Using Deep Learning for Trajectory Classiﬁcation

Nicksson C. A. de Freitas, Ticiana L. Coelho da Silva, Jos

e Ant

onio Fernandes de Mac

edo,

Leopoldo Melo Junior and Matheus Gomes Cordeiro

Insight Data Science Lab, Fortaleza, Brazil

Keywords:

Trajectory, Classiﬁcation, Machine Learning, Deep Learning.

Abstract:

The ubiquity of GPS-enabled smartphones and automotive navigation systems connected to the Internet allows

us to monitor, collect, and analyze large trajectory data streams in real-time. Trajectory classiﬁcation is an

efﬁcient way to analyze trajectory, consisting of building a prediction model to classify a new trajectory (or

sub-trajectory) in a single-class or multi-class. The classiﬁcation trajectory problem is challenging because of

the massive volume of trajectory data, the complexity associated with the data representation, the sparse nature

of the spatio-temporal points, the multidimensionality, and the number of classes can be much larger than the

number of motion patterns. Machine learning methods can handle trajectories, but they demand a feature

extraction process, and they suffer from the curse of dimensionality. On the other hand, more recent Deep

Learning models emerged to link trajectories to their generating users. Although they minimize the sparsity

problem by representing the input data as an embedding vector, these models limit themselves to deal with

multidimensional data. In this paper, we propose DeepeST (Deep Learning for Sub-Trajectory classiﬁcation)

to identify the category from a large number of sub-trajectories extracted from GPS services and check-ins

data. DeepeST employs a Recurrent Neural Network (RNN), both LSTM and Bi-directional LSTM (BLSTM),

which operates on the low-dimensional to learn the underlying category. We tackled the classiﬁcation problem

and conducted experiments on three real datasets with trajectories from GPS services and check-ins. We show

that DeepeST outperforms machine learning approaches and deep learning approaches from state-of-the-art.

1 INTRODUCTION

The recent advances in the sensors and communi-

cation technologies and the popularity of Location-

Based Social Networks (LBSNs) such as Foursquare,

Twitter, and Facebook contribute to the explosive

growth of trajectory data. We strongly believe that

these data provide a unique opportunity for under-

standing the patterns and behaviors of several mov-

ing objects, such as people, animals, transportation

modes, hurricanes, among others.

Trajectory data can be recorded in different for-

mats according to device types. For instance, GPS

tracking devices usually generate raw trajectories as a

consecutive sequence of spatio-temporal points sorted

in time (x, y,t), where x and y represent the spatial co-

ordinates of the moving object at timestamp t. LBSNs

collect information about their users’ visited places,

and a trajectory is a continuous sequence of check-

ins or Points of Interest (POI). Trajectories collected

from LBSN are called semantic trajectories. In the lit-

erature, there exist several research problems for tra-

jectory data. In this paper, we focus on the trajectory

classiﬁcation.

The trajectory classiﬁcation problem consists of

building a prediction model to classify a new trajec-

tory in a single-class or multi-class. The model is

trained and learns the patterns (or classes) from a his-

torical labeled trajectory (or sub-trajectory) data. Ba-

sic examples of trajectory classiﬁcation are: (i) deter-

mining the transportation mode of the moving object

like car, bus, bike, taxi, airplane, and train; (ii) de-

termine what a user’s next stopping point like home,

school, cafe, ofﬁce, and restaurant; (iii) identifying

who is the user of a trajectory (TUL problem).

In general, the trajectory classiﬁcation problem

is challenging because of: (1) the massive volume

of trajectory data continuously generated by multi-

ple users; (2) the complexity associated with the data

representation (how can we represent latitude, lon-

gitude, and timestamp features in our models with-

out losing information?); (3) the sequence of spatio-

temporal points can be sparse, for instance, LBSNs

usually contain samples in days; (4) the nature of

multiple dimensions: as technologies advances, more

and more properties are linked to a trajectory, such as

664

A. de Freitas, N., Coelho da Silva, T., Fernandes de Macêdo, J., Melo Junior, L. and Cordeiro, M.

Using Deep Learning for Trajectory Classiﬁcation.

