Application of Deep Learning in Tourist

Daniel Samwel Makala

and Li Zongmin Li

Department of Computer Science, China University of Petroleum (East), Qingdao, China

Keywords: Deep Learning, Machine Learning, Prediction Analysis, Forecasting, Neural Networks.

Abstract: One of the Giant sectors in Tanzania is Tourism. About 40% of the foreign exchange in Tanzania comes from

this sector. It is the number one job provider to Tanzanians, about 10% of the working class is in the tourism

sector. Since independence, the sector has been growing well until 2019 during the pandemic issue of COVID-

19. However, since 2020 Tanzania has regained and restored the tourist income to normal and expected more

tourists. Government and Authority are in the age of determining the number of tourists to come and the

income associated with the tourist for better planning. Forecasting tourist inflows requires an accurate model

because of the highly changing tourist data due to external factors such as political influence, security issues,

or transportation issues. This study analyses and proposes the CNN to be used for the prediction of tourist

arrival. CNN can handle and process multiple sequences and thus can handle data and multivariate time series.

Using data from 1961 to 2022, of Tanzania's arrival, the proposed model was able to predict with more

accuracy compared to ARIMA, LSTM, and CNN-LSTM by having 0.1 RMSE. However, the study is limited

due to the unavailability of daily tourist income.

1 INTRODUCTION

1.1 Overview of Tanzania Tourism

Tanzania is one of the most beautiful countries in the

attraction of visiting and tourism. She is most famous

for her natural gorgeous and rich cultural heritage.

She is home to the beautiful highest mountain in

Africa, Mount Kilimanjaro, and the world's number

one natural national park, Serengeti National Park.

Not only those two Tanzania has abundant attractions

including Zanzibar Island, Ngorongoro creator

among many others. Because of this beautiful nature,

the demand for quality tourism services has

increased. The Tanzanian tourism industry

contributes a lot to the Tanzanian economy. It

contributes about 17% to the national GDP of the

country and 35% of all foreign exchange revenues.

According to government statistics, the tourism sector

provides direct and indirect employment to the

people. It was recorded in 2018 that this sector gave

out 2.6 million jobs to the people. This means the

tourism sector in Tanzania is one of the major job

providers. (Kyara et al., 2022).

https://orcid.org/0000-0003-3827-2346

https://orcid.org/0000-0003-4785-791X

It is all known that the pandemic of covid 19

disturbs each economic sector everywhere. This

tragedy spared no sector as the tourism sector also

faced this challenge. This is because of many

countries’ restrictions on traveling, and it was

predicted that international tourism would fall by

80% (Henseler et al., 2022). Before COVID-19, this

sector was the leading foreign exchange earner in

Tanzania. From late 2021, the sector is regaining

momentum, and the government has to lift the sector

to its peak. One of the ways used by the government

is the introduction of the royal tour (Tanzania

Tourism Sector - February 2023 Update, n.d.). The

royal tour involves the President of Tanzania as the

tour guide. The president's ultimate guide for a week,

unveiling Tanzania’s history, environment, music

foods, and culture, as well as telling the stories of

Tanzania’s hidden jewels. The CEO of the Hotel

Association in Tanzania said We hold better positive

projections in 2023. If the income of tourists increases

at this rate, then we can call it a full recovery at least

by the end of this year (Tanzania Tourism Sector -

February 2023 Update, n.d.).

With the increase in popularity, it comes with the

responsibility as it is very crucial to manage the

248

Makala, D. S. and Zongmin, L.

Application of Deep Learning in Tourist.

DOI: 10.5220/0013342500004646

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

In Proceedings of the 1st International Conference on Cognitive & Cloud Computing (IC3Com 2024), pages 248-256

ISBN: 978-989-758-739-9

number of visitors entering the countries to maintain

better environments for the attractions as well as to

ensure the environment is not damaged or

overwhelmed by tourists. Thus, it is essential to

determine the volume of people (tourists) visiting the

country over time. Forecasting the income of tourists

will enable stakeholders in the tourism sector to

provide a memorable experience for their tourists and

hence a country gets better recommendations to

attract more tourists. A better prediction model is

needed.

This study uses the Tourism sector in Tanzania as

a case study as Tanzania is one of the countries doing

better in the tourism sector. This case study will

represent the tourism sector around the world.

