Automated Classification of Building Objects Using Machine Learning
Nadeem Iftikhar
a
, Peter Nørkjær Gade, Kasper Møller Nielsen and Jesper Mellergaard
University College of Northern Denmark, Sofiendalsvej 60, Aalborg, Denmark
Keywords:
Building Information Modeling, Machine Learning, Building Object Classification, Digital Tools.
Abstract:
In the construction sector, digital technologies are being employed to enable architects, engineers and builders
in the creation of digital building models. Although these technologies come equipped with inherent classifi-
cation systems, they also bring forth certain obstacles. Frequently, these systems categorize building elements
at levels that exceed their necessary specificity. To illustrate, these classification systems might allocate values
at a broader granularity, such as “exterior wall” rather than at a more precise level, like “exterior glass wall
with no columns”. As a result, the manual classification of building elements at a granular level becomes
essential. Nonetheless, manual classification frequently results in inaccuracies and erroneous semantic details,
while also consuming a significant amount of time. Precise and prompt classification of building objects holds
significant importance for activities like cost planning, construction cost management and overall procurement
processes. To address this, the current paper suggests an automated classification approach for building ob-
jects, focusing on specific types, through the utilization of machine learning. The effectiveness of the proposed
system is showcased using real-world data from a prominent architectural firm based in Scandinavia.
1 INTRODUCTION
The construction sector is experiencing a digital revo-
lution in response to heightened demands for sustain-
ability, safety and user specifications. This shift to-
wards digital transformation necessitates the adoption
of novel procedures to oversee and harmonize digi-
tal workflows. In this context, Building Information
Modeling (BIM)
1
stands out as a renowned instrument
employed by architects, engineers and builders to
generate, oversee and distribute digital models. These
models encompass a substantial volume of building
data, comprising both geometric and informational el-
ements. Hence, it is critical to systematically classify
this building information in accordance with industry
standards, ensuring a streamlined and proficient pro-
cess. A classification system operates similar to a uni-
versal language, facilitating a clear and unambiguous
exchange of digital data across diverse BIM software
platforms and among various BIM users. The act
of classifying building information serves a dual pur-
pose. On one hand, it enhances communication across
all stakeholders engaged in construction projects. Si-
multaneously, it empowers project collaborators (ar-
chitects, engineers and builders) to effectively align
a
https://orcid.org/0000-0003-4872-8546
1
https://www.autodesk.com/industry/aec/bim
projects with requirements, schedules and financial
allocations. While numerous building design tools
possess the capability to automatically recognize and
classify building objects, either within general cate-
gories like “door” or within specific families like “sin-
gle flush door,” it remains essential to classify these
objects at the precise type level conforming to na-
tional or international standards before their appli-
cation in a construction project. Nevertheless, the
majority of building design tools lack the inherent
capacity to automatically classify building objects
at the type level or autonomously assign assembly
codes to them. These assembly codes comply to es-
tablished national or international classification stan-
dards and remain easily modifiable over the course
of the project’s life cycle. Additionally, the assem-
bly codes play a pivotal role in dictating the organi-
zational structure of construction information, align-
ing it with the core building objects. Consequently,
the assignment of classifications at the assembly code
level is often carried out through manual intervention
by architects and engineers. For instance, using the
Cuneco Classification System (CCS)
2
, widely used
in Denmark, a product like a “window” can be cat-
egorized as “[L]%QQA90102.01”, with the accom-
panying description indicating its nature as an “ex-
2
https://ccs.molio.dk/News/About CCS
Iftikhar, N., Gade, P., Nielsen, K. and Mellergaard, J.
Automated Classification of Building Objects Using Machine Learning.
DOI: 10.5220/0012197500003598
In Proceedings of the 15th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2023) - Volume 1: KDIR, pages 331-338
ISBN: 978-989-758-671-2; ISSN: 2184-3228
Copyright © 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
331

terior multidisciplinary window - type 01”. Despite
the availability of software applications designed to
facilitate automated management of CCS codes, the
correct codes still necessitate manual input in the ma-
jority of classification systems.