DOI: 10.5220/0010227906640671

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 664-671

ISBN: 978-989-758-484-8

weather condition, POI category, adverse events, and

so on; and ﬁnally; (5) the number of the classes can be

much larger than the number of motion patterns (e.g.,

in TUL problem there may be more than one hundred

users).

The trajectory classiﬁcation topic has been ex-

ploited for over a decade (Lee et al., 2008). In the

beginning, most of the works in the literature focus on

extracting features from GPS trajectories (e.g., veloc-

ity, distance) and use them as input to machine learn-

ing classiﬁers such as Random Forest and Multilayer

Perceptron (Patterson et al., 2003; Zheng et al., 2008).

These approaches are limited to extract the best fea-

tures from the spatio-temporal data and suffer from

the curse of dimensionality.

More recently, studies involve Deep Learning

models for Trajectory-User Linking (TUL) problem

as (Gao et al., 2017; Zhou et al., 2018, 2019), and

they aim at identifying and linking trajectories to their

generating-users. These studies involving two Deep

Learning models are limited because they cannot han-

dle multidimensional data, but only the spatial dimen-

sion. We believe this the major limitation, considering

that more properties have been linked to trajectories

over the years and can improve the model’s perfor-

mance.

In this work, we propose the DeepeST (Deep

Learning for Sub-Trajectory classiﬁcation) model for

identifying the category from a large number of sub-

trajectories by jointly the embedding of many factors,

i.e., location, time, or any features associated with

a trajectory (or sub-trajectory). Our approach also

employs a Recurrent Neural Network (RNN), more

speciﬁcally, LSTM Schuster and Paliwal (1997) that

has been extensively used to process variable-length

input and can allow highly non-trivial long-distance

dependencies to be easily learned. We also apply the

Bi-directional LSTM (BLSTM) model to take into

account an effectively inﬁnite amount of context on

both sides of a sub-trajectory position. The main rea-

son we use RNNs is the capacity of these networks to

learn complex patterns from a sequence, unlike feed-

forward neural network models.

In this paper, we tackle the trajectory classiﬁca-

tion problem addressing two different classiﬁcation

tasks because we would like to evaluate DeepeST

using GPS and LBSNs data: (1) a problem called

Trajectory-User Linking (TUL), where the classiﬁca-

tion task is to link users to their sub-trajectories using

LBSNs trajectories. Correlating sub-trajectories with

users could help in identifying terrorists/criminals

from sparse spatio-temporal data (e.g., the transient

phone signals and check-ins) and is also helpful in

making better, more personalized recommendations

(Gao et al., 2017); (2) identify criminal patterns to

link criminal activities to GPS sub-trajectories inde-

pendent of the user’s trajectory. We trained DeepeST

with some categories, speciﬁed by the National De-

partment of Public Security, which are related to the

strategic POIs and the rules or speciﬁc constraints that

offenders need to follow, e.g − they need to charge

the ankle bracelet battery and to ensure the correct

functioning of the equipment that transmits his/her lo-

cation, that is, the signal that sends trajectory data to

the police department cannot be blocked; they need to

be at his/her home during the night. Figure 1 shows

an example of an offender’s sub-trajectories, where

the green, blue and purple points represent a sub-

trajectory link to a supermarket, its home and a block

signal, respectively. In this case, the signal block

was detected by the DeepeST based on the Spatio-

temporal pattern, where the same spatial coordinates

(latitude and longitude) are attributed to sequences of

sub-trajectories for a long time before a sub-trajectory

link to its home. In general, offenders prefer to block

the signal in their homes to avoid unsuspecting behav-

ior.

Figure 1: GPS Sub-trajectories classiﬁed by DeepeST.

The remainder of the paper is structured as follows:

Section 2 presents the related works and formally de-

ﬁnes the problem. Section 3 introduces our proposal.

