![ffmulksin
This is a presentation of my capstone project for the IBM
Data Science Professional Certificate via Coursera.[1]
IBM DATA SCIENCE PROFESSIONAL
The code can be found in form of a Jupyter notebook on
my Github account:
https://github.com/ffmulks/data...certification
CODE AVAILABILITY
Florian is a chemistry researcher who strives to use data-
driven decisions to fast-forward experimental and
computational chemical research processes.
DR. FLORIAN F. MULKS
Introduction 2](https://image.slidesharecdn.com/reportffmulks-200714203200/85/A-New-Venue-for-Toronto-IBM-Data-Science-Capstone-Project-by-F-F-Mulks-2-320.jpg)
![ffmulksin
4
Types of Machine Learning
Dimensionality Reduction
Clustering
Machine Learning
Unsupervised Learning
Supervised Learning
Reinforcement Learning
Regression
Classification
QUICK REMINDER
A broad set of machine learning methods was applied in
the scope of this project. Justification and further
explanation is given in the discussion of our results.
This is an overview of the available algorithm groups
classified as machine learning methods. Please refer to
the course material for further details.[1]
USED WITHIN THIS PROJECT](https://image.slidesharecdn.com/reportffmulks-200714203200/85/A-New-Venue-for-Toronto-IBM-Data-Science-Capstone-Project-by-F-F-Mulks-4-320.jpg)
![ffmulksin
6
Business Understanding
Investors or city planners are looking to open a venue in
Toronto that is profitable and contributes toward the local
cultural infrastructure.
LAUNCH ANY VENUE
We are not posing any prior limitations regarding the
location or type of venue. The decision will be purely
based on data-driven predictions.
DECISION CIRCUMSTANCES
Toronto is a cultural melting pot among the top 10 most
populated cities in North America. Rich diversity is found
in venues and citizens.
TORONTO, CANADA
6
Ref. [2]](https://image.slidesharecdn.com/reportffmulks-200714203200/85/A-New-Venue-for-Toronto-IBM-Data-Science-Capstone-Project-by-F-F-Mulks-6-320.jpg)
![ffmulksin
7
Literature Overview
There are plenty of analyses of Toronto’s venue and
population structure due to the privileged role as
standard example in the course capstone.
DATA SCIENCE CAPSTONE
Available studies aim at opening specific venue types,[3] at
finding best places to live,[4] or at declaring an area the
“best” neighborhood.[5]
AIM OF PREVIOUS STUDIES
Most studies start with clustering postal code areas and
proceed to derive decisions based on bar graphs of some
additional data sources based on their goal.
METHODS SUMMARY
7
Ref. [2]](https://image.slidesharecdn.com/reportffmulks-200714203200/85/A-New-Venue-for-Toronto-IBM-Data-Science-Capstone-Project-by-F-F-Mulks-7-320.jpg)
![ffmulksin
Among all venue categories, coffee shops/cafés were
identified by linear regression to be in extreme demand
throughout many central Toronto areas.
HIGH COFFEE DEMAND
Venue clustering showed that there is an extreme coffee
demand in the Church and Wellesley area that even lacks
further venues of similar categories.
BEST LOCATION FOR A STORE
Launch a coffee shop in Church and Wellesley. The area
lacks 4.5 coffee shops and a grand total of 6 venues
catering the need for coffee. This should be very lucrative!
ACTION RECOMMENDATION
Conclusions 23
Ref. [6]](https://image.slidesharecdn.com/reportffmulks-200714203200/85/A-New-Venue-for-Toronto-IBM-Data-Science-Capstone-Project-by-F-F-Mulks-23-320.jpg)
![ffmulksin
24
Thanks for Reading!
Thank you for your attention. If you have any comments,
please get in touch. Constructive criticism is always
welcome.
This project was done to learn the ropes of data science
for application in my research aiming to enable simple
and semi-automatic computational chemical modelling
for experimental researchers. If you are interested in my
work, feel free to contact me or check out my homepage
and social media.