The manual categorization of numerous objects at
a specific level within a building model introduces
complexities, consumes considerable time and intro-
duces potential ambiguities in the classification pro-
cess. Additionally, a previous study (Flager and Hay-
maker, 2007) highlighted that over 50% of the time
and resources invested by architects and engineers are
allocated to the management of design information,
which includes object classification. Moreover, errors
during the classification process or inaccuracies in se-
mantic details can result in flawed construction cost
estimates, incorrect construction practices, erroneous
material choices and related issues. Despite the ar-
ray of challenges associated with BIM object classifi-
cation, including the search for relevant international
or national BIM standards, establishment and storage
of tailored classification tables, collaborative distribu-
tion of classification tables among project teams, nav-
igation through multiple classification systems, time-
intensive processes and absence of automation. The
acceptable automation of BIM object classification at
the precise type level remains an open research con-
cern within the construction sector. This paper at-
tempts to tackle this issue by introducing an approach
that employs supervised machine learning to auto-
mate the classification of objects at the type or assem-
bly code level.
In summary, this paper’s key contributions can be
outlined as follows:
− Offering comprehensive insights through ex-
ploratory data analysis of building information;
− Introducing a comprehensive solution for classi-
fying building elements through the utilization of
diverse supervised machine learning approaches;
− Demonstrating practical implementation using ac-
tual building data extracted from a prominent con-
struction project executed by a leading Scandina-
vian architectural firm.
The paper’s organization is as follows: In Section 2,
a survey of related work is provided. The motivation
driving this work is elaborated in Section 3. The ex-
ploratory data analysis is detailed in Section 4. The
machine learning approach for automated building
object classification is presented in Section 5. Sec-
tion 6 outlines the experimental findings. The paper
concludes in Section 7, also highlighting potential di-
rections for future research.
2 RELATED WORK
The focal point of this section centers on prior re-
search attempts related to automated building object
classification through the application of semantic en-
richment. A state-of-the-art review by (Zabin et al.,
2022) indicated that there is a need to integrate ma-
chine learning into BIM processes. Similarly a study
by (Amor and Dimyadi, 2021) examined evolving
approaches for automated compliance checking and
pointed out that research in semantic enrichment of
BIM models is necessary. In addition, (Zhang and
El-Gohary, 2012) suggested a natural language pro-
cessing approach for automated information extrac-
tion from construction regulatory documents. Sim-
ilarly, an automated compliance checking approach
is presented by (Salama and El-Gohary, 2011). A
prototype software based on an inference rule engine
to semantically enrich the building models is intro-
duced by (Belsky et al., 2016) and its extension is
suggested by (Sacks et al., 2021). Further, (Bloch and
Sacks, 2019) applied both supervised machine learn-
ing and rule-inferencing to correctly classify rooms
types in residential apartments. The machine learning
approach provided better accuracy than rule-based ap-
proach. A data-driven iterative method for automated
classification of objects in BIM is presented by (Wu
and Zhang, 2019). Moreover, a deep learning frame-
work to utilize both geometric and relational informa-
tion of BIM objects for classification is proposed by
(Luo et al., 2022). Likewise, a deep learning frame-
work for classifying objects based on synthetic data
sets created from BIM objects is presented by (Fr
´
ıas
et al., 2022). In addition, (Kim et al., 2019) proposed
an approach for automating the classification of build-
ing element instances within BIM. The approach uti-
lized deep learning based classification technique that
uses images of objects as inputs. A machine learning
based solution to automatically recognise elements in
buildings information models is proposed by (Bassier
et al., 2017). The solution can efficiently classify the
basic components such as floors, ceilings, roofs, walls
and so on. Additionally, (Emunds et al., 2022) pro-
posed a neural network model based on sparse con-
volutions for the classification of IFC-based geome-
try and semantic enrichment of BIM models. Con-
cludingly, the study conducted by (Koo et al., 2022)
examined the viability of deep and machine learn-
ing models in the context of automatically classifying
subtypes of door and wall elements.