Section 4 discusses the experimental evaluation. And

ﬁnally, Section 5 draws the ﬁnal conclusions.

2 TRAJECTORY

CLASSIFICATION

In this paper, we tackle the trajectory classiﬁca-

tion problem, which consists of classifying a sub-

trajectory into a label or class. We claim that we

classify sub-trajectories since our training set is de-

rived from the segmentation of trajectories as we ex-

plain later. For sake of brevity, hereinafter we will

use the term trajectory classiﬁcation in place of sub-

Using Deep Learning for Trajectory Classiﬁcation

665

trajectory classiﬁcation. So, given a set of trajectories

and labels, we want to build a model for predicting

and assigning such labels to every sub-trajectory. The

labels are a set of category features, be the users are

the owner of the trajectory, the transportation modes

(car, bus, bike, walk), or criminal activities.

Trajectory classiﬁcation is one of the widely stud-

ied problems on trajectory pattern mining over the

years. In the beginning, trajectory classiﬁcation fo-

cused on detecting patterns of mobility from raw tra-

jectories. Most of the supervised learning approaches

classify trajectories or sub-trajectories into categories

that represent activities (as hiking and walking) or dif-

ferent transportation modes (like a car, bike, and bus)

(Zheng et al., 2008; Patterson et al., 2003; Fang et al.,

2016). One of the ﬁrst methods for trajectory classi-

ﬁcation was TraClass, proposed by Lee et al. (2008)

that supports only the spatial dimension. Patel (2013)

extended the TraClass to support both the spatial and

the time dimensions. Basically, these works use ma-

chine learning methods demanding a feature extrac-

tion process to categorize raw trajectories or (sub-

trajectories) into different motion patterns consider-

ing features that are extracted from the spatial and

temporal dimensions like velocity, acceleration, and

distance.

More recently, studies have investigated Deep

Learning models to link trajectories to their generat-

ing users. TULER was the ﬁrst model introduced in

(Gao et al., 2017) for identifying and linking a large

number of check-in trajectories to their generating-

users using RNN based models. More speciﬁcally,

TULER receives a sequence of POIs as input and rep-

resents this sequence in a new low-dimensional space

(an embedding vector) similar to word embedding in

natural language (Mikolov et al., 2013). Finally, the

trajectory was characterized via trained RNN mod-

els to link them to their users. TULVAE was pro-

posed in (Zhou et al., 2018) after the TULER and en-

hanced in (Zhou et al., 2019), a generative model to

mine human mobility patterns, which aims at learning

the implicit hierarchical structures of trajectories and

alleviating the data sparsity problem with the semi-

supervised learning. TULVAE achieved a signiﬁcant

performance improvement for the TUL problem in

comparison to existing methods.

In summary, Machine learning methods can han-

dle trajectories, but they demand extraction features

process and suffers from the curse of dimension-

ality. On the other hand, Deep Learning models

as TULER and TULVAE minimize the sparse data

problem by representing the data in an embedding

vector but only support a sequence of POI identi-

ﬁer. These approaches do not cope with other es-

sential features, commonly used to describe seman-

tic trajectories. Therefore, we propose the DeepeST

model for identifying the category from a large num-

ber of raw (sub)trajectory, semantic (sub)trajectories,

or any (sub)trajectory of other domains. DeepeST

also employs a Recurrent Neural Network (RNN),

both LSTM (Schuster and Paliwal, 1997) and Bi-

directional LSTM (BLSTM), which are designed to

recognize the sequential characteristics of data and

thereafter using the patterns to predict the future sce-

nario. Moreover, DeepeST is able to minimize the

computational complexity since it also operates on

the low-dimensional to learn the underlying category

from the sub-trajectory data. Finally, DeepeST is ca-

pable of handling raw trajectories (collected by GPS)

and semantic trajectories (generate from LBSNs) with

several features, since the model is capable of receiv-

ing sequences of attributes linked to a trajectory.