1.2 Background of the Study

As discussed in the earlier segment, it is important to

forecast the number of tourist arrivals in any country

for better tourism service. This article focuses on

identifying the best deep learning model to be used

for forecasting the volume of tourist arrivals with the

case study of Tanzania. Nevertheless, the study also

engaged in comparing the results of the proposed

model with the real data to see how the model has

performed. Lastly, the study uses the model to focus

on the volume of tourist arrivals in Tanzania.

In this study, the deep learning mole is considered

since deep learning has been showing a greater

impact in the prediction analysis of times series. The

CNN model is proposed to be used in the prediction

of tourism arrival in Tanzania. The CNN model has

been performing wonders in the health sector,

especially in image recognition and speech

recognition. Because of the better performance of the

model in images, the study proposed this model to use

the features in the models to capture the trend

movement of the tourism arrival, as it is known the

nature of the number of tourists is very volatile and

non-linear. Therefore, the proposed model will be a

suitable model for capturing all the trend movement

of time series data such as the number of income

tourists. The proposed model will be compared with

another popular traditional model of ARIMA and

Deep learning models of LSTM as well as CNN-

LSTM models.

2 RELATED WORKS

2.1 Introduction

Related work is simply a summary of existing

research and study published materials related to a

certain topic. It gives out what some researchers have

done about that specific topic. It also provides more

sources of information for easy understanding. This

part of the related work surveys the available

literature to identify current trends and gaps in

knowledge, while also providing context for future

studies. Different studies have indicated that the way

of forecasting is categorized in two ways which are

linear way and non-linear ways. (Y. Li & Cao, 2018).

However, according to, there are three ways of

forecasting, these are time series way, Artificial

intelligence, and economics. But time series and

econometric ways are sometimes known as

traditional ways.

For a long time, prediction and forecasting have

been conducted. In previous years linear methods

were highly used although they were not able to

capture all the non-linearity data for prediction.

Sometimes complex models may not perform as

perfectly as simpler techniques. (Choy, 1984).

“Because of the influence of external forces such as

public health, economic crises, and seasonal

variations, the tourism arrival time series data have

become more complex, and nonstationary as a result,

it is tremendously challenging to obtain satisfactory

results when dealing with the prediction of such

dataset” (Goh et al., 2008).

Different researchers and academicians have

discussed the tourism sector in Tanzania. Much more

the discussion is based on how tourism employs

Tanzanians. It is found that about 5% of GDP is from

the tourism sector and this is incline in future time.

The tourism sector provides 10% of the working

population in Tanzania. The sector is a leading sector

in contribution to foreign exchanges. (Wamboye et

al., 2020)

2.2 Traditional Forecasting Models

ARIMA method has been extensively used in

prediction and forecasting. This model can take into

account seasonality variation and that is often used in

the tourism sector, since one of the factors that affect

the income of tourists is the season of the year. Take

the example of (Bumthang, 2018), who researched

forecasting international tourists visiting Bumthang

in Bhutan. The data used are from January 2017 to

June 2017. The out of their research shows that

Season Arima (0,0,0) *(1,1,0) performs very well

with an accuracy of 91%. More to that (Petrevska,

2017) also conducted a study using 58 observations

of the data arrival to Macedonia from 1956 to 2013.

The result shows that the ARIMA (1,1,1,) is suitable

for forecasting tourists in Macedonia and predicted

Application of Deep Learning in Tourist

249

that by 2018 the arrival of tourists will increase by

13.9%.

Besides the ARIMA model, the economic method

known as the Vector Autoregressive model (VAR) is

another prediction technique that is doing very well.

The VAR model relies on one element depending on

the other element. This implies that, to perform with

VAR two or more variables are needed and must

relate to each other. (Vector Autoregression (VAR) -

Comprehensive Guide with Examples in Python -

Machine Learning Plus, n.d.). This model has been

used in tourism demand by different researchers,

(Witt & Witt, 1995; Wong et al., 2006).

(Choy, 1984)“Conducted and investigated a

comprehensive method for tourism demand analysis.

This accurate and systematic approach is based on the

Bayesian global vector autoregressive model

(BGVAR). They deal with the income of

international tourists in nine Countries of Southern

Asia.” The outcome of their research shows BGVAR

to perform better than the other three different VAR

models. They conclude by outlining the superiority of

their model and how powerful it is in forecasting

tourism in Asia and time series in general.

