Pattern Recognition in Multiple Bike sharing Systems for comparability

Guided By,
Prof. Dr. Dirk C. Mattfeld,
M.Sc. Jan Brinkmann
Pattern Recognition in Multiple Bikesharing Systems for
Comparability
Presented By,
Athiq Ahamed

Vorlesungstitel | Semester | Kapitel x | Folie 2
© Prof. Dr. Dirk C. Mattfeld
Motivation for Bikesharing Systems (BSS)
• Increase in the proportion of people living in urban areas
• United Nations predicts by 2050, 86 % of the world will be urbanized
• Most important modes of transport are public and private
• Several problems in urban areas for transport
• Traditional urban transportation does not solve the problems

BSS being a Promising Nominee (Shared Mobility)
http://bike-sharing.blogspot.de/
• BSS is a sustainable short-term bicycle rental service
• Cost effective and flexible form of transportation (Customer, Subscriber)
• First BSS program was deployed on July 28, 1965
Benefits
• Health
• Use of free ride (money)
• Pollution control
• Having fun!!!!!!

Issues and Hypotheses
Issues
• Difficulties in operating and managing these BSS
• Predicting usage (open data)
• Availability of bikes and free racks
User satisfaction !!!!
Solution
• Planning the location properly and predicting bike usage (activity pattern)
• Prior redistribution of bikes
Hypotheses
1. Rentals and returns depend on the spatial and temporal factors.
2. Profile of users influences rentals and returns.
3. Profile of the user influences in business development in the neighborhood.

Knowledge Discovery in Databases
• Preprocessing (Data Cleansing)
• Data reduction / integration / transformation
• Data Mining (Knowledge Discovery)
• Clustering, Classification
• Visualization
• Geo BI
• Several other methods
• Post processing
• Decision support

Related Work
• Froehlich, Neumann and Oliver (2009) : Analyzed BSS usage to infer
human mobility patterns in the city (Barcelona).
• Kaltenbrunner et al (2010): Developed a short-term statistical prediction
model for station occupancy (Barcelona).
• Borgnat et al. (2010): Developed a statistical model for predicting the
occupancy of station (Lyon).
• Oliver O’Brien et al. (2013) : They analyzed 38 bikesharing systems,
widespread all around the world (Europe, Middle East, Asia, Australasia
and the Americas)
• Vogel et al. (2011) [69]: They presented complete analyses of operational
data from Vienna’s BSS (Vienna).

Problems with Existing Systems
• Existing system analyzed data with simple analysis techniques
• Geocoding with population, household, employment data was never taken
by any of the systems
• Most of their future work is comparability
• No in-depth comparability of systems is done
• None of the systems check comparability of two BSS with similar features
• Very fewer systems analyze cluster movements

Why Comparability ? Revisiting
• Comparability is crucial to get insights of activity patterns from multiple
systems
• For predicting or anticipating such a bike activity in future
• Designing a new or extending an existing BSS
• Location for a new system is planned properly
• Serves as an input for several applications
• As a result, bikes or free racks available all the time

Which Systems for Comparability ?
• Population
• Weather
• Household ratio
• Economical aspects
• Tourism

Which Systems for Comparability?
 Citi Bike NewYork and Capital Bike-Share Washington, D.C
 Open data from Citi Bike and Capital Bikeshare websites(2014)
Citi Bike NewYork Capital Bike-Share Washington
332 station ids with 340 station’s 356 stations
Annual (45 minutes free), 24
hour, 7(30 minutes free)
Annual, 30-day, 3 day,1 day( free
for 30 minutes)
80,81,216 trips approx per year
which reduced by 3% approx
after data cleansing
29,45,512 trips approx per year
which reduced by 3% approx
after data cleansing
Thousands of bikes, kiosks,
docking stations……
Thousands of bikes, kiosks
docking stations ……