2.1 Preliminaries

In this section, we will introduce some necessary no-

tations and the basic terminology before we formally

present the problem formulation of the trajectory clas-

siﬁcation problem. Table 1 presents a list of com-

monly used notations.

Table 1: A list of commonly used notations.

Notation Description

A a ﬁnite sequence of attributes linked to a trajectory

l a location point

L = {l

, ...l

} a ﬁnite set of location indexes

T = {t

, ...t

} a ﬁnite set of time slot indexes

= (l

, A

) a spatio-temporal position

T R

= [p

, ... p

len

] trajectory

= [p

, ... p

] a sub-trajectory

y a label of a sub-trajectory

Y = {y

, ...y

} a set of labels

S = {S

, ...S

} A ﬁnite set of sub-trajectories

There are several concepts for trajectory. In this work,

we consider trajectory raw (from GPS tracking de-

vices, it is a consecutive sequence of spatial coordi-

nates ordered by time) and semantic trajectories (gen-

erated by LBSNs that collect information about vis-

ited places of their users as a consecutive sequence of

check-ins or stop episodes, for instance) in a single

deﬁnition, formalized as follows:

Trajectory: Let a trajectory T R

be a sequence of

points sorted in time [p

, .., p

len

]. Here, p

(1 ≤ i ≤

len

) is a tuple (l

, A

), such that l

is a location point

at time t

, and A

= [a

, ..., a

] is a sequence of m at-

tributes linked to the trajectory (e.g velocity, acceler-

ation, geographic information, among others).

For sake of brevity, the location point l

is a spatial

coordinate e.g., latitude and longitude collected from

a GPS device, or l

can refer to check-in or stop lo-

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

666

cation (it can be a POI location, for instance). It is

worth to mention that if a trajectory is not linked to

any semantic information, then A

For simplicity, in this work, we represent each lo-

cation l

composed of latitude and longitude from a

trajectory T R in a spatial grid cell, however it could be

in any well-deﬁned geographical space. We also map

each timestamp t

to a time slot in T = {t

, ..., t

such that T ∈ IR

. A time slot could be a regu-

lar time interval, for instance, some minutes, hours,

days or weeks. Finally, in order to reduce the com-

putational complexity and capture richer knowledge

of sub-trajectory patterns from trajectories, we seg-

ment the trajectories into sub-trajectories. There are

several methods for trajectory segmentation based on

the shape of a trajectory, time interval, and semantic

meanings (Zheng, 2015). Since trajectory segmen-

tation is not at the core part of this work, we adopt

the simplest method based on the time interval to

trajectories. A trajectory [p

, p

, ..., p

], such that

(1 ≤ q

< q

< ... < q

≤ len

, where l

= l

k−1

+1) is

called a sub-trajectory of S

. We are now ready to for-

mulate our classiﬁcation problem for sub-trajectories.

2.2 Problem Statement

Given a set of sub-trajectories

S = {S

, S

, ..., S

}, the

task is to classify the category by linking each sub-

trajectory S

∈

S to a label y ∈ Y = {y

, ..., y

Notice that our problem is generic, Y can be a set

of transportation mode {car, bus, bike, walk}, a set of

users that are the owner of the trajectory or any cate-

gory feature of other domains.

3 DeepeST: DEEP LEARNING

FOR SUB-TRAJECTORY

CLASSIFICATION

In this paper, we propose a deep learning model,

called DeepeST, to the trajectory classiﬁcation prob-

lem. DeepeST receives as input a ﬁt sequence of fea-

tures contain location, time, and any other attributes

annotated in a sub-trajectory.

3.1 DeepeST Architecture

DeepeST architecture is composed of embedding lay-

ers to each input, a concatenation layer, a recurrent

layer (LSTM or BLSTM), and a fully connected layer

with softmax as the activation function. The overview

of DeepeST is illustrated in Figure 2.