These previous studies emphasize on various con-
ceptual aspects and recent advancements in semantic
enrichment and automated classification in construc-
tion industry. It can be concluded from these previous
KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval
332

works that the absence of automation in the classifica-
tion of building objects at specific type level remains
an open research issue. To our best understanding,
this paper stands among a limited number that con-
centrate on automated building object classification
through machine learning. Additionally, this paper
also considers the practical aspects of automated clas-
sification to enhance operational effectiveness and ef-
ficiency within the construction industry.
3 MOTIVATION
Given that the construction industry employs BIM to
create digital presentation of buildings. Hundreds,
or even thousands, of objects are incorporated into
a model before its completion. In the majority of
instances, these objects are required to be manually
classified. As previously noted, manual classification
within the domain of BIM poses a significant chal-
lenge.
Figure 1: BIM complete 3D model.
In this paper, the focus is on automatic classifi-
cation of building objects based on their assembly
codes. The assembly codes used in this paper are
found in BIM7AA standard
3
, where BIM7AA is a
simple encoding method for BIM objects based on
Danish building standards. For instance, a typical
“exterior precast concrete wall” has assembly code
211, and “interior frame structured wall” has as-
sembly code 224, and so forth. While building a
BIM model, an architect/engineer inserts a standard
“wall” and edits it into the desired detailed specifica-
tions; these often change in the duration of a project.
Therefore, the correct assembly code based on Dan-
ish building standards is hard to predict. Hence, the
architects/engineers have to point-and-click their way
through every object in the BIM and classify their as-
sembly codes. Classifying hundreds or thousands of
objects manually is time-consuming and erroneous.
In addition, ambiguous classification results in addi-
3
http://bim7aa.dk/index UK.html
tional and/or unwanted expenses.
Figure 2: BIM 3D model (walls only).
To provide an illustration, Fig. 1 portrays a com-
prehensive 3D BIM model featuring walls, roofs,
spaces, coverings, stairs and columns. Conversely,
Fig. 2 exclusively displays the walls present within
the model. The data set encompassing all wall el-
ements within the model comprises a total of 4025
walls, where each of these walls has over 160 fea-
tures. The features contains a wide range of data types
including numerical, alphanumerical, unique identi-
fiers, binary indicators and categorical attributes.
Table 1: Selected set of wall features.
Wall feature Value Description
Area 92.09 length * height
Base Constraint PLAN 10 floor level
Length 26489.99 wall’s length
Structural 1 yes 1/no 0
Structural Usage 1 non load-bearing 0/load-bearing 1
Type Id 2147048 wall type
Unconnected Height 4462.99 wall’s element height
Volume 49.39 walls’s volume
Width 540 wall’s width
Assembly code 211 exterior precast concrete wall
For the purpose of illustration, only a subset of 10
features out of 160 (for the outer wall distinguished
by its highlighted color in Fig. 2) has been chosen.
A snapshot of feature data is presented in Table 1.
The final row in Table 1, displays the assembly code,
which is intended to be automatically assigned or pre-
dicted.
4 EXPLORATORY DATA
ANALYSIS
In this section, an exploratory data analysis of the
classification problem has been carried out. The pri-
mary objective of this analysis is to investigate the
data set and pinpoint the features that wield an im-
pact on the assembly code. To address the extensive
array of features (totaling 160 in this instance), Prin-
Automated Classification of Building Objects Using Machine Learning
333

cipal Component Analysis (PCA)
4
has been employed
to reduce their dimensionality. To evaluate the signif-
icance of the chosen features, one method involves
creating a heatmap that visualizes their correlations.
By depicting the correlations between these features,
a heatmap offers a graphical depiction of their inter-
relationships, facilitating the evaluation of their im-
portance. The correlation coefficient is measured on
a scale that spans from +1 to 0 to -1. A correlation
of -1 indicates a robust negative correlation between
two values, while a correlation of +1 indicates strong
positive correlation, and a value of 0 implies no cor-
relation between them.
