Classification Algorithms for Identifying Neighborhood Typology
Robert Brasso
Data Scientist
Data Visualizer
Data Analyst
Databricks
Python
scikit-learn
Towson University
1
— Home vacancies remain a major problem that
many medium-large cities deal with. They use up valuable land,
are often uninhabitable, cause a tax burden on the city, and bring
down property values for the neighborhoods they are in. This
paper explores the assumption that there is a relationship
between home vacancy and the level of distress within their
neighborhood. We will use a 2011 Housing Market Typology
report from Baltimore Open Data to analyze whether or not
certain classification algorithms can be implemented to classify
the typology of a neighborhood with the hopes of being able to
apply this algorithm in future studies to help predict
neighborhood statuses based on real-time housing metrics.
I. INTRODUCTION
In 1950, Baltimore’s population was recorded at just under
a million people and it ranked as the sixth largest city in the
United States. Today, Baltimore’s population is around
620,000 representing a 35% decline in the city’s population
over the last 60 years. The result of this dramatic decrease in
population is that the city contains an inflated amount of
vacant homes. According to an article in
2014, Baltimore has upwards of 16,000 vacant homes and
vacant homes outnumber homeless people 6 to 1. While the
city has tried for years to address this issue, most policies have
fallen short of making a large-scale difference in the problem.
One philosophy to solving the problem of vacant homes is to
increase the property value of the surrounding area, thus
incentivizing people to move in and either renovate or replace
these homes.
Using the 2011 Housing Market Topology dataset from
Baltimore Open Data, we are attempting to create a model that
accurately predicts the market category that a neighborhood
will fall into based on a variety of other indicators.
In the paper that follows, we will provide a description of
the dataset, results from our EDA (exploratory data analysis),
develop a training and testing dataset, and then run three
classification algorithms against the testing dataset to
determine the best possible classification model for our
experiment.
II. DATASET DESCRIPTION
Open
Baltimore
Baltimore that provides free datasets relating to the city. The
dataset is called “2011 Housing Market Typology” and was
created in order to inform neighborhood planning efforts, also
informing residents of the local housing market conditions in
their communities. The entire dataset has 626 records,
contains 12 fields, and can be found by going to the following
Below each of the fields will be described in detail.
blockGroup
Baltimore census block groups can be found at
marketCategory
of the following groups…
Regional Choice
markets with high owner-occupancy and
high property values
Middle Market Choice –
above the city average with strong
ownership rates, low vacancies, but slightly
increased foreclosure rates
Middle Market –
as well as high ownership rates. Higher
foreclosure rates, with slight population
loss.
Middle Market Stressed –
home sales than city average and have not
shown significant sales price appreciation.
Vacancy and foreclosure rates are high and
the rate of population loss has increased.
Distressed –
housing stock. Contains high vacancy and
the lowest homeownership rates. Most
substantial population loss.
sales20092010 –
from 2009 to 2010.
salesPriceCoefficientVariance –
standard deviation from 2009-2010 divided the mean
sales price from 2009-2010.
commericalResidentialLandRatio –
and institutional land area divided by residential land
area.
unitsPerSquareMile –
square mile.
residentialPermits –
equal to $50,000 divided by residential lots plus
vacant lots.
Analysis of 2011 Housing Market Topology
Data for the City of Baltimore
Robert Brasso
2
vacantLots –
residential plus vacant lots.
vacantHouseNotices –
divided by residential plus vacant lots.
foreclosureFilings –
2010 as a percentage of privately owned residential
lots.
medianSalesPrice20092010 –
price for homes in census block from 2009-2010.
ownerOccupied –
residential units that are owner occupied.
III. EXPLORATORY DATA ANALYSIS
Beginning our exploration of the dataset, the first piece of
information we needed was the breakdown of the amount of
records for each of the market classifications. 209 of the
census blocks were reported as Distressed, 117 as Middle
Market Choice, 114 as Middle Market Stressed, 94 as
Regional Choice, and 92 as Middle Market.