DeepeST incorporates embedding layers to re-

ceives sequences from the sub-trajectory. An embed-

ding is a relatively low-dimensional space into which

you can translate high-dimensional vectors. Embed-

dings make it easier to do machine learning on large

inputs like sparse vectors representing words. There

are two main reasons we use sub-trajectories embed-

ding: (1) Traditional methods, such as one-hot encod-

ing, are binary, usually sparse (mostly made of zeros),

and very high-dimensional (same dimensionality as

the number of distinct labels) (Chollet, 2018); (2) the

frequency of location in sub-trajectories can follow

a power-law distribution. In general, likewise, words

embedding in natural language (Mikolov et al., 2013).

Sub-trajectory embedding alleviates the curse of di-

mensionality and maintains the input data’s proximity

with similar patterns in a new dimensional space. For

instance, suppose a student user has visited the POI

sequence [Home, Bus Station, University], while an-

other user has visited the POI sequence [Home, Sub-

way, University]. Note that the embeddings of Bus

Station and Subway will be similar because they hap-

pened in the same context (after Home and before

University).

Deep learning models TULER and TULVAE ex-

plore only the spatial dimension to embedding vector.

In these works, we explore the spatial and temporal

dimensions and other different features linked to sub-

trajectories. The more important features we linked to

a sub-trajectory, more information to improve classi-

ﬁcation accuracy and can be used directly and hence

save more time.

We used a recurrent layer that receives input from

a feature vector, but DeepeST presents embedding

layers to each sub-trajectory attribute. So, a concate-

nation layer is deﬁned between the embedding layers

and the recurrent layer to join embedding vectors in a

unique input features that will be used in the recurrent

layer, as shown in Figure 2.

The main reason we use a Recurrent Neural Net-

work (RNN) is the capacity of these networks to learn

complex patterns from a sequence, unlike feedfor-

ward neural networks. DeepeST employs a LSTM

(Hochreiter and Schmidhuber, 1997), which has been

extensively used to process variable-length input and

can allow highly non-trivial long-distance dependen-

cies to be easily learned. We also experimented Deep-

eST with the Bi-directional LSTM (BLSTM) model

(Schuster and Paliwal, 1997), which can take into ac-

count an effectively inﬁnite amount of context on both

sides of a sub-trajectory and eliminates the problem

of limited context that applies to any feed-forward

model. Both LSTM and BLSTM operate at the lo-

cation, time, and all attribute embedding levels to

Using Deep Learning for Trajectory Classiﬁcation

667

SoftmaxConcat

LSTM

BLSTM

DeepeST Model

. . .

,..., S

]

, ...,  l

]

, ..., tn]

[a1

, ...,  a1

]

[am

, ...,  am

]

. . .

Embedding to am

. . .

Embedding to a1

Embedding Time

Embedding location

Figure 2: DeepeST model to sub-trajectory classiﬁcation.

learn the underlying pattern (or label) from the sub-

trajectory data.

The output of LSTM or BLSTM passes by the

softmax function, which converts the recurrent layer’s

result into the set of probabilities to be assigned to

each label. Softmax takes as input a vector of real

numbers and normalizes it into a probability distribu-

tion consisting of o probabilities (for each label) pro-

portional to the input numbers’ exponentials.

After applying softmax, we have probabilities for

each label y ∈ Y on the interval [0, 1].

3.2 Optimization in DeepeST

Overﬁtting is a major problem in RNN due to a large

number of weights and biases. To alleviate overﬁt-

ting, we determined a dropout layer for regulariza-

tion. Dropout is a strategy radically different from

other approaches, since it changes the network struc-

ture itself, instead of the cost function. Suppose we

have a training set X and the corresponding desired

output y. Normally, we train by direct propagation of

X across the network, and then the backpropagation

algorithm computes the error to the gradient. When

we use a layer dropout, this process is modiﬁed. We

eliminate by randomly (and temporarily) some of the

neurons hidden in the network, but leave the input and

output neurons untouched. Heuristically, if we aban-

don different sets of neurons, we are training with var-

ious neural networks. Therefore, dropout can reduce

overﬁtting, whereas other networks adapt in different

ways.