Figure 3: Correlation heatmap.
The correlation heatmap is depicted in Fig. 3.
Upon analyzing the heatmap, it becomes evident that
Structural and Structural Usage exhibit positive cor-
relation, suggesting that one of them can safely be
removed. The remaining attributes, except for Type
Id, display cross-correlation and do not raise any con-
cerns regarding their relevance. Another feature to
look at is Type Id in relation to the Assembly Code
(the target variable). Both of them display some-
what correlations with the other attributes in a similar
fashion, indicating a degree of similarity in behavior.
This similarity raises the possibility of data leakage,
thus motivating the removal of Type Id. Data leakage
occurs when the model is trained using an attribute
that describes the target variable in some manner. In
such cases, the model is prone to making overly op-
timistic predictions. Furthermore, during exploratory
data analysis, it is noted that thirteen different classi-
fications are distributed across 4025 wall instances, as
illustrated in Fig. 4.
The depicted figure illustrates the representation
of different variations of Assembly Codes within the
data set. The Y-axis corresponds to the count of rows,
4
https://www.turing.com/kb/guide-to-principal-
component-analysis
Figure 4: Unequal distribution of assembly codes in the
training data set.
while the X-axis corresponds to the various possible
classifications. Among these classifications, namely
224, 221, 225, 211 and 226, hold dominance with
a combined representation of 3758 instances. Con-
versely, the remaining eight classifications are no-
tably non-existent, accounting for only 267 exam-
ples. This evident imbalance in the data set raises
concerns. Imbalanced data set can result in models
that exhibit substantial bias, making it extremely dif-
ficult or even unattainable for the model to accurately
classify the underrepresented types. To address this
issue stemming from imbalanced data, over-sampling
has been implemented. Over-sampling involves gen-
erating multiple additional entries for the minority
classes, thereby achieving a balanced representation.
An effective approach for this is the the Synthetic Mi-
nority Oversampling Technique (SMOTE) (Chawla
et al., 2002). Through the application of the SMOTE
algorithm, the data set has undergone a substantial ex-
pansion, growing from its original 4025 rows to ap-
proximately 25000 rows.
5 MACHINE LEARNING
APPROACH
In this section, the automated building object clas-
sification system utilizing machine learning is intro-
duced. It features a web application for user interac-
tion and an API for data and model management. Su-
pervised learning algorithms and preprocessing tech-
niques are used to train the classifier. The system’s
performance is evaluated using metrics, and the best
model is then deployed for label predictions on new
data.
5.1 Problem Formulation
In the realm of construction, a variety of build-
ing objects are often encountered, denoted as O =
{o
1
, o
2
, ..., o
n
}. Each object o
i
is distinguished by
a set of features F
i
= { f
1
, f
2
, ..., f
m
} and is assigned
an assembly code C
i
based on a specific classifica-
tion standard. The main goal is to devise a classifier
φ : O → C that assigns each building object o
i
to an
KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval
334

assembly code C
i
, with the aim of minimizing the pre-
diction error E =
∑
n
i=1
L(C
i
, φ(o
i
)). Here, L(C
i
, φ(o
i
))
represents a loss function that quantifies the differ-
ence between the actual assembly code C
i
and the pre-
dicted assembly code φ(o
i
).
5.2 Methodology
To achieve this, supervised machine learning algo-
rithms are proposed for training the classifier φ using
a labeled data set D = {(o
1
, C
1
), (o
2
, C
2
), ..., (o
n
, C
n
)}.
The performance of this classifier is then evaluated
using metrics such as accuracy, balanced accuracy
and F1-score. To further enhance the performance of
the model, PCA is proposed for dimensionality reduc-
tion and SMOTE for addressing data imbalance.