It is true many studies show wide usage of

traditional models, this indicates how powerful

ARIMA is among the traditional. (Huang & Min,

2002) And in studies on the demand for tourists in

Taiwan. (M. Li et al., 2023) Among others, all show

how the ARIMA model is successful, but the model

performs well when dealing with the linearity of the

data only. Similarly, the VAR model to predict the

flow of tourists in Macau, Germany, and British

forecasted demand in Greece, and forecasted tourism

ex-port and export price of the EU-15, respectively.

All the above researchers found VAR to be the best

model. However due to the complexity and large

volume of data available now days Deep learning has

become a solution for the problem of non-linear data.

Different traditional models have been used in

tourism sectors such as elaborated by (Astuti et al.,

2018; Yue et al., 2017).

The traditional models have some weaknesses

that lead to the development of artificial intelligence

models especially deep learning modes. These

include reliance on historical data and linear models,

falling short of capturing the complexities of the

modern world, failure to capture the non-linear

relationship of the variables as well as failure to

handle large volumes of data.

2.3 Deep Learning Models

Artificial Intelligence can be defined by dozens of

definitions. Some research explains (AI) Artificial

intelligence as the power of computers and machines

to impersonate the problem-solving and decision-

making capabilities of the human mind. (What Is

Artificial Intelligence (AI) ? | IBM, n.d.). AI is where

the world is and deep Learning is the feature. The

tourism sector also heavily applied DL techniques in

different ways. (Kontogianni et al., 2022). For a better

plan of tourist policy, accurate tourism demand

forecasting plays an important factor. However,

making predictions about tourism is very complex

and not in linear form. And here is where Deep

Learning Models come in.

(Essien & Chukwukelu, 2022) In their study, the

aim is to “provide an efficient evaluation of the

existing literature on the applications of deep learning

(DL) in hospitality, tourism, and travel as well as an

agenda for future research”. Their study is based on a

review case analysis. Their study concentrates on the

five years of data from 2017 to 2021, basically

journals from Springer, Science Direct, Emerald

Insight, and Wiley Library. They found out that

“Deep learning is mainly used to develop novel

models that create business value by forecasting (or

projecting) some parameter(s) and promoting better

offerings to tourists”.

Other researchers proposed a hybrid model of

SARIMA-CNN-LSTM intending to forecast the

tourist demand using daily data. Here SARIMA

captures the linearity of the data structure, CNN

captures nonlinear data features and LSTM captures

the long-term dependencies in the data. Combining

all these three models ensures they capture every

feature of the data. The results show the proposed

model of the SARIMA-CNN-LSTM has greater

forecasting accuracy compared to the individual

model. We all know forecasting the flow of tourists is

at an important level for the government and

authorities. (Y. Li & Cao, 2018) uses LSTM to

predict the flow of tourists where the proposed model

performs better than ARIM and Back Propagation

Neural Network (BPNN). More to that, (Wu et al.,

2020) whose study aimed at forecasting the daily

arrival of tourists in Macau China, proposed to use the

hybrid mode of the SARIMA-LSTM approach. The

results indicate the hybrid model performs well.

Again (Chang & Tsai, 2017) compare the SVM, NN

and Deep learning applied neural network, by using

the MAPE as a measurement of performance they

find deep learning applied Neural Network has a

MAPE of 2.05% while the other models have a

MAPE of 10%.

IC3Com 2024 - International Conference on Cognitive & Cloud Computing

250

However deep learning models have resolved the

issues raised in traditional models. As we all know

nothing is perfect, the deep learning models have

some shortcomings too which include that, they

perfectly work well when a large dataset is involved.

Also, the algorithm of DL models is much more

complex. To try to resolve these issues the study uses

the CNN model to ensure the mentioned issues are

resolved with the main focus being to get the greatest

model in the prediction of tourist arrival.

3 METHODOLOGIES

3.1 Introduction

The emphasis of this study is to find the best model

for the prediction of tourist arrival using the case

study of Tanzania tourist arrival. To achieve this

target, the deep learning model of CNN is proposed

since the model has been doing an amazing job in

various tasks such as image recognition, playing

complex games, car-self driving, and speech

recognition. The model is then compared with

ARIMA, LSTM, and the hybrid model of CNN-

LSTM. There here our proposed model together with

comparison models is explained in the methodology

section.