4 EXPERIMENTS

In this section, we present the experimental evalua-

tion to evaluate DeepeST in terms of quality predic-

tion. We start by providing details about the datasets,

the baseline algorithms, the evaluation metrics, fol-

lowed by the experimental evaluation. For repro-

ducibility purposes, we made the source code avail-

able on GitHub

4.1 Datasets

To evaluate the performance of DeepeST for the clas-

siﬁcation trajectory problem, we conduct our experi-

ments on three datasets: (1) a public dataset that con-

tains check-ins of users extracted from Brightkite

between April 2008 and October 2010. (2) a pub-

lic dataset that contains trajectories of check-in ex-

tracted from Gowalla

between February 2009 and

October 2010; ﬁnally, (3) a private dataset of of-

fender trajectories extracted from GPS services. To

validate the models, we split the three datasets into

training (70%), validation (15%), and test sets (15%).

We shufﬂe sub-trajectory data, run the baselines algo-

rithms ten times for each dataset, and compared the

models using Accuracy, Macro Precision, Macro Re-

call, and Macro F1-Score.

For Brightkite and Gowalla datasets the classiﬁ-

cation task is to predict the corresponding user who

generated a given trajectory. We use a segmentation

based on time to created weekly sub-trajectories from

each user check-in. Notice that there are no overlap-

ping sub trajectories, there are weekly sub-trajectories

for each user (label). We also created a grid covering

all spatial points of trajectories with a cell of 30m

We selected only sub-trajectories of users who have

at least 15 weekly trajectories because we will have

at least two samples for each user in the validation set

and the test set. Finally, the Brightkite dataset con-

tains 4565 sub-trajectory samples to train, 996 to val-

idation, and 1085 to test. Gowalla dataset contains

2325 sub-trajectories sequences to train, 517 to vali-

dation, and 572 to test. Table 2 describes the attributes

of trajectory points to both Brightkite and Gowalla

datasets.

The Criminal dataset contains trajectories of of-

fenders in June 2019 that moved around Fortaleza,

https://github.com/nickssonarrais/ICAART2021

https://snap.stanford.edu/data/loc-Brightkite.html

https://snap.stanford.edu/data/loc-gowalla.html

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

668

Table 2: Description of the check-ins trajectories to

Brightkite and Gowalla.

Attributes Type Range/example N.

User Nominal {58186, ..., 58190} {197, 100}

Weekday Nominal {Monday, ..., Sunday} 7

Hour Numeric [0, 23] 24

Index grid Numeric {0, ..., 46458} {3742, 11345}

POI Nominal {dsda411, ..., ee8b8e} {4085,15977}

Subtraj ID Numeric {0, ..., 10000} {6646, 2335}

Ceara. The classiﬁcation task is to identify crimi-

nal activities, as well as what the criminal is doing

- at home, blocking the signal from the equipment

that transmits his/her location, selling stolen car parts,

in a hearing with the judge, among other categories.

In this dataset, we randomly selected trajectories of

ninety offenders. To criminal activity dataset we used

a different segmentation based on time and label to

avoid overlapping sub trajectories. We created sub-

trajectories group each user and criminal activity for

up thirty minutes. In order words, for each offender

the algorithm returns sub-trajectories contains a sin-

gle criminal activity for up to thirty minutes. We

deﬁne a 2D grid content cells of 30m

for an area

around Fortaleza city. Table 3 describes the attributes

of trajectory points to the Criminal dataset. Finally,

the Criminal dataset contains 116255 sub-trajectories

sequences to train, 24912 to validation and 24912 to

test.

Table 3: Description of the GPS trajectories to the Criminal

Dataset.

Attributes Type Range/example N.