API
Preprocessing
Test Data Training Data
Train Models
Test Models Best Model
BIM Data
Database
BIM Data
Web App
BIM Modeling Tool
3. Request/Return
4. Save
5. Process
1. Export
6. Collect
7. Split (Labeled Data)
9. Use
8. Use
11. Save/Model Retraining
13. Predict Labels
14. Import (Enriched data)
12. New Data (Unlabeled Data)
2. Request/Response
10. Select
BIM Data
BIM Data
Figure 5: Process flow.
5.3 Process Flow
The process flow of the proposed solution is detailed
in Algorithm 1 and visualized in Fig. 5. The solu-
tion employs multi-classification algorithms for accu-
rate predictions. Prior to implementation, BIM data
must be manually extracted from a BIM software tool
such as, Autodesk Revit
5
or Speckel
6
. Post prediction,
the enriched BIM data is manually imported back into
the BIM modeling tool. The solution comprises two
main components: developing a classifier using la-
beled data and utilizing this classifier for predicting
assembly codes. The classification process, depicted
in Fig. 5, unfolds through the following subsections.
5
https://www.autodesk.com/products/revit/
6
https://www.speckel.io/
Result: Enriched CSV file (O
0
, C
pred
, P)
Step 1: Export BIM data D as CSV file(s)
containing objects O (with features F) and labels
C;
Step 2: Load CSV file(s) into web application and
select features F and target C;
Step 3: Preprocess (F) to get preprocessed
features (
˜
F);
Step 4: Train classifier φ using selected machine
learning algorithms on training subset
(
˜
F
train
, C
train
) and evaluate performance using
metrics L, where L = evaluate(φ,
˜
F
test
, C
test
);
Step 5: Predict labels C
pred
and confidence scores
P for new data using best model:
(C
pred
, P) = φ
best
(O
0
);
Algorithm 1: Automated building object classification.
5.3.1 Export BIM Data
1. CSV file(s) are manually exported from the BIM
modeling tool(s), containing the required data
(D). The CSV file(s) should include the features
and labels of the building objects that need to
be classified according to a specific classification
standard. The features can be numerical, alphanu-
merical, unique identifiers, binary indicators, or
categorical attributes. The labels can be assembly
codes or other types of classifications that follow
a standard encoding method.
5.3.2 Web Application Usage
2. The web application (W ), a user-friendly inter-
face, allows users to interact with the system by
selecting features and requesting predictions. It
communicates with the API, a software interme-
diary that connects the web application with the
database and the machine learning models.
3. The web application provides a “Find Features”
button that prompts the presentation of all po-
tential features extracted from the data. Users
can choose the features (F
i
) that are relevant for
the classification task and the target (C
i
) variable
that represents the desired output. For example,
“Area”, “Length”, “Width”, “Structural Usage”,
etc. can be selected as features and “Assembly
Code” as target.
4. Data (D) containing the chosen features and tar-
get is stored in a database, which is a structured
collection of data that can be accessed and manip-
ulated by the system. The database ensures that
the data is organized and secure.
− A classifier (φ) is created using selected ma-
chine learning algorithms (Random Forest
(RF), Gradient Boosting (GB), and K-Nearest
Automated Classification of Building Objects Using Machine Learning
335

Neighbors (KNN)). These algorithms are able
to learn from labeled data and make predictions
for new data. The subsequent preprocessing
phase is initiated when the user activates the
“Create Classifier” button.
5.3.3 Preprocessing Data
5. The preprocessing phase is automated, eliminat-
ing the need for manual involvement. This phase
retrieves the data (D) containing the previously
chosen features and target from the database.
− Data (D) is cleaned by handling missing or
duplicate values, outliers and inconsistencies.
During this phase appropriate methods are also
applied to handle any errors or anomalies in the
data. For example, missing or duplicate val-
ues can be removed or imputed, outliers can be
detected or treated and inconsistencies can be
resolved or corrected.
− Data (D) is scaled or normalized using appro-
priate methods to transform the values of the
features into a common range. This helps to
reduce the effect of different units or magni-
tudes on the performance of the machine learn-
ing models. For example, scaling or normal-
ization methods can include min-max scaling,
standardization or log transformation.