3.2 Data Collection and Processing

Making predictions about tourism is very complex

and not in linear form. And here is where Deep

Learning Models come in. One of the key aspects of

any research study and to have data. The data may be

collected from different sources such as interviews,

questionnaires, captured from different platforms, or

even face-to-face conversations depending on the

nature of your study. Due to the high increase in the

usage of technology, and the availability of large

volume data, many companies and government

institutions put most of the data over the internet (soft

data). This paper uses data obtained from the

Tanzania National Bureau of Statistics, World Bank,

and Statista. (Tanzania Tourism Sector - February

2023 Update, n.d.), (Tanzania: Number of Tourist

Arrivals 2015-2022 | Statista, n.d.). The annual data

of tourist arrivals from 1961 to 2022 is obtained from

the mentioned websites.

After acquiring the data, the cleaning of the data

phase follows. Here understanding what kind of data

is needed in our model is required. Cleaning and data

processing involve putting the data in the required

form, ensuring no missing or filling up of the missing

data, and so on. Thereafter data are categorized into

training data and testing data. Training data in this

study case are the tourist arrival in Tanzania from

1961 – 2012 and the validation data is for ten years

from 2013 -2022.

3.3 Convolutional Neural Network

Convolution Neural Network is a deep neural

network that has been very famous in image

classification, and image search among others in the

same line. Because of that, it has been very useful in

health centers, especially in cancer detection and MRI

scanners. Nerve the less in recent years it has been

involved in time series prediction.

CNN structure has two sections, the first being the

convolutional and pooling parts and the second party

is fully connected layers. The first section,

convolutional and pooling consists of an input layer

and convolutional kernels. ‘The pooling part does a

reduction of the dimension and puts them into a single

neuron. The convolutional layers perform

convolution operations on the time series of the

preceding layer with convolution filters” (Luo et al.,

2019). The fully connected layer is simply one neuron

in one layer that has a connection to another neuron

in another layer. It is simple MLP. The fully

connected layer is connected to the output layer.

Figure 1 explains the structure of CNN. In CNN the

number of training weights is small causing a more

efficient model and expecting more accurate and

reliable results.

Figure 1Simple Architecture of CNN Model (Gu et al.,

2019)

3.4 Long Short Time Memory

It is an updated recurrent neural network that has the

capability of learning order dependence in a time

series sequence. This means LSTM is a modified

RNN that has been explicitly designed to avoid the

long-term dependency problem, which is the main

issue in RNN models. The LSTM was first discovered

by Hochreiter and Schmidhuber to address the issue

of long-term relief (Hochreiter & Schmidhuber,

1997)

Application of Deep Learning in Tourist

251

LSTM uses artificial neural networks. The

networks have connections to each other as data can

be transferred backward and forward. These neural

networks are also known as RNN (Recurrent Neural

Network). The structure of the LSTM is comprised of

four main sections in a single unit of LSTM

architecture. These four components are a cell, an

input gate, an output gate, and a forget gate. Three

gates control the flow of information into and out of

the cell, and the cell keeps track of values over an

arbitrary period as described in Figure 2.

Figure 2 Simple Block of LSTM (Makala & Li, 2019)

Forget gate: As the name expresses itself, the

forget gate is in control of determining which data

from the previous state is eliminated or kept. This

allows the LSTM to have the ability to keep the long-

term dependencies. This means forget gate assists in

choosing which past information is still important for

the prediction and which ones are not important. The

data from the previous state meets the current input

and are both processed through the sigmoid function

where the output is based on 1 or 0 values. For that

case when output is zero (0), that means the LSTM

forgets that information while the 1 value is kept.

Mathematically the work at Forget gate can be

expressed as in an equation 1 given that ht-1 is the

information from the previous cell and xt is the

current input.

𝑓



= 𝜎(𝑤





ℎ



, 𝑥





+ 𝑏



)

(1)

Input gate: One of the gates in the LSTM block

is the input gate. It does deal with determining which

information is to be added to the cell at a particular

time. The input gate has a layer that uses a function

known as sigmoid, whereby the sigmoid function

generates values between 0 and 1 and these values are

the ones that act as gate and control the information

passing through. Therefore, the main function at the

input gate includes controlling the information

entering the cell state, by filtering the important

information to pass through and discarding the

irrelevant information. In addition to that, there is

tanh function which creates a vector of new input that

could be added to the cell. More to that, the input gate

can handle long-term dependence. Now let the

information from the previous state be denoted by

ℎ𝑡−1 and current input 𝑥𝑡. The general equation at the

input gate will be as:

𝑖



= 𝜎(𝑤





ℎ



, 𝑥





+ 𝑏



)

Memory Cell: it comprises the CEC, having a

recurrent edge with unit weight. The current cell state

is computed by forgetting irrelevant information (if

any) from the previous time step and accepting

relevant information (if any) from the current input.