GET IN TOUCH
inff@mulks.ac
mulks.ac ffmulks
ffmulks
24
Ref. [6]](https://image.slidesharecdn.com/reportffmulks-200714203200/85/A-New-Venue-for-Toronto-IBM-Data-Science-Capstone-Project-by-F-F-Mulks-24-320.jpg)
![ffmulksin
25
References
https://www.coursera.org/professional-certificates/ibm-data-science,
07/08/2020.
[1] Course material:
Wladyslaw Sojka via http://www.sojka.photo, 07/08/2020.
[2] Toronto skyline photography:
https://towardsdatascience.com/exploring-toronto-neighborhoods-to-
open-an-indian-restaurant-ff4dd6bf8c8a,
https://capstoneprojectcoursera.wordpress.com/, 07/08/2020.
[3] Best place to open specific venue:
https://medium.com/...-ibm-capstone-project-52b4292ef410,
https://www.linkedin.com/pulse/capstone-project-battle-
neighborhoods-rohitaksh-gs/, 07/08/2020.
[4] Rental or personal flat:
https://github.com/gnavia007/Coursera_Capstone/,
http://roshangrewal.com/...finding-a-better-place-in-scarborough-
toronto/, 07/08/2020.
[5] Best neighbourhood:
Mike Kenneally via https://unsplash.com, 07/08/2020.
[6] Coffee photography:](https://image.slidesharecdn.com/reportffmulks-200714203200/85/A-New-Venue-for-Toronto-IBM-Data-Science-Capstone-Project-by-F-F-Mulks-25-320.jpg)
A New Venue for Toronto: IBM Data Science Capstone Project by F. F. Mulks
- 1. A New Venue for Toronto Capstone Project for IBM Data Science Professional Certificate via Coursera
- 2. ffmulksin This is a presentation of my capstone project for the IBM Data Science Professional Certificate via Coursera.[1] IBM DATA SCIENCE PROFESSIONAL The code can be found in form of a Jupyter notebook on my Github account: https://github.com/ffmulks/data...certification CODE AVAILABILITY Florian is a chemistry researcher who strives to use data- driven decisions to fast-forward experimental and computational chemical research processes. DR. FLORIAN F. MULKS Introduction 2
- 3. ffmulksin Data science is the study of large quantities of data, which can reveal insights that help make strategic choices. WHAT IS DATA SCIENCE? Data scientists need to be curious, judgemental and argumentative. Finding a good story in the data and telling it well is as important as data treatment. DATA SCIENTISTS Data science is a multi-disciplinary field influenced by computer science, software engineering, mathematics, statistics, economics and business management and uses tools from all of these. METHODS Data Science 3
- 4. ffmulksin 4 Types of Machine Learning Dimensionality Reduction Clustering Machine Learning Unsupervised Learning Supervised Learning Reinforcement Learning Regression Classification QUICK REMINDER A broad set of machine learning methods was applied in the scope of this project. Justification and further explanation is given in the discussion of our results. This is an overview of the available algorithm groups classified as machine learning methods. Please refer to the course material for further details.[1] USED WITHIN THIS PROJECT
- 5. ffmulksin 5 Methodology CRoss-Industry Standard Process for Data Mining (1996), project by the EU and five companies. Most widely used analytics model. CRISP-DM, 1996 More refined 10-step version of CRISP-DM of an IBM data scientist. He adds analytic approach, data collection, data understanding and feedback. JOHN ROLLINS’ MODEL, 2015 Analytics Solutions Unified Method for Data Mining/Predictive Analytics, by IBM. Especially refines the model around reiteration. ASUM-DM, 2015 Evaluation Business Understanding Data Understanding Deployment Data Preparation Modeling Data
- 6. ffmulksin 6 Business Understanding Investors or city planners are looking to open a venue in Toronto that is profitable and contributes toward the local cultural infrastructure. LAUNCH ANY VENUE We are not posing any prior limitations regarding the location or type of venue. The decision will be purely based on data-driven predictions. DECISION CIRCUMSTANCES Toronto is a cultural melting pot among the top 10 most populated cities in North America. Rich diversity is found in venues and citizens. TORONTO, CANADA 6 Ref. [2]