Goal
• To identify patterns in BSS
• To prove the patterns are interesting using Data Mining and Geo BI
• With the hypotheses, one can prove that the patterns are interesting
• When the patterns are interesting hypotheses are proved
• When Hypotheses are proved, systems are comparable

Architecture Of NyDc
Clustering and
ClassificationVisualization
Tasks Tools
DB Postgres
Data
Cleansing
Postgres,
SAS
Clustering RapidMiner
Visualization Tableau
NyDC Tableau
Decision
Support

Overview of the Process
Duration Start
time
Stop
time
Start_id Stop_id Start
name
Stop
name
Start
longitude
Stop
longitude
Start
latitude
Stop
latitude
Bike_id User
type
Birth
year
Gender
Station ID Rental 0-1 ------ Rental 23-0 Returns 0-1 ------ Return 23-0
ID Start / Stop time Average rentals / returns
Station
ID
Rental 0-1 ------ Rental 23-0 Returns 0-1 ------ Return 23-0 Clusters

Data Cleansing (Selection / reduction / Intergration /
Transformation)
 For clustering meaningful attributes is necessary (cleaned)
 Only duration greater than 60 seconds are chosen
 Only summer months are chosen
 Data integrated from multiple data sources and average rentals and returns per
station per hour is calculated
 Input with 48 attributes and one ID after transformation (each hour as an attribute)
Weekday
Casual
Weekday
Subscriber
Weekday
Weekend
Subscriber
Weekend
Casual
Weekend
Data

Citi Bike Weekday

Capital Bikeshare Weekday

Citi Bike Subscriber Weekday

Capital Bikeshare Subscriber Weekday

Citi Bike Customer Weekday

Capital Bikeshare Customer Weekday

Data Mining (Knowledge Discovery)
Clustering
• Unsupervised learning, process of grouping common objects
• Data contains no labels
• Common objects are the ones which are similar (members or attributes)
• Idea is to find some structure/pattern in a collection of unlabeled data
• It is learning by observation, not with example (K-means and K-medoids)
• Goal, high intra-cluster similarity opposite for inter-cluster similarity
Areas
• Almost all the research fields
• Market research to medicines
• Image processing to spatial data

Data Mining (Knowledge Discovery)
Classification
• Classification is a supervised learning technique
• It’s a process of finding a model or function
• Distinguishes the data consisting of class labels.
• The given data is usually divided into training data (known class label) and
test data (unknown), (K-NN and Naive Bayes)
• Recall : It is the measure of completeness ---- TP/(TP + FN)
• Precision : It is the measure of exactness ---- TP/(TP + FP)
• Accuracy: The percentage of test set tuples that are correctly classified
• by the classifier.
Class “A” Class “Not A”
Test says “A” True Positive False Positive
Test says “Not A” False Negative True Negative

K-means: Clustering Algorithm
• A simple clustering algorithm for high intra-cluster similarity and opposite for
inter-cluster similarity
Working
1) It begins by randomly selecting k data points (initial centroids)
2) Creates k empty clusters.
3) It then assign’s exactly one centroid to each cluster.
4) After assigning, it iterates over all instances. It then assigns each data point
to one cluster with the nearest centroid (mean).
5) After each iteration, it computes cluster centroids based on the new data
points.
6) It checks if clustering is good enough (until no change) or it returns to (2).

Complicated Questions
How many clusters ???
• Davies–Bouldin index (DBI)
• Accuracy using Classification
• Experience
Why K-means ?
• Davies–Bouldin index (DBI) shows a low value
• High accuracy, precision, and recall using classification algorithms
Pseudo code for NyDc
• Run clustering algorithms
• Get accuracy using classification algorithms (choose the best one)
• Evaluate using Davies-Bouldin Index
• Use Geo BI to validate the analysis or proving the hypotheses