Equation here are

𝑐



= 𝑡𝑎𝑛ℎ(𝑤





ℎ



, 𝑥





+ 𝑏



)

𝑐



𝑓



∗𝑐



+ 𝑖



∗𝑐



)

Output gate. This is the last part of the block. As

its name is the output gate is where the decision to

give output or to re-enter again of information is

conducted. It controls what information would flow

out of the LSTM unit as the output of LSTM. At this

gate, the main function is regulating the information

that has been processed in hidden layers and becomes

the out of the LSTM unit. Also, this gate enables the

LSTM unit to learn the complexity of the pattern and

hence improve the performance. In a similar manner

as in forget get, the input from the past cell is

combined with the current input and they pass

through the sigmoid function to generate the value

between o and 1 and then are multiplied by the tanh

function. These two functions can be expressed

mathematically as shown in equation 5 and 6

respectively.

𝑜



= 𝜎(𝑤





ℎ



, 𝑥





+ 𝑏



)

And the

ℎ



= 𝑜



∗tanh

(

𝑐



)

Whereby:

𝑖𝑡 indicates the input gate, 𝑓𝑡shows forget gate, 𝑜𝑡

is the output gate, 𝜎 is the sigmoid function, 𝑤𝑥

indicates weight at respective gates, ℎ𝑡−1 shows the

input from the previous LSTM block, 𝑥𝑡 shows

current input and 𝑏𝑜 indicate biases at respective

gates.

Generally, LSTM does have to do very well in

many aspects because of the function it has in each

gate. These gates allow better control of information.

More of that it consists of memory cell which enables

the storage of information. LSTM should be able to

IC3Com 2024 - International Conference on Cognitive & Cloud Computing

252

handle the long-term dependency. It has done well in

natural language processing, image classification,

and time-series forecasting

3.5 ARIMA

Autoregressive Integrated Moving Average is also

known as ARIMA by most people. It is one of the old

and common models in forecasting and prediction

analysis. This model was first introduced by G. Box

and Gwilym in the 1970s (Liu et al., 2011; Sato,

2013). ARIMA's name comes from the combination

of the three models. AR is simply an autoregression

method, I stands for integrated and MA is the Moving

average method. This means It combines three

important components: autoregression, differencing,

and MA to capture the trends and patterns in the data.

The mode generally advancements the version of the

autoregressive moving average (ARMA). It is most

denoted as ARIMA (p, d, q), whereby P is lag order,

which is simply the number of lag observations

included in the model, and d is the degree of

difference which is the number of times that the raw

observations are differenced, and q is the order of

moving average some time known as the size of the

moving average window.

These parameters (p,d,q) can change the model

when is obtained. The parameter value of zero (0),

indicates that that parameter has no use in the

equation and the model. This way, the ARIMA model

can be developed to perform the function of an

ARMA model, or even simple AR, I, or MA models.

Consider an example the value of is (p,0,0) then

ARIMA becomes equal to AR(p) since d and q have

no meaning and values p, d, and q can never be

negative.

In ARIMA the predicting equation is constructed

as follows. First, let y denote the d

difference of Y,

which means:

If d = 0; then 𝑦𝑡 = Y𝑡

(7)

If d = 1: 𝑦𝑡 = Y𝑡 - Yt-1

(8)

If d = 2: y𝑡 = (𝑌𝑡 - Yt-1) - (𝑌𝑡-1 - Yt-2)

(9)

=𝑌𝑡 - 2Yt-1 + Yt-2

(10)

According to (ARIMA Model - Complete Guide to

Time Series Forecasting in Python | ML+, n.d.) the

mathematical presented can be put in words and be

Predicted

Yt = Constant + Linear combination Lags of Y +

Linear Combination of Lagged forecast errors (up to

q lags).

Therefore in ARIMA, it is very important to find

the value of parameters p,d,q. The value of p can be

found by looking at the value of PACF. This is a

Partial Autocorrelation plot. PACF is a correlation

relationship between the series and its lag. The Value

of d can be calculated by looking at several

differencing conducted. The aim of doing

differencing is to make time series data at the

stationary level. If the data given are at stationery that

means d =0. The stationarity of data can be calculated

by the Augmented Dickey-Fuller test.