− Categorical data (D) is encoded using appropri-
ate methods to convert categorical features into
numerical values that can be used by the ma-
chine learning models. Categorical features are
those that have a finite number of possible val-
ues that represent categories or classes. For ex-
ample, encoding methods can include label en-
coding, one-hot encoding or ordinal encoding.
− Dimensionality of data (D) is reduced using
PCA (PCA(D)), which is a technique that trans-
forms a large number of correlated features into
a smaller number of uncorrelated components
that capture most of the variance in the data.
This helps to reduce noise and redundancy in
the data and improve computational efficiency
and performance of the machine learning mod-
els.
− Imbalance of data (D) is addressed using
SMOTE (SMOTE(D)), which is a technique
that generates synthetic samples for the minor-
ity classes in the data set to balance their repre-
sentation with the majority classes. This helps
to reduce bias and improve accuracy and gen-
eralization of the machine learning models.
6. Processed data (D) is reinserted into the database
for further use by the system.
5.3.4 Training and Testing Models
7. The system splits data (D) into training and test-
ing subsets (D
train
and D
test
). The training sub-
set is used to train the classifier using the selected
machine learning algorithms. The testing subset
is used to test the classifier and evaluate its per-
formance using evaluation metrics.
8. The classifier (φ) is trained using selected ma-
chine learning algorithms (RF, GB, and KNN)
and the training subset (D
train
). These algorithms
learn from labeled data and make predictions for
new data. The training subset is used to fit the pa-
rameters of the algorithms and optimize the clas-
sifier.
9. The classifier (φ) is tested using the testing subset
(D
test
) and performance is evaluated using met-
rics (accuracy, balanced accuracy, and F1-score).
These metrics measure how well the classifier can
predict the correct labels for new data.
10. The best model is selected based on highest met-
ric scores among the three algorithms. The best
model is the one that can achieve the highest accu-
racy, balanced accuracy and F1-score on the test-
ing subset.
11. The best model is saved in the database for future
use and potential retraining.
5.3.5 Label Prediction
12. New unlabeled data (D
0
) is loaded from a BIM
modeling tool. The “Predict Labels” button can be
used to utilize the classifier for predicting labels
for new data. The new data is a set of building
objects that need to be classified according to a
specific classification standard. The new data is
loaded from a BIM modeling tool as a CSV file.
− The new data (D
0
) is preprocessed using the
same methods as before. The system prepro-
cesses the new data ensuring that it is compati-
ble with the classifier and has the same format
and structure as the training and testing data.
− The best model is loaded from the database.
This model was selected based on the highest
metric scores in the previous step.
13. Labels (C
0
) for new data (D
0
) are predicted using
the best model. The labels are assembly codes
or other types of classifications that follow a spe-
cific classification standard. Confidence scores
for each prediction are also generated, indicating
how confident the model is about its prediction.
The enhanced data set is showcased in Table 2.
KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval
336

Table 2: Snapshot of the predicted assembly codes with probabilities.
Area Base constraint Length Structural Usage Unconnected Height Volume Width Assembly Code Probability
.. .. .. .. .. .. .. .. ..
19.6495 PLAN 03 4975 0 4533 2.7954 145 224 1
1.4679 PLAN 12 1472.4999 0 1110 0.2128 145 225 1
4.6728 PLAN 11 2010 0 2360 0.3738 80 221 1
11.9191 PLAN 01 2992.5000 1 4533 3.5757 300 221 0.8466
0.4799 ANIMAL STABLE 3199.9999 0 150 0.036 75 223 1
2.4478 PLAN 03 270 1 4533 1.3212 540 221 0.7125
13.7785 PLAN 01 3600 0 4538 1.9978 145 224 1
3.81 PLAN 02 2050 0 2000 0.3619 95 225 0.6966
13.8608 PLAN 03 3552 0 4533 2.0098 145 224 1
11.6124 PLAN 05 2727 0 4333 1.5386 132 224 1
22.8693 ROOF 6564.9997 0 3421 5.4886 240 217 0.6333
5.015 PLAN 12 2055 0 2360 0.3009 60 221 1
2.3101 PLAN 04 3650 1 650 0.4435 192 214 0.98
23.3638 PLAN 03 7115.2569 0 4533 3.3188 145 212 1
.. .. .. .. .. .. .. .. ..
