakang 14110914D 14110562D 14111439D 15050698D 1 Contents 1. Work flow of the prediction tool 2. Model performance evaluation
More informationPREDICTION OF CONCRETE MIX COMPRESSIVE STRENGTH USING STATISTICAL LEARNING MODELS
Journal of Engineering Science and Technology Vol. 13, No. 7 (2018) 1916-1925 School of Engineering, Taylor s University PREDICTION OF CONCRETE MIX COMPRESSIVE STRENGTH USING STATISTICAL LEARNING MODELS
More informationFrom Ordinal Ranking to Binary Classification
From Ordinal Ranking to Binary Classification Hsuan-Tien Lin Learning Systems Group, California Institute of Technology Talk at Dept. of CSIE, National Taiwan University March 21, 2008 Benefited from joint
More informationLecture 6: Decision Tree, Random Forest, and Boosting
Lecture 6: Decision Tree, Random Forest, and Boosting Tuo Zhao Schools of ISyE and CSE, Georgia Tech CS764/ISYE/CSE 674: Machine Learning/Computational Data Analysis Tree? Tuo Zhao Lecture 6: Decision
More informationThroughout the Development
Managing Acceptance Criteria Throughout the Development Lifecycle Shea Watrin, Cecilia Chin, Julie TerWee Amgen Quality Outline Acceptance criteria primer Criteria on day 1 What necessitates change? Is
More informationEstimating individual tree growth with non-parametric. methods. Susanna Sironen U NIVERSITY OF JOENSUU FACULTY OF FORESTRY
Estimating individual tree growth with non-parametric methods Susanna Sironen Introduction to non-parametric methods Growth is predicted as a weighted average of the values of neighbouring observations.
More informationPredictive Modelling for Customer Targeting A Banking Example
Predictive Modelling for Customer Targeting A Banking Example Pedro Ecija Serrano 11 September 2017 Customer Targeting What is it? Why should I care? How do I do it? 11 September 2017 2 What Is Customer
More informationAnt Colony Optimization
Ant Colony Optimization Part 4: Algorithms Fall 2009 Instructor: Dr. Masoud Yaghini Ant Colony Optimization: Part 4 Outline The Traveling Salesman Problem ACO Algorithms for TSP Ant System (AS) Elitist
More information(& Classify Deaths Without Physicians) 1
Advanced Quantitative Research Methodology, Lecture Notes: Text Analysis I: How to Read 100 Million Blogs (& Classify Deaths Without Physicians) 1 Gary King http://gking.harvard.edu April 25, 2010 1 c
More informationModeling the Safety Effect of Access and Signal Density on. Suburban Arterials: Using Macro Level Analysis Method
Modeling the Safety Effect of Access and Signal Density on Suburban Arterials: Using Macro Level Analysis Method Yuan Jinghui 1,2 and Wang Xuesong 1,2 (1. The Key Laboratory of Road and Traffic Engineering,
More informationCrowe Critical Appraisal Tool (CCAT) User Guide
Crowe Critical Appraisal Tool (CCAT) User Guide Version 1.4 (19 November 2013) Use with the CCAT Form version 1.4 only Michael Crowe, PhD michael.crowe@my.jcu.edu.au This work is licensed under the Creative
More informationSupervised Learning Using Artificial Prediction Markets
Supervised Learning Using Artificial Prediction Markets Adrian Barbu Department of Statistics Florida State University Joint work with Nathan Lay, FSU Dept. of Scientific Computing 1 Main Contributions
More information