3.6 CNN-LSTM

As the name referred to CNN-LSTM, it is the hybrid

of the two models of neural convolution network and

the LSTM (long short time memory). CNN-LSTM

was created for the focus on the time series

forecasting issues and to generate text descriptions

from sequences of images and videos, but can also

accommodate the times series prediction.

This hybrid method is a combination of the CNN

model and LSTM. In this model, first, the input data

passes through the Convolution layers in CNN to

produce the vectors, which are then passed through

the LSTM layers to produce output. The aim of

choosing this methodology is to capture all necessary

information from CNN and LSTM in case one model

is missed. Figure 3 explains the pictorial presentation

of the flow of data of the model hybrid CNN-LSTM.

Thus In this model, the CNN part will also be

involved in interpreting the sequence of input data

while LSTM will be put together for the forecasting

process.

3.7 Performance Evaluation

It is very important to determine how the models have

performed and compare them to each other. These

performance measurements tell us how best the

model is doing and based on these results, decision-

Figure 3 Showing simple flow of data in CNN-LSTM

Application of Deep Learning in Tourist

253

makers can decide on the different factors to look at

when computing the outcome of the deep learning

model. For this research work, accuracy was the main

factor. Here we consider three main measurement toll

that are:

RMSE: This is Root Means Square Error, that

simply the square root of the square of different actual

values and predicted values it is one of the popular

metrics in prediction analysis. RMSE gives the results

in the same SI unit as the data used in prediction. For

example, in the forecasting of the price of oil, the

RMSE result will be in the same price unit.

Mathematically:

𝑅𝑀𝑆𝐸 =



∑

(𝑝𝑟𝑒𝑑𝑖𝑐𝑡−

𝐴

𝑐𝑡𝑢𝑎𝑙)







𝑁

(11)

MAE: This Means Absolute Error is the metric

used to measure how much big an error occurred in

the forecasting model. Similar to RMSE the results of

the MAE are also in the same unit with the value used

in the forecasting model.

𝑀𝐴𝐸 =

𝑛



𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛−

𝐴

𝑐𝑡𝑢𝑎𝑙







(12)

R-squares (R2) is sometimes referred to as the

coefficient of determination. It is mostly used in a

regression model which determines the

proportionality of variance in the dependent variable

that explains the independent variable. Mostly

presented as a percentage value. [46] The higher the

value the better performance.

𝑅−𝑆𝑞𝑢𝑎𝑟𝑒𝑑= = 1 −

𝑆𝑆



𝑆𝑆







(13)

𝑆𝑆



= sum of squares due to regression, and

𝑆𝑆



The total sum of sq.

4 RESULTS AND DISCUSSION

Before running into the training, the data has to be

divided into the training set and evaluation set. Then

proposed model is trained with the training in order

to learn the movement and trend of the given data.

with tunning at different parameters of the model

such as different epochs, and the number of layers,

the model with epochs of 500 and the 1 convolutional

layer of 256 filters provides better results. The

outcome of this research article shows that the CNN

performed better compared to all other models,

however, the hybrid model of CNN-LSTM is also

better. Looking at the evaluation tools, the less the

value of the RMSE indicates the better performance

of the model. CNN model has approximately 0.1

RMSE which can also explain how perfect the model

is with R-square, whereas CNN has 99% as shown in

Fig 4. Figure 4 shows how results of each model vs

the true value. Looking CNN, and True values are

very closely followed by the hybrid model of CNN-

LSTM

This study expected all of the deep learning to

perform better, unfortunately, LSTM did poorly. The

main reason of poorly performance of LSTM is

caused by having fewer datasets which are annual

data. Since the CNN did well followed by the CNN-

LSTM, with having r-square of 0.99 and 0.98

respectively, the see pictorial for how they relate with

the actual value as shown in Figure 4. If you look over

you will observe in both cases the predictions are

almost similar to the actual value, however, by

looking at the evaluation matrix tools R-Square,

RMSE, and Means Absolute Error the CNN did better

as indicated in Table 1.

Table 1: Shows the performance of the models.

RMSE MAE R-Square

149113.89 133688.35 0.83

ARIMA 408036.81 257622.42 0.26

LSTM 464294.04 389014.22 0.22

CNN 0.09 0.04 0.99

CNN-LSTM 1389 1052.74 0.98

IC3Com 2024 - International Conference on Cognitive & Cloud Computing

254

Figure 4 Shows how each model has performed with compared with the true value

5 CONCLUSIONS

CNN has performed well compared to other models.

