Unsupervised Partial Domain Adaptation for Occupants Behavior

Modeling in Smart Buildings

Jawher Dridi

, Manar Amayri

and Nizar Bouguila

Concordia Institute for Information Systems Engineering (CIISE),Concordia University, Canada

{jawher.dridi, manar.amayri, nizar.bouguila}@concordia.ca

Keywords:

Activity Recognition, Occupancy Estimation, Smart Buildings, Domain Adaptation, Unsupervised Learning,

Deep Learning, Machine Learning, Sensor Data.

Abstract:

Smart buildings rely on activity recognition (AR) and occupancy estimation (OE) tasks to provide residents

with several services such as optimal energy management, HVAC (Heating, ventilation, and air conditioning)

systems optimization, and security. Estimating the number of occupants and recognizing their activities is

performed using sensor data which is scarce. The collection and labeling of smart building data are tedious,

costly, and time-consuming, pushing researchers to consider solutions based on domain adaptation (DA) to

transfer knowledge from source domains where data is abundant to target domains where data is scarce. In

particular, unsupervised domain adaptation (UDA) has been considered to solve the unavailability of labeled

data in target domains. Previous research has focused on standard UDA methods where label space is identical

between source and target domains which is not the case for real-world datasets. This work considers unsuper-

vised partial domain adaptation (UPDA) methods where target classes are a subset of source classes. We adapt

and evaluate two UPDA techniques called Adversarial Re-weighting for Partial Domain Adaptation (ARPDA)

and Selective Adversarial Networks for Partial Domain Adaptation (SAN w PDA). We have compared their

performance to Adversarial Re-weighting for Standard Domain Adaptation (ARSDA) and Selective Adversar-

ial Networks for Standard Domain Adaptation (SAN w SDA) as well as several previous UDA methods. The

impressive results with scores up to 98% prove the efﬁciency of the adapted UPDA techniques. We provide

the code in the following repository: https://github.com/JawDri/UPDA-for-OE-and-AR.git.

1 INTRODUCTION

Smart buildings (Kazmi et al., 2017), powered by

the Internet of Things (IoT) and machine learning

(Dridi et al., 2022), offer several advantages for resi-

dents which help enhance life conditions and reduce

bills. 0E and AR are among the most interesting

smart building tasks that can help provide optimal en-

ergy management, HVAC (Heating, ventilation, and

air conditioning) systems optimization, and security

(Dridi et al., 2023a; Dridi et al., 2022; Prabhakaran

et al., 2022; Dridi et al., 2023b). Indeed, estimating

the number of occupants in different areas can be used

to distribute energy optimally across the building and

reduce energy waste in unoccupied places (Zamzami

et al., 2019). HVAC systems can also be optimized

by adjusting heating, ventilation, and air conditioning

based on the number of occupants and their activities

https://orcid.org/0000-0001-6062-2897

https://orcid.org/0000-0002-5610-8833

https://orcid.org/0000-0001-7224-7940

in a particular place (Dridi et al., 2023a). Recognizing

activities in buildings helps provide more security for

residents by identifying unauthorized actions using

sensor data (Dridi et al., 2022). Smart building data,

in particular labeled data, is scarce and hard to collect

due to several factors such as cost, privacy, and time

(Dridi et al., 2023b). All these issues have pushed re-

searchers to consider unsupervised domain adaptation

(UDA) methods that collect knowledge from source

domains where labeled data is available and transfer

it to target domains where labeled data is unavailable

(Dridi et al., 2023b). Researchers aim, by sharing

knowledge across domains, to solve data scarcity is-

sues and to enhance target model performances. Pre-

vious works on UDA have considered the same la-

bel space between source and target domains while

sharing knowledge which is not the case in most real-

world scenarios. Indeed, collected datasets may con-

tain different or related classes and not necessarily the

same labels which can lead to negative transfer while

sharing knowledge across domains. In this work, we

Dridi, J., Amayri, M. and Bouguila, N.

Unsupervised Partial Domain Adaptation for Occupants Behavior Modeling in Smart Buildings.

