data-driven mobility services - uni-hildesheim.de · (xbg), feed forward neural network (fnn). i...
TRANSCRIPT
Data-driven Mobility Services
Data-driven Mobility ServicesVolkswagen Financial Services and University of Hildesheim
Lars Schmidt-Thieme
Information Systems and Machine Learning Lab (ISMLL)Institute for Computer Science, University of Hildesheim, Germany
November 23, 2018
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 1 / 25
Data-driven Mobility Services
Outline
1. Data-driven Mobility Services
2. Predictive Parking (WP1)
3. Automatic Damage Assessment for Cars (WP2)
4. Residual Value Forecast for Cars (WP3)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 2 / 25
Data-driven Mobility Services 1. Data-driven Mobility Services
Outline
1. Data-driven Mobility Services
2. Predictive Parking (WP1)
3. Automatic Damage Assessment for Cars (WP2)
4. Residual Value Forecast for Cars (WP3)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 2 / 25
Data-driven Mobility Services 1. Data-driven Mobility Services
Data-driven Mobility Services
I cooperative research project betweenI Volkswagen Financial Services (VWFS) andI ISMLL, University of Hildesheim
I 4/2018 – 3/2021
1. Data Analytics for banking services
2. Enhance customer experience through intelligent servicesI Predictive Parking (WP1)I Automatic Damage Assessment for Cars (WP2)
3. Forecasting the car leasing and financing risk assessmentI Residual Value Forecast for Cars (WP3)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 2 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Outline
1. Data-driven Mobility Services
2. Predictive Parking (WP1)
3. Automatic Damage Assessment for Cars (WP2)
4. Residual Value Forecast for Cars (WP3)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 3 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
The Parking ProblemAn increasingly difficult task
I Studies show around 25 billion car trips annuallyI Every car spends on average 162 hours a week being parkedI Governed by multiple factors
(a) Weatherhttps://bit.ly/2OXCZLA
(b)Destinationhttps://bit.ly/2R1YBs7
(c) Traffichttps://bit.ly/2AbU5Aa
The impactI congestionI air pollutionI psychological problems (stress, anxiety, etc.)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 3 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Predictive Parking: Problem definition
I Given a desired target parking location and time point,predict if it will be free.
I individual places vs. smaller lots of parking placesI occupancy rate: percent occupied places in a lotI crowded threshold: occupancy rate when considered crowdedI forecasting horizon: 5 mins, 15 mins, 30 mins . . . ahead
I Given a desired target parking location and time point,recommend a free parking place as close as possible.
I Parking availability: show which parking places are free nowI usually measured by sensors
I Pay and/or reserve a parking place
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 4 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Related Projects
I Infrastructure for Connected and Automated Road Transport(ICT4CART)
I IBM, Fiat, BMW, Nokia, Bosch, T-Mobile Austria and othersI 20M EUR EC project, 9/2018–8/2021
I Companies and start-ups
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 5 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Predictive Parking: Problem Definition
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 6 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Predictive Parking: Prediction Model Design
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 7 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Modelling with Convolutional Neural Network
Model Architecture
I 4 x Convolutional (Feature Map of 64)
I 2 x Fully Connected Layers (512 and 256 Hidden Units)
Model Hyperparameters
I History Length
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 8 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Evaluation of Precision to find the best prediction result
Precision
I Ability to find an empty location that is indeed empty
I Higher Precision leads to higher profit $$
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 9 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Evaluation of Recall to find the best prediction result
Recall
I Ability to find an occupied location that is indeed occupied
I Lower Recall leads to unsatisfied customers !!
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 10 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Data AugmentationI Additional data:
I Location identifiers (Model HL)I Statistics of the target location (Model HS1)
I sliding mean, minimum, maximum etc.
I mappend to the latent feature representation ZL−1
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 11 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Spatial-Temporal Prediction
I model spatial and temporal relations simultaneously
I correlations among regions sharing similar temporal patterns.
I Data reshaped into sequence of map grids over the time.
I Each pixel in the grid is defined by longitude and latitude values.
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 12 / 25
Data-driven Mobility Services 2. Predictive Parking (WP1)
Spatial-Temporal Prediction
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 13 / 25
Data-driven Mobility Services 3. Automatic Damage Assessment for Cars (WP2)
Outline
1. Data-driven Mobility Services
2. Predictive Parking (WP1)
3. Automatic Damage Assessment for Cars (WP2)
4. Residual Value Forecast for Cars (WP3)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 14 / 25
Data-driven Mobility Services 3. Automatic Damage Assessment for Cars (WP2)
Automatic Damage Assessment for Cars: Problemdefinition
Goal of the Use Case
I Idea is to streamline the lease-end process.
