Clustering Analysis for Property Valuation:

 From Complex Unlabeled Data to Brightly Lit Information

Written by Helvita Dorojatun

(Valuer and Data Analytic Enthusiast)

Introduction

How to make smart decision when you are faced with a huge of multi-variable data? Understanding the details of each data certainly allows us to make an accurate decision but time constraints are always there. Systematic steps are necessary for extracting useful information and knowledge from data. One of the systematic procedures that can be taken is using Clustering algorithm. In this article, we will propose Clustering Algorithms as an alternative to improve the quality and quantity of the valuer's work.

Clustering is an unsupervised learning technique that perform the process of grouping similar objects or the partitioning of a data set into subsets. Clustering algorithms can analyze data sets with numerical or categorical variables. Some examples of using the clustering algorithms are determining the right specific customer, Zoning, Property market segmentation, determining patterns in real estate, and recognizing pattern in unprepared data. 

There are four types of clustering algorithms:

1.   Partition based clustering for divide a set of data into a number of clusters

2.   Hierarchical clustering to determine the close relationship between data  

3.   Density Based Clustering to find out the main patterns from data that is full noise

4.   And the last but not the least is model based clustering which is an exploratory algorithm that has many variations and still continues to develop.

Why Clustering Algorithm is Relevant in Valuation Process?

Valuation can be described as procedure of developing objective and unbiased opinion about the value of an asset. This process is carried out by comparing the subject property with comparable properties. The closer between the subject and the comparable data, the more reliable the value will be. The clustering algorithm is a method commonly used by researchers to group database on close connectivity.

The clustering algorithm was developed from an ancient formula for measuring the shortest distance between two points. Euclidean distance is calculated as the square root of the sum of the squared difference between latitude and longitude components in two vectors. This formula was growing to measure more than two attributes in the data. The way this algorithm works is applicable to the property valuation process which uses a lot of proximity to quantify the influence of a variable. The use of mathematical formulas in this method makes the clustering algorithm work objectively. This will certainly ensure consistency and accountability in the process of property valuation.

Application of data mining in property valuation is becoming mandatory as data production increases. The Ideal data in the property valuation at least consists of several comparables, identical property attribute, up to date information, verified outcome from fair market value, and data from active market transaction. The expectation of this ideal data will be faced with a several challenges, namely: the imperfect of real estate market, the availability of data in less active market and the significantly different between the available data and the valuation subject (see RICS:2019). Data identification and collection are now getting faster and cheaper as access to data increases, so that currently the valuation process emphasizes more on data analysis. This fact ultimately encourages valuer to expand knowledge base in managing data efficiently (see: Appraisal Institute 2020; 81).

How Does Clustering Algorithm Work?

The first and most widely used Partition-Based Clustering algorithm is K Means. This algorithm forms data groups based on the proximity between the data and the cluster centeroid points obtained through an iterative process. This algorithm will form smaller distance within cluster members than distance between clusters. An explanation of the K-means clustering algorithm is presented in the figure below.

How Clustering Algorithms Can Address Challenges in Property Valuation Data?

Property valuation data in general has several uniqueness, namely it is mixed with outliers, has attributes in categorical form and has different units or scale forms. This condition can be overcome by modifying some mathematical equations.

Outliers on a dataset have an impact on the results of clustering in K mean. When the centeroid in K Means is determined by the average of the cluster members, the question arises, what if there is an outlier in the cluster members? The outliers will pull the average to the right or left of the pool actual data set. This condition initiates K Medoids and K Median approach are born.

The K-medoids algorithm uses the actual value that has the smallest dissimilarity to all other objects in the cluster. While in the K Median, centeroid is selected from the median data (mid points). K Medoids and K Median are designed to be more robust outliers in the data or large range data (high variability).

Are categorical attributes a problem for data mining? We often have property data attributes in categorical form. In general, this can be solved by transforming the form of data into a numerical.  This procedure can produce biased conclusions as the centeroid in K Means will be determined based on the average value of the cluster members.