DOI: 10.5220/0013073400003953

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2025), pages 69-76

ISBN: 978-989-758-751-1; ISSN: 2184-4968

adapt partial UDA methods to share knowledge from

source to target domains where target labels are a sub-

set of source classes. The considered UPDA tech-

niques reduce the effect of negative transfer caused

by the source labels. The adapted methods are Ad-

versarial Re-weighting for Partial Domain Adaptation

(ARPDA) (Gu et al., 2021) and Selective Adversar-

ial Networks for Partial Domain Adaptation (SAN w

PDA) (Cao et al., 2018). We have compared their

performance to Adversarial Re-weighting for Stan-

dard Domain Adaptation (ARSDA) (Gu et al., 2021)

and Selective Adversarial Networks for Standard Do-

main Adaptation (SAN w SDA) (Cao et al., 2018)

as well as several previous UDA methods. Adver-

sarial Re-weighting is based on adversarial learning

and it learns the contribution of each source sample

to the update of the target model networks by assign-

ing them source weights. By re-weighting source do-

main data, it mitigates negative transfer, then it aligns

source and target data distributions by minimizing

a Wasserstein distance. Adversarial Re-weighting

has been evaluated with partial and standard UDA

(ARPDA and ARSDA). SAN (Selective Adversar-

ial Networks) is also based on adversarial learning,

it eliminates source samples with outlier labels and

encourages samples with shared labels, to mitigate

negative transfer while reducing discrepancy between

source and target domains. SAN has been evalu-

ated with partial and standard UDA (SAN w PDA

and SAN w SDA). The adapted methods have been

evaluated on smart buildings datasets for several AR

and OE tasks (Dridi et al., 2023a). This research

has several contributions as follows. It is the ﬁrst to

adapt ARPDA, ARSDA, and SAN approaches from

2-dimensional space to 1-dimensional and evaluate

them with partial and standard UDA. It has provided

new architectures for the features extractor, classi-

ﬁer, and discriminator modules that ﬁt IoT data. The

newly adapted approaches can be applied to any 1-

D data and are not restricted to smart building data.

Partial UDA methods that have been adapted solve a

real issue related to negative transfer which is com-

mon in smart buildings data. A comparison analy-

sis between the ﬁndings of the adapted methods with

partial and standard UDA as well as previous UDA

methods. The adapted UPDA methods have outstand-

ing performances with scores up to 98%. The rest of

the paper is divided into 3 sections. In section 2, we

present some works related to OE, AR, and partial

DA. In section 3, we explain the adapted partial UDA

methods: ARPDA and SAN w PDA. In section 4, we

present the experimental setup and discuss the results.

2 LITERATURE REVIEW

AR and OE tasks, based on smart building data, can

contribute to the generation of several advantages

such as energy management (Dridi et al., 2023a). Do-

main adaptation solves the problem of data scarcity

which is common in smart buildings by sharing

knowledge across domains (Dridi et al., 2023b). Sev-

eral works have been done on AR, OE, and DA.

2.1 Occupancy Estimation (OE)

Occupancy estimation (Amayri et al., 2019), is the

task of counting people in an area such as a room,

apartment, or building. Several works have been

done to predict the number of occupants using dif-

ferent types of smart building data. (Chen et al.,

2017) has developed an OE method based on hid-

den Markov models (HMMs) and logistic regression

to optimize HVAC systems and ensure safety within

buildings. The approach is called an inhomogeneous

hidden Markov model with multinomial logistic re-

gression (IHMM-MLR), and it uses environmental

sensors (Chen et al., 2017), such as humidity and CO2

concentration sensors, to collect the required data.

2.2 Activity Recognition (AR)

Activity recognition (Ali and Bouguila, 2020), is a

smart building task that aims to understand the be-

havior and actions of people by recognizing their ac-

tivities. AR is beneﬁcial by providing several advan-

tages such as security and HVAC systems optimiza-

tion. (Wang et al., 2016) has used smartphone in-

ertial sensors, such as gyroscopes, to collect needed

data for AR. The choice of smartphone sensors is due

to their low cost (Wang et al., 2016). Activities of

people have been recognized using collected data and

machine learning classiﬁers such as the Na

ıve Bayes

classiﬁer.

