Pet Adoption Status Prediction Based on Multiple

Machine Learning Models

Honglin Lu

Computer Science, University of Washington, Seattle, U.S.A.

Keywords: Pet Adoption, Machine Learning Models, Artificial Intelligence.

Abstract: It is urgent and challenging to reasonably select machine learning algorithms and models to accurately predict

the adoption status of pets. In this paper, the prediction is a classification task for whether a pet will be

adopted or not which aims to reach the highest possible accuracy. Several models are used to predict the pet

adoption status based on some information in the given dataset on Kaggle and the evaluation is mainly based

on accuracy. For each model, pre-processing includes separating numerical and categorical columns and

dropping useless columns or editing some columns if necessary. Multiple parameters are involved in tests for

each model. The Artificial Neural Network (ANN) is designed to have four layers with the first three layers

using ReLu as activation functions and the last using Sigmoid. Different amounts of middle layer neurons and

epochs are tested in order to select the parameter with the highest accuracy to represent the model. By

comparing the accuracy of each testing result, it indicates the best performing model is the four-layer ANN

model with both the number of neurons in the middle layer and the amount of epoch to be around 60 and the

number of neurons in the first and third layers to be 20.

1 INTRODUCTION

Pets, nowadays, have become an indispensable part

of many people's lives. According to the pet

ownership statistic on Forbus in 2024, 66% of the

U.S. Households own at least one pet, which is about

86.9 million homes (Megna, 2024). Common

domestic pets include dogs, cats, birds, fish, rabbits,

lizards, etc. These pets play important roles in their

owners' lives, most commonly as companions. To

achieve a sense of belonging, the relationship

between pets and people is an important source of

social support and attachment (Hajek, 2024).

Moreover, pets can effectively improve owner’s

health from a variety of aspects, such as reducing

systemic hypertension, lowering hyperlipidemia and

promoting physical activity (Hussein, 2021). Many

professional jobs also require well-trained pets to

complete, such as guide dogs and detection dogs.

With the increasing demand for pets, pet adoption has

gradually become a current affair for pet protection

organizations and adoption agencies and shelters.

Therefore, it is critical and urgent to provide a

reasonable and accurate prediction of whether a pet

https://orcid.org/0009-0004-6168-9523

will be adopted, which can not only enhance the pet

adoption rate, but also provides superior resource

allocations in fundings and food.

As digitalization becomes more prevalent around

the world, Machine Learning (ML) and Artificial

Intelligence (AI) are developing rapidly. In order to

make ML and AI efficient, accurate, flexible and

diverse, various algorithms for intelligent machine

training, for example, random forest, decision tree,

logistic regression, etc., have been gradually

designed, developed, and widely applied in almost

every field. With the help of AI in Computed

Tomography (CT) image diagnosis, case screening,

intelligent analysis and diagnosis, and auxiliary

clinical diagnosis and treatment decision-making can

be more accurately and quickly (Hussein, 2021). In

the field of chemistry, a robust ML model can

expedite the identification of promising candidates

for experimental validation by efficiently predicting

the reaction outcome for any new choice of

substrates/catalyst (Zeng, 2024). Interestingly, in

recent years, there has been a lot of data research and

analysis on pets based on machine learning.

Mohammad et al. applied AI and ML techniques with

362

Lu, H.

Pet Adoption Status Prediction Based on Multiple Machine Learning Models.

DOI: 10.5220/0013331300004558

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

In Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management (MLSCM 2024), pages 362-367

ISBN: 978-989-758-738-2

the biological perspective to predict and classify

canine personalities based on data derived from C-

BARQ database (Amirhosseini, 2024). Seon-Chil and

Sanghyun attempted using AI to predict a dog’s

health conditions based on the data provided by the

sensor attached on dogs, which resulted in an

accuracy of 87.5% with a discrepancy only in beagle

compared to the veterinarians’ analyses (Kim, 2024).

This paper will focus on the prediction of whether the

pet will be adopted or not from the shelter based on

data of some basic information and traits of the pet.

Multiple algorithms will be utilized to predict the

adoption status and further discussion on the

parameters and layers will be addressed.

The data set file is from Kaggle, “Predict Pet

Adoption Status Dataset”. This paper will first briefly

analyze the data set, including some basic

construction of the data and data visualization.

Multiple models, such as Random Forest, Decision

Tree, Logistic Regression, Neural networks with

Rectified Linear Unit and Sigmoid, will be tested with

different values for each parameter, for example,

n_estimators, min_samples_leaf, max_depth,

min_samples_split, etc. and by comparing the general

result of these models, the best option with higher

accuracy and efficiency will be selected with further

modification on the parameters and model structure

to ensure its best performance. Some possible future

developments and applications of this model

approach will also be addressed later.

