多離散連續（MDC）選擇分析的深度神經網絡和結構計量經濟學模型的比較分析.pdf

1、COLLABORATE.INNOVATE.EDUCATE.A Comparative Analysis of Deep Neural Network and Structural Econometric Models for Multiple Discrete-Continuous(MDC)Choice Analysis1XYZ Xth,2019June,2023Aupal MondalChandra R.BhatCOLLABORATE.INNOVATE.EDUCATE.2IntroductionThere is an emerging trend of using machine learn

2、ing models,to analyze individual decisions,in consumer behavior and transportation-related studies.These machine learning(ML)models are consistently found to achieve much higher predictive performance compared to the traditional discrete choice models(DCM).Deep Neural Network(DNN)and Random Forest(R

3、F)algorithms,in particular,are found to regularly outperform any other ML models and DCMs in predictions.COLLABORATE.INNOVATE.EDUCATE.PredictionsInterpretability3Introduction There is a constant debate about the computation perspectives and prediction accuracy offered by ML methods and the interpret

4、ability and behavioral foundations ingrained in theory-driven choice models.Theory-driven choice models based on economic and domain-specific theory,emphasize interpretability,make explicit assumptions,can work with smaller datasetsMachine Learning models “black-box”models,prioritize predictive accu

5、racy,can learn complex patterns,generally require large datasets.COLLABORATE.INNOVATE.EDUCATE.4IntroductionA few studies have attempted to address the interpretability issue in ML models,while a few have focused on utilizing a“synergistic”approach to harness the interpretability of theory-driven cho

6、ice models and the predictive accuracy of ML-based models.However,most earlier explorations compare the ML-based methods to the simple MNL model as the“strawman”.In any case,comparative studies on ML models versus theory-driven choice models have primarily focused on single discrete choice models.Cu

7、rrent study:Provides a systematic comparative evaluation of machine learning methods and structural econometric models in the context of multiple discrete-continuous choice situations.COLLABORATE.INNOVATE.EDUCATE.5Multiple discrete-continuous choice situation5Alternative 1Alternative 1Alternative 2A

8、lternative 2Alternative 3Alternative 3Alternative 4Alternative 4Several consumer choice situations are characterized by the choice of multiple alternatives(or goods)at the same time.This involves whether to choose an alternative decision(discrete component)and(if chosen),to what extent to consume an

9、 alternative decision(continuous component).Such a situation can be termed as a“multiple discrete-continuous”(MDC)choice scenario.COLLABORATE.INNOVATE.EDUCATE.6 Examples of MDC-type choice situations:Recreational Activity participation(time-use)Vehicle type holdings and use(Vehicle Miles of Travel)M

10、onthly expenses or consumption in a specific category Multiple discrete-continuous choice situationCOLLABORATE.INNOVATE.EDUCATE.71+=3=,1 0 is quasi-concave,increasing,and continuously differentiable utility function is the consumption quantity(1is an essential outside good)kand are parameters associ

11、ated with good,is the budget and is the unit price,1=1112()lnln1KkkkkkxUx=+x()kkk+=zexpwhere,is parameterized as The following is a typical functional form for the utility function that is maximized by a consumer in an MDC situationMultiple discrete-continuous choice situationCOLLABORATE.INNOVATE.ED

12、UCATE.8 The MDC extreme value(MDCEV)model and its variants are widely used to model such multiple discrete-continuous”situations(Bhat,2008).Multiple discrete-continuous choice situationSeveral advanced variants have been proposed in the literature-relaxing the IID assumption across alternatives(Gene

13、ralized Extreme Value,Nested structure,Mixed modeling).Probit-based kernel(assuming the error terms to be distributed multivariate normal)leads to the MDC-Probit or MDCP model.-easier to incorporate random coefficients and covariance across alternatives.-recent analytical approximation methods make