- 7. ffmulksin 7 Literature Overview There are plenty of analyses of Toronto’s venue and population structure due to the privileged role as standard example in the course capstone. DATA SCIENCE CAPSTONE Available studies aim at opening specific venue types,[3] at finding best places to live,[4] or at declaring an area the “best” neighborhood.[5] AIM OF PREVIOUS STUDIES Most studies start with clustering postal code areas and proceed to derive decisions based on bar graphs of some additional data sources based on their goal. METHODS SUMMARY 7 Ref. [2]
- 8. ffmulksin 8 Toronto Postal Codes and Venues We will use the 103 postal codes starting with M in Toronto, Canada, scraped from Wikipedia together with assigned boroughs and neighborhoods. DATA SET Venue data was requested from the Foursquare API within 500 m radii of postal code centroids. A limit of 100 venues per postal code request was applied. VENUES The collected dataset contains names, addresses, geographic locations, and venue categories. We have extracted 2130 venues from 271 categories spread over 103 areas defined by 500 m radii around the center of postal codes. DATA OVERVIEW
- 9. ffmulksin 9 Geographic Venue Distribution We computed the total number of venues and the mean distances between them by calculating the earth surface distances (Haversine distances). FEATURE CONSTRUCTION Most investigated 500 m radii only contain less than 20 venues. Venue distances appear to be rather invariant among the different postal code areas. DISTRIBUTION Models are expected to be of limited representativeness as limited data amounts per postal code radius exist. FIRST UNDERSTANDING Click buttons to control figure display: distribution plot box plots
- 10. ffmulksin 10 Venue Density Distribution High venue counts should lead to high venue densities which could be a good measure of competitiveness but also lucrativity of a location. VENUE DENSITY Due to little data of neighborhoods with high venue counts, no good correlation can be found. A slight decrease in density is found with higher counts. DISTRIBUTION Distances seem broadly spread in the low venue count neighborhoods. The mean actually slightly increases with increasing venue counts. Let us see if we can model the structure of the venue density. WEAK CORRELATION EXPECTED
- 11. ffmulksin 11 Venue Density Regression Linear regression was applied to the venue density and we created polynomial features of the distances and venue counts to investigate the data structure. REGRESSION MODELS A poor correlation with R2 0.01 is found with a linear model, the cubic model delivers an R2 score of 0.06 (only explaining 6% of the data variation). CORRELATION We cannot employ the venue density for a priori binning of our data, but the distance between the venues seems a valuable feature to keep for further analyses. FEATURE TREATMENT Click buttons to control figure display: cubic regression linear regression
- 12. ffmulksin 12 Venue Categories Distribution Foursquare defines categories to explain what exactly the customer can expect when visiting a venue. FOURSQUARE CATEGORIES The categories are extremely detailed. The vast majority of the 271 venue categories occurs only once in the whole observed region in Toronto. DISTRIBUTION Clustering and Principal Component Analysis (PCA) will be needed to reduce the dimensionality. 271 categories with only 103 samples (postal codes) make for ill-defined matrices reducing the choice of applicable algorithms. HIGH DIMENSIONALITY
- 13. ffmulksin 13 Clustering Algorithms Several algorithms and initialization types were employed to find similarities in neighborhoods based on their venue counts and categories. CLUSTERING Most algorithms only allocate downtown areas in smaller clusters. The vast majority is assigned to one or sometimes two “low venue count” clusters. DISCOVERED PATTERNS K-Means (k-means++ initialization) was chosen for further investigation as it was capable of showing some structures even outside of downtown Toronto. ALGORITHM CHOICE K-Means (Random, PCA) K-Means (Random)K-Means (k-means++)DBSCANAgglomerative Clustering Click buttons to change algorithm: k = 52 clusters found