2.3 Partial Domain Adaptation (PDA)

Domain adaptation (Dridi et al., 2023b), is a tool

used to solve data scarcity issues by sharing knowl-

edge from domains where data is available to other

domains where data is scarce. In smart buildings,

data scarcity is a common issue that has pushed

researchers to employ domain adaptation methods.

Most research focuses on standard domain adaptation

where source and target domains share the same label

space which is not the case with real-world datasets.

Collected smart building data contains related labels

and a few common labels between domains. (Li

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems

Figure 1: Model and discriminator architectures for UPDA approaches.

et al., 2020) has developed a PDA method called Deep

Residual Correction Network (DRCN) which reduces

the negative transfer created by outlier source labels.

In this research, we aim to estimate the number

of occupants in buildings and recognize their activi-

ties. Since smart building data is scarce, we use do-

main adaptation methods, particularly unsupervised

approaches. Commonly, real-world datasets do not

share the same label space, but they may share a sub-

set of classes. Therefore, we consider partial domain

adaptation to mitigate the effect of negative transfer

created by outlier source labels. We consider Un-

supervised Partial Domain Adaptation for Estimat-

ing Occupancy and Recognizing Activities in Smart

Buildings.

3 METHODS

In this work, we consider partial UDA methods that

deal with real-world scenarios where datasets con-

tain a common label subspace and multiple unre-

lated labels. Let us consider a labeled source do-

main S = {x

, y

}

i=1

and an unlabeled target domain

T = {x

}

i=1

, with n

and n

the number of source

and target samples, respectively. Let Y

and Y

the source and target label spaces, respectively, where

⊂ Y

. Figure 1 gives the newly created architec-

tures for the feature extractor, classiﬁer, and discrim-

inator modules of the adapted methods.

3.1 Adversarial Re-Weighting for

Partial Domain Adaptation

(ARPDA)

ARPDA applies a feature transformation using a fea-

ture extractor F, then it re-weights the source samples

using adversarial learning based on their importance

for the target label space (Gu et al., 2021). ARPDA

applies a reweighted cross-entropy objective using the

generated weights of source data, and a conditional

entropy objective to update the target model (the fea-

ture extractor F and the classiﬁer C) to allow knowl-

edge transfer across the domains (Gu et al., 2021).

The overall objective is a combination of a target

domain conditional entropy and a source domain re-

weighted cross-entropy losses as deﬁned in Eq.(1).

L(θ

, θ

, W ) =

∑

i=1

J (C(F(x

;θ

);θ

), y

∑

j=1

H(C(F(x

;θ

);θ

))

(1)

where H(·) is a conditional entropy, J (·, ·) is a

cross-entropy objective, (θ

, θ

) are the parameters

of the model, and W = {w

}

i=1

is a weights vector

of source samples (Gu et al., 2021). The conditional

entropy loss encourages the separation of classes, and

the re-weighted cross-entropy loss predicts the classes

of input data (Gu et al., 2021). The Wasserstein

distance is used to measure the relatedness between

source and target samples in order to re-weigh source

instances based on their importance. The shared label

space between source and target domains is supposed

to have close data distribution (Gu et al., 2021).

3.2 Selective Adversarial Networks for

Partial Domain Adaptation (SAN w

PDA)

SAN w PDA (Cao et al., 2018), is an UPDA method

that combines both the reduction of negative transfer

caused by outlier source samples and the promoting

of positive transfer generated by the rest of source

data. It is based on adversarial learning (Cao et al.,

2018) which has been used in several research that

tackles standard UDA where the source and target la-

bel spaces are the same. For a typical standard UDA

based on adversarial learning the objective deﬁned in

Eq.(2) can be an example.