2 METHOD

2.1 Dataset Preparation

The data information is from Kaggle, which offers an

in-depth overview of the factors that may affect a

pet's chances of being adopted from a shelter. The

data contains a total number of 2, 007 pets, or 2, 007

rows with 13 columns of pet characteristics and

attributes, such as pet type, breed, color, size, etc.

There are a total of four pet types: dog, cat, bird and

rabbit (the amount of each is shown in Figure 1) with

seven different breeds shown in Figure 2.

Figure 1: Pet type counts (Photo/Picture credit: Original).

Besides the pet type and breed, the rest data are

briefly shown in Figure 3 with bar diagrams. The

prediction is designed to be a classification task,

where for each pet, it will be predicted as adopted or

otherwise unadopted.

Before the model constructions and training, data

preprocessing is required for all the testing models. In

order to simplify the data, the first and last columns

which represent the pet ID and pet adoption

likelihood will be dropped and the target, adoption

likelihood, needs to be stored. For all models, 80% of

the data is splitted for training and the rest 20% for

testing; the random state is set to 42. Then, by

separating the numerical and categorical columns,

Figure 2: Breed percentile pie chart (Photo/Picture credit: Original).

Pet Adoption Status Prediction Based on Multiple Machine Learning Models

363

Figure 3: Other basic information data visualization (Photo/Picture credit: Original).

each null value in the numerical column will be filled

with mean and each categorical column will be

converted to a binary column.

2.2 Machine Learning Models-Based

Prediction

This study used four models: Random Forest (RF)

(Rigatti, 2017), Decision Tree (DT) (Song, 2015),

Logistic Regression (LR) (LaValley, 2008) and

Artificial Neural Network (ANN). RF, DT and LR are

constructed using sklearn, whereas tensorflow is

utilized for ANN construction. Models will be mainly

evaluated by their accuracies on predictions and the

time efficiency will also be slightly taken into

consideration.

2.2.1 Random Forest

RF operates by constructing a multitude of decision

trees during training. Generally speaking, by

averaging multiple decision trees each built on a

diverse subset of the data (bootstrap samples), RF has

a stable prediction performance and usually low risk

of overfitting; the data that is left Out-of-Bag (OOB)

can be used to estimate the model's performance,

providing an unbiased estimate of the test error

without needing a separate validation set. Individual

trees are trained separately but parallel which makes

RF scalable for large datasets. Here, for the adoption

prediction classification tasks, each tree in the forest

votes for a class, and the class with the most votes is

chosen as the final prediction.

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

364

In this study, RF is optimized using

GridSearchCV with four parameters. n_estimators

represent the number of trees in the forest, ranging

from 100 to 1000, where more trees can improve

accuracy but meanwhile increase computation time.

The maximum depth (max_depth) of each tree is set

to None (unlimited) or restricted to 10, 20, or 30. The

minimum number of samples (min_samples_split)

required to split an internal node is set to 2, 5, or 10,

influencing the model's complexity.

min_samples_leaf sets the minimum number of

samples required to be at a leaf node, ranging from 10

to 100, which controls overfitting by ensuring each

leaf has enough data.

2.2.2 Decision Tree

DT is a tree-like structure where each internal node

represents a decision based on a feature, each branch

represents the outcome of the decision, and each leaf

node represents the final output or decision. As

mentioned above in RF, the data will be splitted into

subsets until the outputs are pure at the leaf node. The

splitting usually happens from the root into two or

more homogeneous data sets based on the most

significant feature and recursively on each derived

subset. After the tree has been fully grown, if needed,

trees can be pruned by removing branches that have

little importance to avoid overfitting. Moreover, DT

can rank features by importance and the missing

values can be either ignored or by considering

multiple possible outcomes.

As stated, DT is part of RF. Therefore, the

parameters are similar, and for DT in this study,

max_depth from 10 to 70 or None, min_samples_split

between 2 and 20, and min_samples_leaf from 1 to 8.

Besides, the split strategy (splitter) is either "best" for

optimal splits or "random" for random splits.

2.2.3 Logistic Regression

LR is commonly used in binary classification tasks,

which fits the pet adoption prediction. The model

assumes a linear relationship between the

independent variables (the input) and the log-odds of

the dependent variable (the outcome). The logistic

function, also called sigmoid, will transform any real-

valued input to a value between 0 and 1 with the

formula, where z is the linear combination of input

features and their corresponding coefficients.