- 14. ffmulksin 14 K-Means Number of Clusters k More clusters will always capture more variation of the data. The cluster number at which the steepness of this accuracy gain drops is usually chosen. ELBOW METHOD While we find the correct location of downtown neighborhoods at all k, the elbow plot shows no good k due to high similarity of most areas. RESULTS Due to the high similarity of our observed areas, clustering does not deliver useful insights with the employed data set. HIGH AREA SIMILARITY 9− 7 +− 5 +− 3 + Click buttons to control figure display: elbow plot choose k
- 15. ffmulksin 15 2d-Principal Component Analysis (PCA) The axes are found to correlate to venue counts from certain categories, indicating similarities between venues of certain categories. NEIGHBORHOODS The matrix was transposed to find structures in our categories rather than areas. The x-PCA represents high density, the y-PCA low density neighborhoods. CATEGORIES Certain venues are likely to appear in certain density neighborhoods. PCA is valuable for grouping the categories. We will reduce the dimensionality of our data set to 30 (11% of 271 categories) to create more meaningful variables. DATA STRUCTURE Click buttons to control figure display: categories neighborhoods
- 16. ffmulksin 16 Model Development The difference between true and MLR-based predictions for venue counts were used to evaluate the demand for venues in Toronto neighborhoods. PREDICTIVE MODEL Categories were clustered to identify similar venue types catering the same needs. These were used to correct for such saturated demands. REAL DEMAND CORRECTION The resulting predictions show both the location and the category for venues that are in demand in Toronto. DEMAND PREDICTION Real Venue Demand Venue Counts in Categories Category Clusters Venue Demand Predicted Venue Counts Cluster Demand Predicted Cluster Venue Counts K-Means Multiple Linear Regression Difference True vs. Predicted Product Venue x Cluster
- 17. ffmulksin 17 Demand Prediction MLR with 5-fold cross-validation was performed to predict every single of our 273 variables (271 categories + venue count + mean distance). MULTIPLE LINEAR REGRESSION (MLR) The resulting prediction matrix emulates average healthy Toronto neighborhoods based on the number of neighboring venues of certain categories. VENUE COUNT PREDICTION Differences between true and predicted values can be employed to predict market oversaturation and, more importantly, market demand for venue types. VENUE DEMAND PREDICTION
- 18. ffmulksin 18 Category Clustering Some categories are very similar. Predicted coffee shop demands might fully be saturated due to existing cafés. DEMAND SATURATION We used K-Means clustering with our transposed data to find 30 category clusters to reduce the dimensionality of our data. DIMENSIONALITY REDUCTION The clusters nicely capture other venues that might satiate the needs for e.g. coffee such as hotels, restaurants and also some surprising categories. CLUSTERS CAPTURE INTERACTIONS General Travel, Modern European Restaurant, Steakhouse, Restaurant, Plaza, Cuban Restaurant, Gift Shop, New American Restaurant, Brazilian Restaurant, Japanese Restaurant, Smoke Shop, French Restaurant, Pub, Art Gallery, Shopping Mall, Speakeasy, Tea Room, Italian Restaurant, Salon / Barbershop, Food Court, Vegetarian / Vegan Restaurant, Seafood Restaurant, Concert Hall, Nightclub, Gluten-free Restaurant, Soup Place, American Restaurant, Bakery, Department Store, Gastropub, Hotel, Coffee Shop, Opera House, Food Truck, Lounge, Asian Restaurant, Art Museum, Cupcake Shop, Train Station, Beer Bar, Colombian Restaurant, Café, Record Shop, Bookstore, Deli / Bodega, Building, Men’s Store, Fast Food Restaurant, Wine Bar, Dog Run, Monument / Landmark, Museum, Thai Restaurant, Salad Place COFFEE SHOP CLUSTER
- 19. ffmulksin 19 Demand in Venue Categories We analyzed the demand as difference between predicted and true venue counts over the 30 category clusters. CLUSTER DEMAND The highest demand for venue is found in the cluster containing coffee shops. The map shows neighborhood demands in the coffee cluster. COFFEE CLUSTER HIGHLY DEMANDED Most suburban areas have a fulfilled demand in the coffee cluster. We found that many downtown areas would support several more venues in the cluster. STRONG DEMAND IN DOWNTOWN AREA Click buttons to control figure display: bar chart−+zoom coffee cluster