(θ

, θ

) =

∑

∈S

)), y

)−

+ n

∑

∈(S

T )

), d

(2)

where G

is a feature extractor with parameters

, G

is a classiﬁer with parameters θ

, G

is a

Unsupervised Partial Domain Adaptation for Occupants Behavior Modeling in Smart Buildings

discriminator with parameters θ

, λ is a balancing-

parameter, L

is a label prediction loss, L

is a domain

discriminator loss, and d

is a domain label. For par-

tial UDA, the domain discriminator’s objective is up-

graded by assigning a discriminator G

for each of all

the target labels (K). Since the target data is unlabeled

it is not evident to assign each discriminator G

to tar-

get input data x

. Therefore, pseudo-labels ˆy

= G

)

are given to each data sample using the source knowl-

edge (Cao et al., 2018). A new probability-weighted

domain discriminator objective is deﬁned in Eq.(3).

′

+ n

∑

k=1

∑

∈(S

T )

ˆy

), d

), (3)

where d

is a domain label, L

is a cross-entropy

objective, and G

is a domain discriminator module

for the k-th source class (Cao et al., 2018). The con-

sidered multi-discriminator domain adversarial net-

work reduces negative transfer and promotes positive

transfer by aligning each data sample with data points

that are close or with the same classes, and by us-

ing probability weights for each domain discrimina-

tor which ﬁlters unrelated data sample classes (Cao

et al., 2018). The proposed objective can be further

enhanced by enhancing the positive transfer. The loss

deﬁned in Eq.(4) reduces further the weights of the

domain discriminators of outlier source classes (Cao

et al., 2018).

′

+ n

∑

k=1

(

∑

∈T

ˆy

)∗

∑

∈(S

T )

ˆy

), d

(4)

where

∑

∈T

ˆy

represents the weights for each

label k which is large for common source and tar-

get classes and small for outlier labels (Cao et al.,

2018). Since the reduction of negative transfer and

the outliers ﬁltration depends heavily on ˆy

= G

we introduce a conditional-entropy H(·) as in (Gu

et al., 2021) for further performance enhancement.

The overall objective function of SAN w PDA is de-

ﬁned in Eq.(5).

C(θ

, θ

) =

∑

∈S

)), y

∑

∈T

H(G

)))−

+ n

∑

k=1

(

∑

∈T

ˆy

)

∑

∈(S

T )

ˆy

), d

(5)

where λ is a balancing parameter.

4 EXPERIMENTAL SETUP AND

RESULTS

4.1 Experimental Setup

For OE, we used our private datasets (Amayri and

Ploix, 2018) collected in two ofﬁces at Grenoble In-

stitute of Technology. For USDA, we considered

3 levels of occupancy: no occupant, one occupant,

and two occupants. For UPDA, we considered 5

levels of occupancy by adding three occupants and

four occupants levels (Amayri and Ploix, 2018). The

datasets have been collected using ambient sensors

such as power consumption sensors (Amayri and

Ploix, 2018). For AR, we used the Washington State

University (WSU) Center for Advanced Studies in

Adaptive Systems (CASAS) datasets (Cook, 2010).

They have been collected using ambient sensors such

as door contact sensors (Cook, 2010). For USDA, we

considered 5 activities: preparing breakfast, prepar-

ing lunch, preparing dinner, watching TV, and toilet-

ing. However, for UPDA, we considered 7 activities

by adding bathing and sleeping out-of-bed activities

(Cook, 2010). We have used accuracy as a metric

for balanced datasets and F1 score as a metric for

unbalanced datasets (Dridi et al., 2023a; Dridi et al.,

2022; Dridi et al., 2023b). Table 1 shows some of the

values of the used parameters in the adapted meth-

ods: number of epochs, batch size, optimizer, learn-

ing rate, gamma value linked with the learning rate,

weight decay for L2 penalization, and momentum.

We have compared the obtained scores from UDA

methods with a supervised machine learning method

(SMLM) which is a decision tree classiﬁer trained and

evaluated on target data. SMLM is considered as a

reference in this research to evaluate the efﬁciency of

the adapted methods compared to supervised learning

methods.

