Mode Detection, Age Pattern Transition, and its Consequences on Carbon Emissions Team 5 Deepank Verma (D3) Neenu Thomas (D2) Omkar Deepak Karmarkar (D1) INDIAN INSTITUTE OF TECHNOLOGY BOMBAY, INDIA
BACKGROUND: MODE CHOICE DETERMINANTS Trip start time Gender Purpose Lesser number of Private mode Walk and bicycle more preferred Public modes for majority of work users during the working hours. mode by female. trips Walk and bicycle least preferred Public and private modes Private modes majorly for leisure for late night trips. more preferred mode by male . trips. . Travel time Trip end time Age 39.17% & 25.47% of short trips Private mode for late night trips Most preferred mode (less than 30 minutes travel) are made on foot & by Car respectively. Walk and bicycle least preferred Children aged less than 15 years - Walk Working aged people – Public transport for late night trips. Above 60% of long trips are made Elderly - Private mode on train.
Data Analysis • PT Data Objective 1 : Comparing Mode Detection using • Total entries : 1461514 Traditional and Machine Learning techniques. • After cleaning : 1441493. • Data random split: 85:15, Training set entries: 1225269, Test set entries: Predictor Variables 216224 Purpose (5 categories) • PP Data Gender (2 categories) Age (17 categories) • Total entries : 1522 Start Time • After cleaning : 1502 End Time • Data random split: 70:30, Training set Travel Time entries: 1051, Test set entries: 451 Target Variable • Target Variable (same) Mode (4 categories) • Predictor Variables (same)
MULTINOMIAL LOGIT MODEL (MNL MODEL) Testing Set: 216224 nos. Total • The most widely used mathematical form for choice Predicted samples probabilities in behavioural travel demand analyses 72885 Key Strengths: • The MNL model is simple to perform • Computationally efficient 63074 • Permits a simple behavioural interpretation of its Actual parameters Key Weaknesses: 31675 • Independence of Irrelevant Alternatives (IIA) property • No correlation between error terms (i.i.d. errors) • Random taste variation can not be represented, 48590 Log-Likelihood: -1464800 McFadden R 2 : 0.24345 Likelihood ratio: 942710
Testing Set: 216224 nos. Total Random Forests (RF) Age_50-54 Predicted samples Age_85 more Age_70-74 Age_65-69 72885 • A decision tree is a decision support tool that uses a Age_35-39 tree-like graph or model of decisions. Age_80-84 Age_25-29 • RFs train each tree independently, using a random 63074 Age_45-49 Actual sample of the data. Age_55-59 Age_60-64 31675 Age_30-34 Key Strengths: Age_75-79 • RF is much easier to tune. Only two hyperparameters (a) Age_40-44 depth of trees (6) and (b) number of estimators (25). 48590 Age_20-24 • More robust than a single decision tree, and less likely to Purpose_Leisure Purpose_Shopping over fit on the training data. Purpose_Ret. Home • Does not require preparation of the input data. Purpose_Others • Works with unscaled data and missing values. Age_15-19 Male • Provides information on Feature importance. Female Purpose_Work Key Weaknesses: Age_5-9 Age_10-14 • For data including categorical variables with different StartTimeHr RF number of levels, RFs are biased against attributes with EndTimeHr more levels. TravelTimeMin 0 0.1 0.2 0.3 0.4 0.5 0.6
Testing Set: 216224 nos. Total Extreme Gradient Boosting (XGB) Female Predicted samples Age_50-54 Age_55-59 Age_45-49 72885 • XGB build trees one at a time, where each new tree Age_30-34 helps to correct errors made by previously trained Age_40-44 Age_60-64 tree. 63074 Age_85 more Actual Age_35-39 Key Strength: Age_25-29 • It performs the optimization which makes the use of Age_80-84 31675 Age_65-69 custom loss functions much easier. Age_20-24 • Boosting focuses on unbalanced datasets by Age_75-79 48590 strengthening the impact of the positive class. Age_70-74 Age_15-19 • Provides information on Feature importance. Age_10-14 Purpose_Others Key Weaknesses: Purpose_Shopping Age_5-9 • Training generally takes longer because of the fact Purpose_Leisure that trees are built sequentially. Purpose_Ret.… • XGB is harder to tune than RF. Three parameters to Purpose_Work Male tune: (a) number of estimators (100), (b) depth of trees StartTimeHr (6) and (c) Learning rate (0.1) XGB EndTimeHr TravelTimeMin 0 0.05 0.1 0.15 0.2 0.25
Total Testing Set: 216224 nos. Artificial Neural Network (ANN) Predicted samples 72885 • ANN is based on a collection of connected units (nodes) called artificial neurons, which loosely model the neurons in a biological brain. 63074 Actual Key Strengths: 31675 • Ability to learn and model non-linear and complex relationships. • Efficiently processes large amount of training samples. 48590 Key Weakness: • Hardware dependence • Unexplained behavior of the Model. • Comparatively sub-standard performance in smaller datasets. • Large number of hyperparameters to tune. (a) No. of Hidden Layers (3), (b) Activation functions(Relu, Relu, Softmax), (c) Loss functions (Cross- entropy), (d) Dropout rate (0.8), (e) Optimizer (Adam), (f) Learning rate (0.01).