This shows that the deep learning model has more

chance of doing better compared to other models

regardless of the limited dataset used in this study.

Deep learning is the current and future technology

since it has already done wonders in self car driven,

speech recognition, as well as image classification.

The deep learning techniques have the potential to

revolutionize the tourism industries across the world

by providing accurate information and hence the DL

will provide efficient solutions for the market as well

as the service provided in the tourism industries. By

having the proper framework and designing of the

model that can accurately predict customer behaviors

and be able to capture the trend of income tourists, the

Tanzania tourism sector may have the greater benefit

from the increase of revenue as well as improving the

customer service to the tourist. More to that Deep

learning technology can be used in improving safety

levels, within public places through facial recognition

logarithms, whereby it helps officials and authorities

to keep the cities safe. Also helps against terrorism

and hence attracts more tourists.

Nevertheless, the Tanzania government has to

overcome and address all the challenges associated

with technology adaptation. These challenges include

a lack of skilled personnel, technological

infrastructure, as well as cyber security. This study

believes the Tanzania Government will soon be able

to utilize deep learning technology solutions to

improve the tourism sector.

In conclusion, this study paper has shown a

meaningful input and influence on the tourism sector

by forecasting the arrival of tourism in Tanzania. The

dataset of tourist arrivals in Tanzania represents all

the countries as Tanzania has been used as a place of

where the research has been conducted. This study

intended to fill the gaps that arise with other models,

especially the traditional ones. However, there are

some limitations to this research such as the

availability of data, thus giving out opportunities for

improvements. We further welcome more research to

dig more into the application of the Convolutional

Neural Network model in the prediction of times

series including tourist arrivals.

DECLARATION

The authors of this work declare that they have no

known kind of interest that influences this work paper

ACKNOWLEDGMENTS

The authors of this work would like to acknowledge

the support received during this work from the Deep

Learning group at China University of Petroleum,

Qingdao.

REFERENCES

ARIMA Model - Complete Guide to Time Series

Forecasting in Python | ML+. (n.d.). Retrieved August

15, 2023, from

https://www.machinelearningplus.com/time-

series/arima-model-time-series-forecasting-python/

Astuti, S. W., -, J., & Fahrudin, R. (2018). Forecasting

Tourist Visits Using Seasonal Autoregressive

Integrated Moving Average Method. IOP Conference

Series: Materials Science and Engineering, 407(1),

Application of Deep Learning in Tourist

255

012148. https://doi.org/10.1088/1757-

899X/407/1/012148

Bumthang, in. (2018). ARIMA model to forecast

international tourist visit. J. Phys.: Conf. Ser, 1039,

12023. https://doi.org/10.1088/1742-

6596/1039/1/012023

Chang, Y. W., & Tsai, C. Y. (2017). Apply deep learning

neural network to forecast number of tourists.

Proceedings - 31st IEEE International Conference on

Advanced Information Networking and Applications

Workshops, WAINA 2017, 259–264.

https://doi.org/10.1109/WAINA.2017.125

Choy, D. J. L. (1984). Forecasting tourism revisited.

Tourism Management, 5(3), 171–176.

https://doi.org/10.1016/0261-5177(84)90036-0

Essien, A., & Chukwukelu, G. (2022). Deep learning in

hospitality and tourism: a research framework agenda

for future research. International Journal of

Contemporary Hospitality Management, 34(12), 4480–

4515. https://doi.org/10.1108/IJCHM-09-2021-

1176/FULL/XML

Goh, C., Law, R., & Mok, H. M. K. (2008). Analyzing and

Forecasting Tourism Demand: A Rough Sets

Approach. Journal of Travel Research, 46(3), 327–338.

https://doi.org/10.1177/0047287506304047

Gu, H., Wang, Y., Hong, S., & Gui, G. (2019). Blind

channel identification aided generalized automatic

modulation recognition based on deep learning. IEEE

Access, 7, 110722–110729.

https://doi.org/10.1109/ACCESS.2019.2934354

Henseler, M., Maisonnave, H., & Maskaeva, A. (2022).

Economic impacts of COVID-19 on the tourism sector

in Tanzania. Annals of Tourism Research Empirical

Insights, 3(1), 100042.

https://doi.org/10.1016/J.ANNALE.2022.100042

Huang, J. H., & Min, J. C. H. (2002). Earthquake

devastation and recovery in tourism: the Taiwan case.