- 20. ffmulksin 20 Coffee Cluster Demand in Neighborhoods The coffee cluster showed the highest demands over all clusters. We now looked at the distribution of this demand over the different locations. DEMAND LOCATION Toronto seems to have a healthy amount of venues in the category but their location is suboptimal with both over- and undersaturated neighborhoods. DEMAND DISTRIBUTION There are several neighborhoods that strongly demand more venues in the coffee cluster. Let us look into the details to find out which venue category would be the most lucrative to launch. AREAS WITH HIGH DEMAND FOUND
- 21. ffmulksin 21 Coffee and Café Demand Coffee shops and cafés are demanded the most within their cluster. Their combined demand is shown in the maps. CATEGORY DEMAND Toronto needs coffee! The two almost identical categories coffee shops and cafés are in extreme demand especially in the Church and Wellesley area. COFFEE DEMAND We now need to make sure that this neighborhood is not saturated with e.g. tea rooms and other venues that serve coffee as well. INTERACTION WITH OTHER VENUES Click buttons to control figure display: bar chart−+zoom
- 22. ffmulksin 22 Real Demand The product of the market demands (negative market saturation) of the whole cluster with the demand of only coffee shops and cafés was taken. DEMAND PRODUCT In some cases, the coffee demand is largely saturated by other venues (larger yellow than red circle). One extreme demand product is found. SATURATION OF COFFEE DEMAND The neighborhood Church and Wellesley lacks 4.5 coffee shops/cafés and is even lacking 6.0 venues from the whole cluster which yields an outstanding demand product of 26.7. OUTSTANDING DEMAND FOUND Cafés: Cluster: Product: 2.1 5.3 10.9 Cafés: Cluster: Product: 4.5 6.0 26.7 Cafés: Cluster: Product: 1.4 6.2 8.8 Cafés: Cluster: Product: 1.8 4.2 7.5 Cafés: Cluster: Product: 4.5 6.0 26.7 Click buttons to control figure display: −+zoom
- 23. ffmulksin Among all venue categories, coffee shops/cafés were identified by linear regression to be in extreme demand throughout many central Toronto areas. HIGH COFFEE DEMAND Venue clustering showed that there is an extreme coffee demand in the Church and Wellesley area that even lacks further venues of similar categories. BEST LOCATION FOR A STORE Launch a coffee shop in Church and Wellesley. The area lacks 4.5 coffee shops and a grand total of 6 venues catering the need for coffee. This should be very lucrative! ACTION RECOMMENDATION Conclusions 23 Ref. [6]
- 24. ffmulksin 24 Thanks for Reading! Thank you for your attention. If you have any comments, please get in touch. Constructive criticism is always welcome. This project was done to learn the ropes of data science for application in my research aiming to enable simple and semi-automatic computational chemical modelling for experimental researchers. If you are interested in my work, feel free to contact me or check out my homepage and social media. GET IN TOUCH inff@mulks.ac mulks.ac ffmulks ffmulks 24 Ref. [6]
- 25. ffmulksin 25 References https://www.coursera.org/professional-certificates/ibm-data-science, 07/08/2020. [1] Course material: Wladyslaw Sojka via http://www.sojka.photo, 07/08/2020. [2] Toronto skyline photography: https://towardsdatascience.com/exploring-toronto-neighborhoods-to- open-an-indian-restaurant-ff4dd6bf8c8a, https://capstoneprojectcoursera.wordpress.com/, 07/08/2020. [3] Best place to open specific venue: https://medium.com/...-ibm-capstone-project-52b4292ef410, https://www.linkedin.com/pulse/capstone-project-battle- neighborhoods-rohitaksh-gs/, 07/08/2020. [4] Rental or personal flat: https://github.com/gnavia007/Coursera_Capstone/, http://roshangrewal.com/...finding-a-better-place-in-scarborough- toronto/, 07/08/2020. [5] Best neighbourhood: Mike Kenneally via https://unsplash.com, 07/08/2020. [6] Coffee photography:
- 26. ffmulksin 26