Offender Nominal {58186, ..., 58190} 90

Weekday Nominal {Monday, ..., Sunday} 7

Hour Numeric [0, 23] 24

Index grid Numeric {0, ..., 46458} 36690

Criminal Activity Nominal {home, ..., blocked signal} 9

Subtraj ID Numeric {0, ..., 10000} 166079

4.2 Baselines Algorithms

We compare DeepeST with four state-of-the-art ap-

proaches from the ﬁeld of machine learning and deep

learning classiﬁcation: XGBoost (Chen and Guestrin,

2016), Random Forest (Breiman, 2001), BITULER

(Gao et al., 2017), and TULVAE.

To ﬁnd the optimal set of hyperparameters for

each model, we apply the grid-search technique to

combine several hyperparameters. For the Deep-

eST models (the one that uses LSTM, the one with

BLSTM), BITULER and TULVAE, we keep 64 as the

batch size and 0.001 as the learning rate and vary the

units (un) of the recurrent layer, the embedding size

to each attribute (es) and the dropout (dp). For TUL-

VAE, we also vary the latent variable (z). We deter-

mine an early stopping callback, that is a stop training

when, in our case, the accuracy has stopped improv-

ing. We set the early stopping as 20 for patience ar-

gument to minimize overﬁtting, i.e., the number of

epochs that produced the model’s accuracy with no

improvement after which training should be stopped.

For further details, we refer to Keras library

For the XGBoost model, we vary the number of

estimators (ne), the maximum depth of a tree (md),

the learning rate (lr), the gamma (gm), the fraction

of observations to be randomly samples for each tree

(ss), the sub sample ratio of columns when construct-

ing each tree (cst), the regularization parameters (l1)

and (l2). We also set the early stopping round to 20

for XGBoost.

For Random Forest, we vary the number of trees

(ne), the maximum number of features to consider

at every split (mf), the maximum number of levels

in a tree (md), the minimum number of samples re-

quired to split a node (mss), the minimum number of

samples required at each leaf node (msl), and ﬁnally,

the method of selecting samples for training each tree

(bs).

For more details about parameters, we shows refer

to Git Hub repository

4.3 Performance Comparison

There are two variations for DeepeST with respect

to the network layers: one with LSTM (DeepeST-

LSTM) and another one with BLSTM (DeepeST-

BLSTM). Our experiments tackle two main objec-

tives: evaluate individually the DeepeST models with

the state of art machine learning and deep learn-

ing models to the trajectory classiﬁcation problem

from check-ins, as we experimented for Gowalla and

Brightkite datasets. And ﬁnally, evaluate individu-

ally Deepest variation and the machine learning ap-

proaches to GPS-based trajectories. The idea behind

is to measure how resilient DeepeST is to learn the

categories in comparison to the baselines.

Table 4 and 5 shows the best set of parameters

from the grid search, it also summarizes the perfor-

mance comparison between the variants of DeepeST,

XGBoost, RandomForest, BITULER and TULVAE

for the Brightkite dataset. The two best values are

highlighted in bold and the third one is shown as un-

derlined. For what concerns to DeepeST variations,

a sub-trajectory S is a sequence with each of the fol-

lowing attributes (ig

, hr

, poi

, wk

), where ig is the in-

dex grid cell and poi is the POI identiﬁer, wk is the

weekday, and hr is the hour); BITULER and TUL-

VAE only deal with one feature, so the input is a se-

https://keras.io/

https://github.com/nickssonarrais/ICAART2021

Using Deep Learning for Trajectory Classiﬁcation

669

quence of POI identiﬁer as presented in Gao et al.