14、estimation efficient.COLLABORATE.INNOVATE.EDUCATE.9The Finite Discrete Mixture of Normals(FDMN)-MDCPAllows a comprehensive discrete and continuous stochastic structure to accommodate unobserved heterogeneity.Incorporates a hybrid approach that combines a continuous response surface for the parameter

15、 coefficients(the continuous component)with a latent segmentation approach(the discrete component).Ignoring the continuous component of the mixing(i.e.,only considering a latent segmentation model)or ignoring the discrete component of the mixing(i.e.,only considering a random parameter model)leads t

16、o substantial bias in the parameter estimates.COLLABORATE.INNOVATE.EDUCATE.10The Finite Discrete Mixture of Normals(FDMN)-MDCPUnivariate example of the FDMN approach(considering one variable)COLLABORATE.INNOVATE.EDUCATE.11The Finite Discrete Mixture of Normals(FDMN)-MDCPThe FDMN semi-parametric appr

17、oach Advantages over non-parametric methods:-Non-parametric methods generally assign equal probabilities to latent segment membership.-Consistency(for non-parametric methods)is achieved only in very large samples,parameter estimates have high variance,and the computational complexity/effort can be s

18、ubstantial.-Non-parametric methods are relatively profligate in parameters.COLLABORATE.INNOVATE.EDUCATE.12Input DataOutput LayerOutputLayer 1 .Layer NHidden LayersNeuronDeep Neural Network(DNN)ApproachCOLLABORATE.INNOVATE.EDUCATE.13.Deep Neural Network(DNN)Approach-MDCDiscrete outputsContinuous outp

19、utsMultiple discrete-continuous outputsInput dataInput dataClassification blockRegression blockCOLLABORATE.INNOVATE.EDUCATE.14Empirical contextFocus:Monthly recurring household recreation and entertainment-related subscription and membership expenditures(REX).Why focus on REX?Closely linked with phy

20、sical and mental well-being(health-care focus)Constitute a significant fraction of consumer retail expendituresCOLLABORATE.INNOVATE.EDUCATE.15Empirical contextData:The Consumer Expenditure(CEX)Survey administered by the US Census Bureau of Labor Statistics The Interview Survey(a component of the CEX

21、 survey)collects data on quarterly expenditures on larger-cost items and all recurring expenditures.Use all quarters data for 2018 and 2019 years(pre-COVID).Sample size:About 7481 Consumer Units(CU)or households.Sociodemographic information extracted for each CU.Annual expenses(summed over four quar

22、ters)used for each CU in our analysis.COLLABORATE.INNOVATE.EDUCATE.16Empirical context REX categories:Six MDC alternatives pertaining to monthly(or annually)recurring recreational expenses.1.Sports or health club membership2.Video streaming subscriptions3.Audio/music streaming subscriptions4.Digital

23、 books/newspaper subscription5.Online apps/gaming subscription6.Seasonal tickets(amusement parks/sports/plays)COLLABORATE.INNOVATE.EDUCATE.17Estimations/Results are in the worksFuture works include:Estimation and prediction using the structural FDMN-MDCP model and DNN-MDC modelCompare sociodemograph

24、ic variable effects from the results of the FDMN-MDCP and the DNN-MDC modelsDiscuss elasticity/average treatment effects(ATE)Policy implications to be identified for:healthcare professionals in terms of encouraging the public to adopt a more active recreational lifestyle.advertising agents in terms

25、of strategies to effectively promote recreational“products”.COLLABORATE.INNOVATE.EDUCATE.18Summary Ongoing debate about the predictive accuracy of ML models and the interpretability of theory-driven models.Earlier studies use the simple MNL model for comparative evaluation.Even then,research has onl

26、y focused on single discrete choice models.We analyze the use of ML-based models and structural models in multiple discrete-continuous choice situations.Future work:Estimation and comparative evaluation of FDMN-MDCP and DNN-MDC models.Empirical demonstration Recurring monthly/annual recreational expenses COLLABORATE.INNOVATE.EDUCATE.Thank you!