Tourism Management, 23(2), 145–154.

https://doi.org/10.1016/S0261-5177(01)00051-6

Kontogianni, A., Alepis, E., & Patsakis, C. (2022).

Promoting smart tourism personalised services via a

combination of deep learning techniques. Expert

Systems with Applications, 187, 115964.

https://doi.org/10.1016/J.ESWA.2021.115964

Kyara, V. C., Rahman, M. M., & Khanam, R. (2022).

Investigating the environmental externalities of tourism

development: evidence from Tanzania. Heliyon, 8(6),

e09617.

https://doi.org/10.1016/J.HELIYON.2022.E09617

Li, M., Zhang, C., Sun, S., & Wang, S. (2023). A novel deep

learning approach for tourism volume forecasting with

tourist search data. International Journal of Tourism

Research, 25(2), 183–197.

https://doi.org/10.1002/JTR.2558

Li, Y., & Cao, H. (2018). Prediction for Tourism Flow

based on LSTM Neural Network. Procedia Computer

Science, 129, 277–283.

https://doi.org/10.1016/J.PROCS.2018.03.076

Liu, Q., Liu, X., Jiang, B., & Yang, W. (2011). Forecasting

incidence of hemorrhagic fever with renal syndrome in

China using ARIMA model. BMC Infectious Diseases,

11, 218. https://doi.org/10.1186/1471-2334-11-218

Luo, Z., Cai, X., Tanaka, K., Takiguchi, T., Kinkyo, T., &

Hamori, S. (2019). Can We Forecast Daily Oil Futures

Prices? Experimental Evidence from Convolutional

Neural Networks. Journal of Risk and Financial

Management, 12(1), 9.

https://doi.org/10.3390/jrfm12010009

Makala, D., & Li, Z. (2019). ECONOMIC

FORECASTING WITH DEEP LEARNING: CRUDE

OIL. MATTER: International Journal of Science and

Technology, 5(2), 213–228.

https://doi.org/10.20319/mijst.2019.52.213228

Petrevska, B. (2017). Predicting tourism demand by

A.R.I.M.A. models. Economic Research-Ekonomska

Istraživanja, 30(1), 939–950.

https://doi.org/10.1080/1331677X.2017.1314822

Sato, R. C. (2013). Disease management with ARIMA

model in time series. Einstein, 11(1), 128.

https://doi.org/10.1590/S1679-45082013000100024

Tanzania: number of tourist arrivals 2015-2022 | Statista.

(n.d.). Retrieved July 29, 2023, from

https://www.statista.com/statistics/1248890/tourist-

arrivals-in-tanzania/

Tanzania Tourism Sector - February 2023 Update. (n.d.).

Retrieved July 26, 2023, from

https://www.tanzaniainvest.com/tourism

Vector Autoregression (VAR) - Comprehensive Guide with

Examples in Python - Machine Learning Plus. (n.d.).

Retrieved July 27, 2023, from

https://www.machinelearningplus.com/time-

series/vector-autoregression-examples-python/

Wamboye, E. F., Nyaronga, P. J., & Sergi, B. S. (2020).

What are the determinants of international tourism in

Tanzania? World Development Perspectives, 17,

100175. https://doi.org/10.1016/J.WDP.2020.100175

What is Artificial Intelligence (AI) ? | IBM. (n.d.). Retrieved

July 27, 2023, from

https://www.ibm.com/topics/artificial-intelligence

Witt, S. F., & Witt, C. A. (1995). Forecasting tourism

demand: A review of empirical research. International

Journal of Forecasting, 11(3), 447–475.

https://doi.org/10.1016/0169-2070(95)00591-7

Wong, K. K. F., Song, H., & Chon, K. S. (2006). Bayesian

models for tourism demand forecasting. Tourism

Management, 27(5), 773–780.

https://doi.org/10.1016/J.TOURMAN.2005.05.017

Wu, D. C. W., Ji, L., He, K., & Tso, K. F. G. (2020).

Forecasting Tourist Daily Arrivals With A Hybrid

Sarima–Lstm Approach.

Https://Doi.Org/10.1177/1096348020934046, 45(1),

52–67. https://doi.org/10.1177/1096348020934046

Yue, Y., Wei, M., & Yuan, S. (2017). Forecast related to

linear regression of China’s tourism market. Journal of

Interdisciplinary Mathematics, 20(6–7), 1367–1371.

https://doi.org/10.1080/09720502.2017.1386472

IC3Com 2024 - International Conference on Cognitive & Cloud Computing

256