Allowing data attributes in categorical form and processing them directly becomes an alternative.  The clustering algorithm used to extract categorical data is K Mode. This algorithm uses the value that appears most frequent value in a set of data (Mode) as a centeroid. Basically, modifying the way the centeroid calculates will result in a compromise zone between the analysis we want and the data we have.

Data analytics enthusiast frequently must deal with data that has a very varied measurement scale. The difference in scale greatly affects the results. Then what treatments can be done to obtain unbiased results? Normalization and standardization are one of the steps that can be taken to produce the best data processing (see: Dorojatun, 2022).

How Many Clusters of Property Data Are Optimal?

The number of Clusters can be determined by an analyst who is well aware of the required number of data classes (e.g., zoning, market type). However, mathematically, the number of data clusters can be determined using several approaches. Some of the approaches discussed in this article are Silhouette Score and Davis Boudin Indexes

The scale of silhouette score has a range of between -1 and 1. Silhouette score near +1 indicate that distance between clusters is far away. Score 0 indicates that the distance between two neighboring clusters very close (class overlap problem) or the cluster member has the possibility assigned to the wrong cluster. So, the most optimal cluster is the cluster with the highest Silhouette score.

Determining the number of clusters is done using Davies Boudin index. The most recommended cluster is the cluster with the smallest DBI value. However, when this technique shows a linear downward trend in DBI scores with increasing cluster numbers, analysts combine it with the Elbow method. Stages of Clustering Algorithm by orange data mining platform is presented in the figure below.

Implementation of Cluster Analysis in Property Valuation

1.     Segmentation

Market segmentation is one of the two questions that must be answered in market analysis.  In property valuation context, valuer can define properties in several segments to obtain information on where demand is coming from and where competitors are from based on its characteristics (see: The Appraisal Institute, 2020). To support the accuracy of value, the comparable data and the property being valued must be in the same market segment. Property markets can be segmented by clustering algorithm with using demographic variables such as age, education level, salary, family size, gender, race, and marital status.

2.     Comparable Data Evaluation

Determining the best three of comparable data. In the sales comparison approach, Value indication is generated from the process of adjusting the difference between the comparable and subject properties. The most credible indication of market value of an asset is in line when the difference with the comparable is very close. The Clustering Algorithm mathematically and consistently will show the comparable data that has the most identical attribute to the subject being valued. The Clustering algorithm will work very well especially when the comparison variable is quantified in proximity. For example, the valuation for residential property is affected by the proximity to public transportation, proximity to open space and parks, proximity to arterial roads, etc.

3.     Pattern Recognition

Transaction pattern studies often must be done by valuers in huge of data with multi variables and multi categories. Clustering can provide this information, as well as show where you stand among comparable data. An example of the benefits of pattern recognition is to determine zoning in comparable data for property valuation.

Conclusion

Clustering is a data mining technique that tells us how far and close the relationship between data is through mathematical equations. Through modifications to the centeroid determination method, this algorithm continues to evolve to find the form that best fits the analysis we want. Through this method, valuer can work systematically and efficiently to improve the accuracy of the valuation results.

Reference:

Appraisal Institute. (2020). The Appraisal of Real Estate (15th ed.). Chicago, ISBN 9781935328780

Dorojatun, H. (2022). Seri Artikel DDDM KPKNL Mamuju: Normalisasi dan Standardisasi dalam Data Mining. Retrieved from https://www.djkn.kemenkeu.go.id/artikel/baca/15943/Seri-Artikel-DDDM-KPKNL-Mamuju-Normalisasi-dan-Standardisasi-dalam-Data-Mining.html.

RICS. (2019, October). Comparable evidence in real estate valuation, RICS professional standards and guidance (1st ed.). ISBN 978 1 78321 373 3

(This article is part of KPKNL Mamuju's Data Driven Decision Making (DDDM) article series for the Ministry of Finance)