− An enriched CSV file containing new data (D
0
),
predicted labels (C
0
) and confidence scores (P)
is generated. This file can be used to update or
enrich the BIM model with accurate and con-
sistent classifications.
14. The enriched CSV file is manually imported into a
BIM modeling tool to complete the classification
process.
By adhering to this process flow, the suggested solu-
tion facilitates the establishment and utilization of a
classifier, allowing automated prediction of assembly
codes within building objects.
6 EXPERIMENTS
Within this section, an evaluation of the models em-
ployed to classify building objects is conducted. The
experiments are carried out using real-world data de-
rived from a building project.
6.1 Setup
For conducting multi-class classification, three super-
vised machine learning algorithms (Random Forest
(RF), Gradient Boosting (GB) and K-Nearest Neigh-
bors (KNN)) were employed for experimentation. RF,
GB and KNN were chosen for their unique strengths
in handling building object classification. RF’s ro-
bustness to overfitting and ability to handle high-
dimensional data make it suitable for large data sets.
GB is known for its high accuracy and provides fea-
ture importance, crucial for understanding influential
characteristics in building object classification. KNN
offers a simple yet effective approach, making no as-
sumptions about the data distribution, which is ben-
eficial when classifying objects with similar features.
Collectively, these algorithms provide a comprehen-
sive and robust approach to the classification task.
The experiments were conducted on the extended
data set comprises 25000 wall objects, each with a list
of over 160 features. The algorithms have run on a
single-node hardware platform with a 8th Generation
Intel Core i7-8565U 1.8 GHz processor, 32GB DDR4
RAM and 1TB SSD. The reported outcomes were de-
rived from running each algorithm 20 times, and the
results were averaged over the best 5 executions.
6.2 Test Results
Table 3 displays the classification model test results.
It is emphasized that accuracy may not be reliable for
imbalanced data sets, hence balanced accuracy and
F1-score are also considered. Accuracy is the ratio
of correct predictions to the total predictions made.
Balanced accuracy represents average recall per class,
while F1-score is the harmonic mean of precision and
recall, with all having an optimal value of 1 and a
minimum value of 0.
Table 3: Evaluation metrics for classification models.
Model Accuracy Balanced accuracy F1-score
RF 0.93 0.87 0.94
GB 0.90 0.83 0.91
KNN 0.82 0.71 0.81
Based on the evaluation metrics presented, the RF
model stands out as the most suitable choice, demon-
strating generally high scores (above 85%), which are
deemed acceptable in building object classification
contexts.
Automated Classification of Building Objects Using Machine Learning
337

7 CONCLUSIONS AND FUTURE
WORKS
In the construction sector, seamless collaboration be-
tween stakeholders, such as architects, engineers, and
builders, is essential to avoid miscommunications
arising from different terminologies. Classification
systems address this by offering a standardized lan-
guage throughout the project life-cycle, from concep-
tion to maintenance. This process involves assigning
unique codes to each object within a BIM model, thus
facilitating accurate quantity evaluations, cost estima-
tions and comprehensive project planning. In this pa-
per, an automated approach for classifying building
objects at a specific type level has been presented, uti-
lizing machine learning algorithms such as Random
Forest, Gradient Boosting, and K-Nearest Neighbors.
The effectiveness of this classification technique was
verified with a real-world data set, showing encour-
aging results. The proposed system, although promis-
ing, has limitations including data quality dependency
and possible inaccuracies due to algorithm assump-
tions. Its scalability and adaptability to other projects
or classification schemes are yet to be confirmed, with
its current evaluation limited to a specific project.