Usually, as well as in this study, the threshold is set

to 0.5, where a result greater than 0.5 predicts the

positive class, otherwise negative class. Besides

binary classification, LR can be extended to

multiclass classification using techniques like One-

vs-Rest (OvR) or softmax regression.

In this study, the inverse of regularization strength

options (solver) include ‘newton-cg', 'lbfgs' and 'sag',

which impacts convergence speed and model

accuracy.

2.2.4 Artificial Neural Network

ANN is characterized by its human brain’s structure,

usually consisting of multiple interconnected layers

of nodes, or "neurons," that process and transmit

information. The model stimulates how the human

brain processes input information and makes

reactions and decisions. Each neuron receives input,

passes it into the assigned activation function which

allows ANN to model non-linear relationships and

then conveys the output to the next layer. Connections

between neurons have weights, which determine the

strength of it. This weight and the bias added to the

input of neurons will keep being updated when the

error is propagated backward through the network.

This ANN model will be constructed with 3 layers

of Rectified Linear Unit (ReLU) and 1 more layer of

sigmoid as output layer. The first and last layers of

ReLU are set to have the same number of neurons

(20) as the input shape. The number of neurons in the

middle layer of ReLU will be the best option between

40 to 80. Adaptive Moment Estimation (Adam) is

chosen as the optimizer for faster convergence and

acceleration in the relevant direction and dampen

oscillations, with the loss set as 'binary_crossentropy'

and metrics set as 'accuracy'. Epochs will vary in

range from 30 to 90.

3 RESULT AND DISCUSSION

The training and testing accuracies for RF, DT and

LR models are shown in Table 1, and the test result

for ANN is listed in Table 2 and Table 3.

Table 1: Accuracy and AUC in training and testing.

RF DT LR

Accuracy

in Training 0.9495327103 0.9426791277 0.9158878505

Accuracy

in Testing 0.9129353234 0.9179104478 0.8855721393

AUC in

Train 0.9668213326 0.9842749416 0.9343185955

AUC in

Test 0.9207912458 0.9102553311 0.9225028058

Pet Adoption Status Prediction Based on Multiple Machine Learning Models

365

Table 2: Accuracy with epoch from 30 to 90 and number of middle layer neuron from 40 to 80.

Number of

Middle Layer

Neuron

Epoch = 30 Epoch = 40 Epoch = 50 Epoch = 60 Epoch = 70 Epoch = 80 Epoch = 90

40 0.9606 0.9655 0.9873 0.9936 0.9927 0.9954 0.9998

50 0.9718 0.9781 0.9883 0.9948 0.9974 1.0000 1.0000

60 0.9432 0.9772 0.9924 0.9984 0.9986 1.0000 1.0000

70 0.9557 0.9811 0.9906 0.9984 0.9998 1.0000 1.0000

80 0.9597 0.9913 0.9971 0.9983 1.0000 1.0000 1.0000

Average

Accurac

0.9582 0.9786 0.9911 0.9967 0.9977 0.9991 1.0000

Table 3: Loss with epoch from 30 to 90 and number of middle layer neuron from 40 to 80.

Number of

Middle

Layer Neu

Epoch = 30 Epoch = 40 Epoch = 50 Epoch = 60 Epoch = 70 Epoch = 80 Epoch = 90

40 0.1099 0.1066 0.0536 0.0331 0.0261 0.0215 0.0062

50 0.0918 0.0725 0.0458 0.0253 0.0201 0.0043 0.0034

60 0.1523 0.0714 0.0364 0.0171 0.0168 0.0093 0.0023

70 0.1071 0.0659 0.0361 0.0141 0.0060 0.0041 0.0046

80 0.1117 0.0473 0.0218 0.0183 0.0066 0.0043 0.0038

Average

Accurac

0.1146 0.0727 0.0387 0.0216 0.0151 0.0087 0.0041

Each data in Table 2 and Table 3 was the average

result of 5 training sessions with the corresponding

epoch and middle layer neurons. Clearly, by

comparing the data in two tables, ANN achieves a

general higher accuracy above 0.95 than the other

three models which have accuracy around 0.90.

Focusing on ANN, it can be seen that with the same

amount of middle layer neurons, as the epoch

increases, the accuracy generally increases and the

loss decreases. However, after reaching about 60

epochs, the increasing trend for accuracy and

decreasing trend for loss becomes somewhat

unpredictable, which shows that even though the data

with higher epochs have a better accuracy and less

loss, it may risk potential overfitting. Similarly, with

the same number of epochs, the suitable range for the

number of neurons lies around 60, where there is a

trend of increasing when less than 60 neurons and

almost no improvement when greater than 60

neurons. Therefore, the suitable model chosen in this

study, achieving an accuracy above 0.99, is a four-

layer ANN consisting of three layers of ReLu and one

more output layer as Sigmoid with both the number

of the middle ReLU layer neuron and epoch around

60 and the neurons in the first and last ReLU layers

set to 20.