4.2 Experimental Results

4.2.1 5-label AR

5-label AR is the task of recognizing 5 human activi-

ties in buildings. For standard UDA, where the source

and target label spaces are the same, the considered

activities are: toileting, watching TV, cooking break-

fast, lunch, and dinner. For partial UDA, where the

target label space is a part of the source label space,

the added two activities to the source domain: bathing

and sleeping. Table 2 gives the obtained scores for

the adapted UPDA and USDA techniques, supervised

machine learning methods (SMLM) as well as pre-

vious research on UDA (Dridi et al., 2023a). AR-

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems

Table 1: Parameters used in the implemented methods.

Parameter ARSDA ARPDA SAN w SDA SAN w PDA

epochs 1000 1000 1000 1000

batch-size 36 36 36 36

optimizer SGD SGD SGD SGD

lr 1e-3 1e-3 1e-2 1e-2

gamma 1e-3 1e-3 1.0 1.0

momentum 0.9 0.9 0.9 0.9

weight-decay 5e-4 5e-4 5e-4 5e-4

Table 2: AR accuracy for 5 balanced classes and F1 score for 5 unbalanced classes.

Method Accuracy (%) F1-score (%)

ARSDA 84.67 73.98

ARPDA 67.33 62.32

SAN w SDA 78.00 63.79

SAN w PDA 68.13 54.16

SMLM 99.33 98.66

DSN (Dridi et al., 2023a) 20.66 23.12

CAT (Dridi et al., 2023a) 38.00 55.64

CAT+RevGrad (Dridi et al., 2023a) 49.00 51.15

CoWA-JMDS (Dridi et al., 2023a) 30.68 18.20

CoWA-JMDS w/o WM (Dridi et al., 2023a) 24.40 10.85

DaC (Dridi et al., 2023a) 56.08 36.65

AaD (Dridi et al., 2023a) 33.19 14.67

SDA, which is a standard UDA method where source

and target label spaces are the same, has the best per-

formance for both balanced and unbalanced datasets

with 84.67% accuracy and 73.98% F1 score. The de-

crease in performance with unbalanced datasets is ex-

pected (Dridi et al., 2023a; Dridi et al., 2022; Dridi

et al., 2023b) since label proportions change leads

to performance degradation in most cases. ARPDA,

which is a partial UDA version of ARSDA where the

source domain has more labels than the target domain,

has a great performance for both balanced and unbal-

anced datasets but less than the performance of AR-

SDA. The decrease in performance is expected since

ARPDA faces the challenge of negative transfer cre-

ated by outlier labels such as bathing. With unbal-

anced datasets, we see a further decrease in perfor-

mance which is expected for different label propor-

tions with 62.32% of F1 score compared to balanced

datasets with 67.33% of accuracy. The performance

of ARPDA is good when compared to supervised ma-

chine learning method performance (SMLM). Also,

ARPDA has exceeded several previous standard UDA

methods such as DSN (Dridi et al., 2023a) and CAT

(Dridi et al., 2023a). The great performance is thanks

to adversarial learning that helped reduce the effect

of outlier source samples. SAN w PDA has also

given great performance comparable to ARPDA with

68.13% of accuracy for balanced label proportions

and 54.16% of F1 score for unbalanced datasets.

SAN w SDA, where source and target label space are

the same, has better performance than SAN w PDA

which is expected since we are not facing the chal-

lenge of negative transfer. SAN w PDA has exceeded

several standard UDA methods (Dridi et al., 2023a)

which is a great achievement for a method that has

the challenge of negative transfer.

4.2.2 3-Label AR

3-label AR is the task of recognizing 3 activities that

are common between source and target domain for

standard UDA. The activities are: toileting, watch-

ing TV, and cooking dinner. For partial UDA, we add

outlier labels for the source domain (cooking break-

fast and lunch). Table 3 gives all the obtained scores.

Compared to the supervised machine learning method

(SMLM) which trains and evaluates a classiﬁer us-

ing labeled target data, the adapted standard and par-

tial UDA have shown an outstanding performance.