Evaluation Metrics PT data F1-score • Overall Accuracy: The ratio of a total number of correctly identified Classes MNL RF XGB ANN pixels to the total number of considered pixels. Public 0.77 0.79 0.79 0.79 Private 0.46 0.5 0.52 0.51 • The overall accuracy metrics is influenced by unbalanced and 0.34 Bicycle 0.03 0.02 0.3 prominent classes. Walk 0.52 0.51 0.58 0.59 • F1-score: is a harmonic mean of Precision and Recall. Ov. Acc. 54.72 57.03 60.71 60.44 Precision is the proportion of positive detections of the • Kappa 0.368 0.388 0.451 0.45 classifier which were actually correct, PP data • Recall refers to the proportion of actual positives which were F1-score detected correctly. Classes MNL RF XGB ANN Public 0.66 0.76 0.72 0.69 • Kappa index of Agreement is used in assessing the performance of 0.66 Private 0.53 0.63 0.55 different models. Bicycle 0.46 0.52 0.57 0.52 Walk 0.49 0.49 0.51 0.48 • The Kappa value (k) of model classifier suggests that the classifier is k*100 percent better than random assignment of classes. Ov. Acc. 56.09 64.52 64.75 58.76 Kappa 0.379 0.491 0.499 0.406
Potential Use Cases Real-time congestion estimation/pricing : By the use of mode detection models trained on telemetry data from various smartphone applications. Advertising : Push Notifications influencing future mode choice behavior by cab aggregators based on trained model. Health: Evaluation of exposure to pollution by estimating the choice of mode.
Objective 2 : Analyzing the Mode Pattern considering the Shifting Age Composition 10
AGE PATTERN AND MODE CHOICE OBSERVATIONS Mode use share 100% Higher share of on-road motorised modes by elderly people 90% 80% 70% Increasing traffic volume of on-road vehicles 60% Increasing carbon share by elderly 50% Comfort and convenience of elderly people 40% 30% 20% 10% Measures to reduce carbon share by transportation 0% Measures to ease the travel for elderly 0-14 15-64 Above 65 Rail Bus Car Bicycle Walk
MODE USE AND CHANGING EMISSION LEVEL Age wise share of different modes Carbon Emission by Transport Modes 100% 90% 80% 70% Hayashiya, H. Urban Rail Transit (2017) 3: 183. https://doi.org/10.1007/s40864-017-0070-4 60% 50% 40% 30% 20% Source: Hayashiya, H. Urban Rail Transit (2017) 3: 183. https://doi.org/10.1007/s40864-017-0070-4 10% Age-wise Share of Carbon Emission through Transport Modes 0-14 0% 2010 5.7 67.7 26.6 Rail Bus Car Bicycle Walk 15-64 0-14 15-64 Above 65 Above 65 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
CARBON EMISSION THROUGH TRANSPORT MODES Scenario 1 Changing Age Pattern a. Changing Population trend 100 b. Same Mode pattern 90 17.4 c. Same Mode based emission per PKM 22.5 27.8 29.6 33.2 35.7 80 Scenario 2 70 a. Changing Population trend b. Same Mode pattern 60 c. Changing Mode based emission per PKM 50 68.1 (Reference: Decadal average) 64.1 60 59.2 40 55.8 53.6 Scenario 3 30 a. Changing Population trend 20 b. Same Mode pattern 10 c. Changing emission per PKM in car while 14.6 12.2 10.8 13.4 11.3 11 emission per PKM in other mode are constant 0 2000 2010 2020 2030 2040 2050 Scenario 4 Age Group a. Changing Population trend b. Shifting Car users to Public transport (Bus) 0-14 15-64 Above 65 c. Changing Mode based Emission per PKM Source: National Institute of Population and Social Security Research
Recommend
More recommend