Clustering Accuracy Evaluation
Recall
Cluster 0 Cluster 1 Cluster 2 Cluster 3
K-means 83.33 87.5 99.17 95.08
K-medoids 93.17 85.07 84.62 80
EM 87.6 85.71 97.85 85.71
Precision
Cluster 0 Cluster 1 Cluster 2 Cluster 3
K-means 94.59 100 96.77 95.08
K-medoids 92.5 93.44 82.5 75.36
EM 94.64 85.71 86.67 92.31
Accuracy
Naive
Bayes
K-NN
K-means 91.46 96.32
K-medoids 87.33 87.92
EM 91.83 89.56

Clustering Validation
For understanding it clearly these clusters are named
• Commuter cluster (active day rental and return)
• Tourist or mix cluster (late afternoon and evening)
• Leisure cluster and utility cluster (active night and early morning)
• Residential or outer city cluster (low activity all time)
Proof for hypothesis one
• Sub-hypothesis 1: Temporal factors: time of the day plays an important
role
• Sub-hypothesis 2: Spatial factors: Location plays an important role

Sub-hypothesis 1- Temporal Validation (Citi Bike)

Sub-hypothesis 1- Temporal Validation (Capital Bikeshare)

Examples for Validation
Commuter
519 - Grand central terminal (railroad terminal)
Dupont station- Dupont circle
Tourist
2006 -Central park
Smithsonian - National mall Washington
Leisure
293- Lafayette Street
U St and 13 St NW- U Street.
Residential
Brooklyn
Arlington county

Sub-hypothesis 2: Spatial validation
Citi Bike Weekday

Sub-hypothesis 2: Spatial validation
Capital Bikeshare Weekday

Proof for Hypothesis 2
• If he or she is a subscriber they are regular commuters
• Subscribers are educated and rich
• Similar morning pickup and return evening pattern (workers) validated in
white color job map
• 41 % of the subscribers are master degree holders and 63 % are under 35
• Subscribers spend more money on BSS than the customers

• Nine in ten survey respondents were employed
• USA census reports shows that only about seven in ten adults in
Washington, D.C., are employed
• Customers are tourists or shoppers visiting neighborhood
• Customers show a less average activity in weekdays and more in the
weekends.
• Late pick ups and active afternoons proves them to be tourists or
shoppers or for household activities

• The riders visit spending locations more frequently
• High average activity in weekends proves to be tourists or leisure user.
• Cyclist visits a supermarket 3.2 times per week
• Motorist visit 2.5 times and spends more money
• If there are more customers or tourist then it’s likely to be better business
in the neighborhood
• If there are more subscriber’s its likely to have a high usage of bikes

NyDc
• Since hypotheses are proved, patterns are interesting
• Since interesting patterns are similar they are comparable
• Final comparability model (Nydc) can be used for various applications
• Could serve as a benchmark for comparability study
• Several useful features
• Prediction model prototype was developed using NyDc
• This model can be mapped to the new location
Population Location Temporal Hot spot Household ... Cluster
average
Location decisions

Conclusion
• In-depth analysis is done by separating data
• Bike activity patterns are obtained for future prediction
• Hypotheses ( patterns are interesting ) are proved for solving the issues of
BSS
• Yes, the two system’s are comparable to a greater extent
• Can be mapped to other cities BSS design one with similar attributes or
prediction
• Business development, city dynamics or providing location based services.

Future Work
• Developing an analogy based bike sharing information system
• Using NyDc to develop a new Artificial intelligence recommendation
system
• Developing better algorithm for predictions (taking more features)
• Personalized data capturing for recommender system (automatic path
calculation with temporal information)
• NyDc comparison to other city somewhere in Asia.
• Human dynamics in different parts of the world.

Pattern Recognition in Multiple Bike sharing Systems for comparability

More Related Content

Viewers also liked (12)

Similar to Pattern Recognition in Multiple Bike sharing Systems for comparability (20)

Recently uploaded (20)

Pattern Recognition in Multiple Bike sharing Systems for comparability