Future research should focus on improving data
quality and feature selection, experimenting with var-
ious machine learning algorithms, optimizing system
scalability and conducting assessments across a range
of projects and classification schemes.
ACKNOWLEDGEMENTS
We sincerely thank Jan Buthke and the team at LINK
Arkitektur for their significant contribution to this pa-
per, providing real data set that greatly enriched the
quality and relevance of our research.
REFERENCES
Amor, R. and Dimyadi, J. (2021). The promise of au-
tomated compliance checking. Developments in the
built environment, 5:100039.
Bassier, M., Vergauwen, N., and Genechten, B. V. (2017).
Automated classification of heritage buildings for as-
built bim using machine learning techniques. In IS-
PRS Annals of the photogrammetry, remote sensing
and spatial information sciences, pages 25–30.
Belsky, M., Sacks, R., and Brilakis, I. (2016). Seman-
tic enrichment for building information modeling.
Computer-Aided Civil and Infrastructure Engineer-
ing, 31(4):261–274.
Bloch, T. and Sacks, R. (2019). Comparing machine learn-
ing and rule-based inferencing for semantic enrich-
ment of bim models. Automation in Construction,
91:256–272.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer,
W. P. (2002). Smote: synthetic minority over-
sampling technique. Journal of artificial intelligence
research, 16:321–357.
Emunds, C., Pauen, N., Richter, V., Frisch, J., and
van Treeck, C. (2022). Sparse-bim: classification
of ifc-based geometry via sparse convolutional neu-
ral networks. Advanced Engineering Informatics,
53:101641.
Flager, F. and Haymaker, J. (2007). A comparison of mul-
tidisciplinary design, analysis and optimization pro-
cesses in the building construction and aerospace in-
dustries. In 24th International Conference on Infor-
mation Technology in Construction, pages 625–630.
Fr
´
ıas, E., Pinto, J., Sousa, R., Lorenzo, H., and D
´
ıaz-
Vilari
˜
no, L. (2022). Exploiting bim objects for syn-
thetic data generation toward indoor point cloud clas-
sification using deep learning. Journal of Computing
in Civil Engineering, 36(6):04022032.
Kim, J., Song, J., and Lee, J. K. (2019). Recognizing
and classifying unknown object in bim using 2d cnn.
In 18th International Conference on Computer-Aided
Architectural Design Futures, pages 47–57. Springer.
Koo, B., Jung, R., , and Yu, Y. (2022). Automatic classi-
fication of wall and door bim element subtypes using
3d geometric deep neural networks. Advanced Engi-
neering Informatics, 47:101200.
Luo, H., Gao, G., Huang, H., Ke, Z., Peng, C., and
Gu, M. (2022). A geometric-relational deep learn-
ing framework for bim object classification. In Com-
puter Vision–ECCV 2022 Workshops, pages 349–365.
Springer.
Sacks, R., Ma, L., Yosef, R., Borrmann, A., Daum, S.,
and Kattel, U. (2021). Semantic enrichment for
building information modeling: Procedure for com-
piling inference rules and operators for complex ge-
ometry. Journal of Computing in Civil Engineering,
31(6):04017062.
Salama, D. M. and El-Gohary, N. M. (2011). Semantic
modeling for automated compliance checking. In In-
ternational Conference on Computing in Civil Engi-
neering, pages 641–648. ASCE Library.
Wu, J. and Zhang, J. (2019). New automated bim ob-
ject classification method to support bim interoper-
ability. Journal of Computing in Civil Engineering,
33(5):04019033.
Zabin, A., Gonz
´
alez, V. A., Zou, Y., and Amor, R. (2022).
Applications of machine learning to bim: A system-
atic literature review. Advanced Engineering Infor-
matics, 51:101474.
Zhang, J. and El-Gohary, N. (2012). Extraction of construc-
tion regulatory requirements from textual documents
using natural language processing techniques. In In-
ternational Conference on Computing in Civil Engi-
neering, pages 453–460. ASCE Library.
KDIR 2023 - 15th International Conference on Knowledge Discovery and Information Retrieval
338