ARPDA has an accuracy of 94% which is a bit lower

than the standard UDA method (ARSDA), and this

proves the efﬁciency of adversarial learning to remove

negative transfer created by outlier source data. SAN

w PDA has also given outstanding performance ex-

ceeding ARPDA (95.99%) which is also a bit lower

Unsupervised Partial Domain Adaptation for Occupants Behavior Modeling in Smart Buildings

Table 3: AR accuracy for 3 balanced classes and F1 score for 3 unbalanced classes.

Method Accuracy (%) F1-score (%)

ARSDA 95.33 97.97

ARPDA 94.00 97.98

SAN w SDA 98.67 97.98

SAN w PDA 95.99 97.99

SMLM 100 100

DSN (Dridi et al., 2023a) 39.00 45.54

CAT (Dridi et al., 2023a) 78.00 80.00

CAT+RevGrad (Dridi et al., 2023a) 87.00 75.79

CAT+rRevGrad (Dridi et al., 2023a) 65.50 80.68

ATDOC+NC (Dridi et al., 2023a) 84.00 87.35

CoWA-JMDS (Dridi et al., 2023a) 65.07 46.79

CoWA-JMDS w/o WM (Dridi et al., 2023a) 58.00 33.28

DaC (Dridi et al., 2023a) 81.93 79.78

AaD (Dridi et al., 2023a) 58.09 25.18

SHOT-IM (Dridi et al., 2023a) 88.80 93.31

SHOT-Pseudo-labeling (Dridi et al., 2023a) 78.80 87.70

than SAN w SDA (98.67%). For unbalanced label

proportions, all the adapted standard and partial UDA

methods have almost the same performance with out-

standing scores around 98%. The excellent F1 scores

are greater than the performance of balanced datasets

which can be due to the additional information gained

about the labels’ proportion difference as explained

before (Dridi et al., 2023a; Dridi et al., 2022; Dridi

et al., 2023b). The adapted UPDA methods have the

advantage of excessively reducing the effect of neg-

ative transfer which can be seen by the close perfor-

mance with USDA methods, and they have the ad-

vantage of exceeding multiple previous USDA meth-

ods with a large performance gap such as DaC (Dridi

et al., 2023a).

4.2.3 3-Label OE

3-label OE is the task of predicting 3 levels of oc-

cupancy for standard UDA which are no occupant,

one occupant, and two occupants. For partial UDA

methods, we add 2 levels which are 3 occupants and

4 occupants levels. Table 4 gives all the obtained

scores for current tasks. ARPDA and SAN w PDA

have shown very good performances for estimating

the number of occupants even with outlier samples

of source domains. Thanks to promoting the positive

transfer, SAN w PDA has 57.33% accuracy and 74.55

F1 score for balanced and unbalanced datasets, re-

spectively. There is a remarkable drop in performance

for SAN w PDA compared to SAN w SDA which is

expected due to the effect of source outlier labels. The

increase in performance with unbalanced datasets is

thanks to gathered information about label proportion

differences as explained in (Dridi et al., 2023a; Dridi

et al., 2022; Dridi et al., 2023b). ARPDA has ex-

ceeded SAN w PDA with very good performances for

both balanced and unbalanced datasets. ARPDA with

61.33% accuracy and 69.87% F1 score has exceeded

multiple USDA methods such as CoWA-JMDS (Dridi

et al., 2023a) even though it is a UPDA method, and it

has close performance to ARSDA where source and

target label spaces are the same. For partial UDA,

ARPDA has the best performance for the current sce-

nario thanks to the efﬁcient use of adversarial learning

to reduce the effect of negative transfer created by out-

lier source data. The excellent performance obtained

for the unbalanced datasets with partial and unsuper-

vised DA methods proves the efﬁciency of the consid-

ered techniques that can overcome the combination of

several challenges and provide very good results.

Figure 2: SAN results for balanced and unbalanced datasets

with standard and partial domain adaptation.

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems

Table 4: OE accuracy for 3 balanced classes and F1 score for 3 unbalanced classes.