I Automate the damage identification and classification
I Reduce the variation in the cost of repair assessment
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 14 / 25
Data-driven Mobility Services 3. Automatic Damage Assessment for Cars (WP2)
Automatic Damage Assessment for Cars: MachineLearning Problem
Deep Learning Model
(Object Detection &Regression)
{roof, red, 1Z, etc}
Data Ouput
damage & boundingboxes
repair cost
I Given a set of Images and meta features with associated repair cost
I Identify the damage area and classify the damage type and action
I Estimate the cost of repair for the damage
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 15 / 25
Data-driven Mobility Services 3. Automatic Damage Assessment for Cars (WP2)
Dataset OverviewTable: Records Statistics
Name Count
Total Reports 39,000Total Images 1,613,333Damage Images 565,716Damage Images with Cost 342,029
Name Count
Image Types Abbildung, Beschadigung,GebrauchsspurNachlackierung, unfall, Fehlteil, Vorschaden
Body Parts - Level 1 61Body Parts - Level 2 1455Damages 43Actions 24
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 16 / 25
Data-driven Mobility Services 3. Automatic Damage Assessment for Cars (WP2)
Damage Classification (Cropped Images)
0 200 400 600 800 1000Iterations
0.5
0.6
0.7
0.8
0.9
1.0
Accu
racy
Classification accuracy
Mobilenet_0.5Inception_v3Mobilenet_0.75
Test AccuracyI Our model: 93.5%I Major AI Consultancy: 85%
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 17 / 25
Data-driven Mobility Services 3. Automatic Damage Assessment for Cars (WP2)
Object Detection Results
0 25000 50000 75000 100000 125000 150000 175000 200000iterations
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
mAP
./RESNET/faster_RCNN_data1
./RESNET/faster_RCNN_data2
./RESNET/faster_RCNN_data3
./INCEPTION/faster_RCNN_data1
./INCEPTION/faster_RCNN_data2
./INCEPTION/faster_RCNN_data3
I data1 : First iteration of image annotationI data2 : Second iteration of image annotationI data3 : Third iteration of image annotation
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 18 / 25
Data-driven Mobility Services 3. Automatic Damage Assessment for Cars (WP2)
Cost Regression: Improved results with additional featuresI Models: Linear Regression (LR), Random Forest (RF), XGBoost
(XBG), Feed forward Neural Network (FNN).I Features: Color, MBV code, body part, damage, action
FNN XGB RF LR Average
100
120
140
160
180
200
RMSE
(Cos
t)
Average RMSE Score across 3-folds
(p) results with mbv withoutgrouping
FNN XGB RF LR Average
80
100
120
140
160
180
200
RMSE
(Cos
t)
Average RMSE Score across 3-folds
(q) results with mbv with grouping
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 19 / 25
Data-driven Mobility Services 4. Residual Value Forecast for Cars (WP3)
Outline
1. Data-driven Mobility Services
2. Predictive Parking (WP1)
3. Automatic Damage Assessment for Cars (WP2)
4. Residual Value Forecast for Cars (WP3)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 20 / 25
Data-driven Mobility Services 4. Residual Value Forecast for Cars (WP3)
Residual Value Forecast for Cars: Problem Definitation
I Given the car’s configuration settings, list price and contract startdate we want to estimate the car’s residual value at a future date Xwith estimated mileage per-year Y
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 20 / 25
Data-driven Mobility Services 4. Residual Value Forecast for Cars (WP3)
Residual Value Forecast for Cars: Machine LearningProblem
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 21 / 25
Data-driven Mobility Services 4. Residual Value Forecast for Cars (WP3)
Residual Value Forecast for Cars: Dataset
I DatasetI 650K Instances (From 2010 to 2018)I 34 Total different vehicle models
I Chronological Data SplitsI Train : Registration≥2008 , Sale≤2010 — Test: Registration=2011I Train : Registration≥2008 , Sale≤2011 — Test: Registration=2012I ..I Train : Registration≥2008 , Sale≤2016 — Test: Registration=2017
I ModelsI Random ForestI Gradient Boosted TreesI VWFS RV baseline MethodI Average Model (Naive Model)
I Evaluation metricI Mean Absolute Error (MAE)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 22 / 25
Data-driven Mobility Services 4. Residual Value Forecast for Cars (WP3)
Residual Value Forecast for Cars: Data Splits
2008 2009 2010 2011 2012 2013 2014 2015 2016 2017Contract Date
Train TestExperiment 1
Train TestExperiment 2
Train TestExperiment 7
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 23 / 25
Data-driven Mobility Services 4. Residual Value Forecast for Cars (WP3)
Residual Value Forecast for Cars: Results
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 24 / 25
Data-driven Mobility Services
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 25 / 25
Data-driven Mobility Services
Comparison of the model to off-the-shelf benchmarks
RECALL compared to Gradient-boosted decision trees (Benchmark 1)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 26 / 25
Data-driven Mobility Services
Comparison of the model to off-the-shelf benchmarks
PRECISION compared to an LSTM with 128 units (Benchmark 2)
Lars Schmidt-Thieme, University of Hildesheim, Germany
November 23, 2018 27 / 25