The next task was to take a look at all of the specific
attributes and identify how they interact with each other. In
figure 2.1 (end of paper) you will see a scatter matrix of all of
the attributes. The attribute names have been adjusted slightly
for readability. From a review of the very large scatter matrix,
there were a handful of relationships I wanted to explore
further. Below are a series of box plots used to analyze these
relationships.
The first relationship examined was between market
categories and Vacant Land Percentage. Based on the above
figure, it appears there is a higher percentage of vacant homes
in the distressed and middle market distressed neighborhood
blocks.
Additionally, on the next boxplot we see that the owner-
occupied rates for those neighborhoods is significantly less
than the others.
Additionally, it appears that the sales price coefficient may
be a significant indicator of distressed and middle market
stressed communities.
While there are likely more relationships that could be
identified, it is fair to say that the attributes in this dataset
provide some insight to the classification they have and
3
therefore will be a good candidate for our classification
algorithm testing.
IV. CLASSIFICATION MODELS
Our goal with testing the following classification algorithms
is to isolate and algorithm that will accurately classify housing
categories based on a variety of statistics. Once we have
identified the best algorithm, we can identify how the values
entered into the algorithm impact the classification, however
that is beyond the scope of this paper. Future research can use
this algorithm to identify the key indicators that describe
neighbors and attempt to provide early detection when
neighbors begin to fall into distressed and middle market
distressed categories
For testing this data against the classifier algorithms, we
used a 70/30 split to create our training and testing datasets.
The three algorithms we will be analyzing are K Nearest
Neighbors, Random Forest, and Naïve Bayes. Below are
descriptions of the results from each of our algorithm testing
A. K Nearest Neighbors
For K Nearest Neighbors we set the distance measure as
Euclidian distance and first set N = 2. The initial results from
N = 2 were promising, but in order to maximize the accuracy,
recall, and precision, we wanted to test how various N values
impacted the results. Below is a chart of the N values and their
corresponding metrics.
N - Value Accuracy Precision Recall F-Score
1 .9521 .95 .95 .95
2 .9521 .95 .95 .95
3 .9468 .95 .95 .95
4 .9628 .96 .96 .96
5 .9521 .95 .95 .95
6 .9521 .95 .95 .95
7 .9415 .94 .94 .94
After testing several values of N, we see that the best results
came from N=4. While the algorithm clearly works well with
all values of N under 7, performance appeared to be optimized
at N=4. The lowest performance was when N=7 and leads me
to believe that even higher values of N would result in lower
scores. To test that, I re-ran the algorithm with N=50 and
while the scores were lower than N=4, they were almost
identical to N=7.
B. Random Forest
Similar to our testing for K Nearest Neighbors, we will
make a chart of the accuracy, precision, recall, and F Scores
where the N_Estimator value is changed. Below is our test
results for Random Forest.
N_Estimator Accuracy Precision Recall F-Score
5 .8830 .89 .89 .89
10 .8883 .89 .89 .89
50 .9149 .92 .91 .91
100 .9521 .95 .95 .95
150 .9309 .93 .93 .93
200 .9309 .93 .93 .93
300 .9309 .93 .93 .93
For the Random Forest algorithm, it appears that the best
we can do is .9521 accuracy with a .95 f-score. While these
values are very strong, they fall short of K Nearest Neighbors
when N=4. Our optimized value for Random Forest appears to
be when N_estimator = 100.
C. Naïve Bayes
Plugging our training and testing datasets into the Gaussian
Naïve Bayes model, we received the following scores;
accuracy = .9309, precision = .94, recall = .93, and f-score =
.93.
V. CONCLUSION
After reviewing all of our data it became evident that all of
the algorithms we tested do a fair job at classifying data into
the various housing categories. However, the clear winner of
our testing is K Nearest Neighbors which got us up to a .96 f-
score with .9628 accuracy. Beyond that, K Nearest Neighbors
was also the most accurate in all of those scores for each of the
individual classification categories. Comparatively, while the
Naïve Bayes model was close in terms of overall scores, it
performed relatively poorly when classifying regional choice,
providing a precision score of .79, which was the lowest score
seen for any category during this testing. This leads me to
believe that the best algorithm of the ones we tested for this
dataset is K Nearest Neighbors, while Naïve Bayes was the
worst.
REFERENCES
FIGURES
2.1