Method Accuracy (%) F1-score (%)

ARSDA 76.67 82.91

ARPDA 61.33 69.87

SAN w SDA 71.33 79.15

SAN w PDA 57.33 74.55

SMLM 94.00 91.93

DSN (Dridi et al., 2023a) 34.80 57.01

CoWA-JMDS (Dridi et al., 2023a) 40.59 54.72

CoWA-JMDS w/o WM (Dridi et al., 2023a) 49.05 55.51

Table 5: OE accuracy for 2 balanced classes and F1 score for 2 unbalanced classes.

Method Accuracy (%) F1-score (%)

ARSDA 94.00 87.73

ARPDA 85.33 88.47

SAN w SDA 92.67 92.84

SAN w PDA 86.00 88.58

SMLM 96.66 95.30

DSN (Dridi et al., 2023a) 56.00 63.87

ATDOC+NC (Dridi et al., 2023a) 82.65 86.19

CoWA-JMDS (Dridi et al., 2023a) 76.55 72.31

CoWA-JMDS w/o WM (Dridi et al., 2023a) 85.88 72.31

4.2.4 2-Label OE

2-label OE is the task of estimating two levels of

occupants for standard UDA: no occupant and one

occupant. For UPDA, we add an outlier level for

source data which is: the 2 occupants level. Table

5 illustrates the performances of current and previ-

ous methods. All the adapted methods for standard

and partial UDA have given excellent scores that are

so close to the performance of SMLM which is a

great achievement. ARSDA has an accuracy of 94%

that exceeds ARPDA by around 8.5% which is ex-

pected due to the effect of negative transfer. SAN w

SDA has an accuracy of 92.67% that exceeds SAN w

PDA by around 6.5%. For unbalanced datasets, we

see a small increase in performance for most methods

which is explained by the additional information pro-

vided by the label proportion changes as explained in

(Dridi et al., 2023a; Dridi et al., 2022; Dridi et al.,

2023b). Overall the scores are excellent and they

have exceeded several previous standard UDA meth-

ods such as ATDOC+NC (Dridi et al., 2023a). Ob-

taining scores greater than 90% even the challenges

of negative transfer, unlabeled data, and unbalanced

label proportions, prove the efﬁciency of the consid-

ered methods.

4.3 Discussion

In this section, we chose the SAN method to discuss

further because it has the best performances for both

partial and standard UDA on average. Figure 2 gives

a graphic illustration of the obtained scores for both

SAN w PDA and SDA for all smart building tasks.

In the ﬁrst view, we notice that all methods follow

the same trend for the different tasks with small dif-

ferences in performance for each smart building task.

It is clear that standard UDA has better performance

than partial UDA for all the methods and tasks which

is expected since UPDA methods face the challenge

of negative transfer created by outlier source labels.

The conclusions can be seen by the fact that USDA

charts are above UPDA charts for both balanced and

unbalanced scenarios. Also, it is clear that the de-

crease in task complexity increases performance such

as moving from 5-label AR to 3-label AR which is

expected as explained before in (Dridi et al., 2023b).

Adapted methods performances for UPDA are greater

than 90% for multiple scenarios which is an excel-

lent achievement for the current research that pushes

these techniques to real-world applications. The low-

est score is around 55% which is acceptable for the

current research that deals with data scarcity, unla-

beled target data, negative transfer, and unbalanced

label proportions.

Unsupervised Partial Domain Adaptation for Occupants Behavior Modeling in Smart Buildings

5 CONCLUSION

In conclusion, partial UDA methods have been

adapted and evaluated on AR and OE tasks aiming

to provide energy management, security, and HVAC

systems optimization for smart buildings. This work

has several contributions. Indeed, it is the ﬁrst to

adapt ARPDA, ARSDA, and SAN approaches from

2-dimensional space to 1-dimensional and evaluate

them with partial and standard UDA. This research

has provided new architectures for the features extrac-

tor, classiﬁer, and discriminator modules that ﬁt IoT

data. The newly adapted approaches can be applied to

any 1-D data and are not restricted to smart building

data. Partial UDA methods that have been adapted

solve a real issue related to negative transfer which is

common in smart buildings data. Also, a comparison

analysis between the ﬁndings of the adapted meth-

ods with partial and standard UDA as well as previ-

ous UDA methods. The adapted UPDA methods have

outstanding performances with scores up to 98%. In

future work, we consider UPDA methods applied to

smart building tasks with more outlier labels in source

domains and compare the ﬁndings with the current re-

search.