There might be a bunch of reasons that ANN

outperformed in this study. As an activation function,

ReLU introduces non-linearity into the model,

allowing it to capture complex relationships in the

data. In contrast, DT may struggle with complex non-

linear patterns and LR models cannot even process

non-linear patterns. Furthermore, ANNs can

implicitly learn interactions between features through

their layered structure. Even though RF does consider

interactions through the aggregation of multiple

decision trees, they might still miss certain subtle

interactions that a neural network can learn. By using

gradient descent and backpropagation, ANN can

efficiently optimize complex loss landscapes and find

patterns that might be hard for non-gradient-based

methods to identify, such as tree-based methods using

greedy algorithms to split nodes, which might not find

the optimal splits for complex data structures.

Due to the limit of resources, device and time,

some complex parameters for RF, DT and LR are not

included in the test, such as criterion, max_feature,

tol, max_iter, etc. By adding these parameters, the

models may be enhanced in accuracy but will

definitely take much longer time to process.

Meanwhile, since each testing result recorded in

Table 2 and Table 3 is just the average of 10 training

results, there can be some bias due to the random

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

366

seed. More trials for each parameter may reflect a

more reliable and representative overview of the

accuracy of the setting. In order to possibly enhance

the accuracy of the prediction, in addition to the pet

information already included in the data set, some

other pet features and specific basic information can

be added into the data set such as sterilization,

biological antibody, inherent disease, etc. A very

representative or decisive information is likely to

greatly promote the accuracy and efficiency of the

model.

4 CONCLUSIONS

In this study, ML is used to predict the adoption status

for pets based on some specific pet characteristics and

information. After the preprocessing which removes

some useless columns and separate the categorical

and numerical columns, multiple models, including

RF, DT, LR and ANN are then applied into ML. In

order to find the most suitable model with the highest

accuracy, different parameters are tested and the

result shows that the 4 layers (3 layers of ReLu and 1

layer of Sigmoid) ANN model with epoch and the

number of neurons in the second ReLu layer both

around 60 and the neurons in the first and last ReLU

layers set to 20. For any future study with similar data

sets, it is essential to choose and evaluate the model

data-dependently, since if the data set is not complex

enough to warrant a deep learning approach, simpler

models, such as linear regression, naive bayes, etc.

may outperform an ANN or any model mentioned in

this study. Adding more representative and decisive

pet information is also encouraged in future ML

processes.

REFERENCES

Amirhosseini, M. H., Yadav, V., Serpell, J. A., Pettigrew,

P., & Kain, P. 2024. An artificial intelligence approach

to predicting personality types in dogs. Scientific

Reports, 14(1), 2404.

Hajek, A., Peltzer, K., Veronese, N., König, H. H., & Gyasi,

R. M. 2024. Pet ownership and psychosocial outcomes

among the oldest old in Germany during the Covid‐19

pandemic. Findings based on the nationally

representative “Old Age in Germany (D80+)”.

International journal of geriatric psychiatry, 39(7),

e6127.

Hussein, S. M., Soliman, W. S., & Khalifa, A. A. 2021.

Benefits of pets’ ownership, a review based on health

perspectives. J. Intern. Med. Emerg. Res, 2, 1-9.

Kim, S. C., & Kim, S. 2024. Development of a Dog Health

Score Using an Artificial Intelligence Disease

Prediction Algorithm Based on Multifaceted

Data. Animals, 14(2), 256.

LaValley, M. P. 2008. Logistic regression.

Circulation, 117(18), 2395-2399.

Megna, M. 2024. Pet ownership statistics 2024. Forbes.

https://www.forbes.com/advisor/pet-insurance/pet-

ownership-statistics/

Rigatti, S. J. 2017. Random forest. Journal of Insurance

Medicine, 47(1), 31-39.

Singh, S., & Sunoj, R. B. 2023. Molecular machine learning

for chemical catalysis: Prospects and challenges.

Accounts of Chemical Research, 56(3), 402-412.

Song, Y. Y., & Ying, L. U. 2015. Decision tree methods:

applications for classification and prediction. Shanghai

archives of psychiatry, 27(2), 130.

Zeng, Q., Sun, W., Xu, J., Wan, W., & Pan, L. 2024.

Machine Learning-Based Medical Imaging Detection

and Diagnostic Assistance. International Journal of

Computer Science and Information Technology, 2(1),

36-44.

Pet Adoption Status Prediction Based on Multiple Machine Learning Models

367