REFERENCES

Ali, S. and Bouguila, N. (2020). Online learning for

beta-liouville hidden markov models: Incremental

variational learning for video surveillance and action

recognition. In 2020 IEEE International Conference

on Image Processing (ICIP), pages 3249–3253. IEEE.

Amayri, M. and Ploix, S. (2018). Decision tree and

parametrized classiﬁer for estimating occupancy in

energy management. In 2018 5th International Con-

ference on Control, Decision and Information Tech-

nologies (CoDIT), pages 397–402. IEEE.

Amayri, M., Ploix, S., Najar, F., Bouguila, N., and Wurtz,

F. (2019). A statistical process control chart ap-

proach for occupancy estimation in smart buildings.

In 2019 IEEE Symposium Series on Computational

Intelligence (SSCI), pages 1729–1734. IEEE.

Cao, Z., Long, M., Wang, J., and Jordan, M. I. (2018).

Partial transfer learning with selective adversarial net-

works. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 2724–

2732.

Chen, Z., Zhu, Q., Masood, M. K., and Soh, Y. C. (2017).

Environmental sensors-based occupancy estimation in

buildings via ihmm-mlr. IEEE Transactions on Indus-

trial Informatics, 13(5):2184–2193.

Cook, D. J. (2010). Learning setting-generalized activity

models for smart spaces. IEEE intelligent systems,

2010(99):1.

Dridi, J., Amayri, M., and Bouguila, N. (2022). Transfer

learning for estimating occupancy and recognizing ac-

tivities in smart buildings. Building and Environment,

217:109057.

Dridi, J., Amayri, M., and Bouguila, N. (2023a). Unsuper-

vised domain adaptation with and without access to

source data for estimating occupancy and recognizing

activities in smart buildings. Building and Environ-

ment, 243:110651.

Dridi, J., Amayri, M., and Bouguila, N. (2023b). Unsuper-

vised domain adaptation without source data for esti-

mating occupancy and recognizing activities in smart

buildings. Energy and Buildings, page 113808.

Gu, X., Yu, X., Sun, J., Xu, Z., et al. (2021). Adversarial

reweighting for partial domain adaptation. Advances

in Neural Information Processing Systems, 34:14860–

14872.

Kazmi, H., Mehmood, F., and Amayri, M. (2017). Smart

home futures: Algorithmic challenges and opportu-

nities. In 2017 14th International Symposium on

Pervasive Systems, Algorithms and Networks & 2017

11th International Conference on Frontier of Com-

puter Science and Technology & 2017 Third Inter-

national Symposium of Creative Computing (ISPAN-

FCST-ISCC), pages 441–448. IEEE.

Li, S., Liu, C. H., Lin, Q., Wen, Q., Su, L., Huang, G., and

Ding, Z. (2020). Deep residual correction network for

partial domain adaptation. IEEE transactions on pat-

tern analysis and machine intelligence, 43(7):2329–

2344.

Prabhakaran, K., Dridi, J., Amayri, M., and Bouguila, N.

(2022). Explainable k-means clustering for occupancy

estimation. Procedia Computer Science, 203:326–

333.

Wang, A., Chen, G., Yang, J., Zhao, S., and Chang, C.-Y.

(2016). A comparative study on human activity recog-

nition using inertial sensors in a smartphone. IEEE

Sensors Journal, 16(11):4566–4578.

Zamzami, N., Amayri, M., Bouguila, N., and Ploix, S.

(2019). Online clustering for estimating occupancy

in an ofﬁce setting. In 2019 IEEE 28th International

Symposium on Industrial Electronics (ISIE), pages

2195–2200. IEEE.

SMARTGREENS 2025 - 14th International Conference on Smart Cities and Green ICT Systems