(2017); Zhou et al. (2019). For XGBoost and Ran-

domForest, a sub-trajectory is a unique concatenated

sequence [ig

, ig

, ..., wk

n−1

, wk

] with all attributes

(ig

, hr

, poi

, wk

From the results reported in Tables 4 and 5, we

can see that the embedding of four features com-

bined with RNN layers in sub-trajectory classiﬁcation

yields accuracy improvements over the baselines. In

summary, in comparison with machine learning ap-

proaches, we can notice that DeepeST outperforms

XGBoost and Random Forest across all metrics by up

to 11% in Brightkite and by up 31% in Gowalla con-

sidering F1-Score. DeepeST takes advantage of the

LSTM/BLSTM and operates at embedding levels to

learn the underlying user categories from check-ins

sub-trajectory data. It is worth to mention that RNN

models (LSTM/BLSTM) are proper models to learn

from temporal sequences as sub-trajectories. We can

notice that DeepeST also outperforms BITULER and

TULVAE across all metrics by up to 5% in Brightkite

and by up 7% considering F1-score. This is because

DeepeST built a more representative model using a

set of variables instead of only using the points of

interest identiﬁcation. As we can see, only POIs

identiﬁcation are not sufﬁcient to distinguish differ-

ent users. It is worth noting that the results could be

higher if there were more important features to sep-

arate the classes from different users in the dataset

(maybe features based on external events and weather

conditions). The expert’s view of the application do-

main can be essential to increase the performance of

the model. In this case, we applied only the spa-

tial and time features (weekday, hour, index grid,

and POI) extracted from the original Gowalla and

Brightkite datasets. DeepeST-LSTM only preserves

the past information because the only inputs it has

seen are from the past. DeepeST-BLSTM run your

inputs in two ways, one from past to future and one

from future to past. In our experiments, DeepeST-

BLSTM achieved slightly more signiﬁcant results

than DeepeST-LSTM for Brightkite dataset since

DeepeST-BLSTM takes into account an effectively

inﬁnite amount of context on both sides of a sub-

trajectory position and eliminates the problem of lim-

ited context that applies to any feed-forward model.

On the other hand, DeepeST-LSTM achieved slightly

more signiﬁcant results than DeepeST-BLSTM for

Gowalla dataset. It is important to highlight that

the results between DeepeST-BLSTM and DeepeST-

LSTM are very close. However, using future informa-

tion can usually be easier and faster for the network

to understand the next label.

Table 6 shows the best set of parameters from

the grid search, it also summarizes the performance

comparison between the variants of DeepeST, XG-

Boost, RandomForest. It is important to mention that

BITULER and TULVAE were not included in the

experiments since they are applied from a sequence

of POI identiﬁer (one-dimensional data) in check-ins

trajectories. To DeepeST variations, a sub-trajectory

S are sequence for each attribute in (ig

, hr

, wk

where ig is the Index grid, hr is the hour of day,

and wk is the weekday. For XGBoost and Ran-

domForest, a sub-trajectory is a unique concatenate

sequence [ig

, ig

, ..., wk

n−1

, wk

] with all attributes

(ig

, hr

, wk

). The two best values are highlighted in

bold and the third one is shown as underlined. We

can notice that DeepeST outperforms XGBoost and

Random Forest on all of the metrics by up 25% con-

sidering F1-macro. DeepeST again takes advantage

of the LSTM/BLSTM and operates at the location

and time embedding levels to learn the underlying

criminal categories from the offenders’ sub-trajectory

data. DeepeST-BLSTM achieved slightly more sig-

niﬁcant results than DeepeST-BLSTM for the Crimi-

nal dataset.

5 CONCLUSION

In this paper, we investigate the trajectory classiﬁ-

cation problem to classify a category from a set of

labels. A category can be anything, such as trans-

portation mode (car, bus, bike, walk), criminal activ-

ity (home, ..., blocked signal), or a user who is the

owner of the trajectory. We propose a new model,

called DeepeST (Deep Learning for Sub-Trajectory

classiﬁcation), to identify any category from a large

number of sub-trajectories extracted from GPS ser-

vices or generate from LBSNs. DeepeST employs a

Recurrent Neural Network (RNN), both LSTM and

Bi-directional LSTM (BLSTM), which operate on the

low-dimensional to learn the underlying category. We

attacked two trajectory classiﬁcation tasks and con-

ducted extensive experiments on three real datasets

to evaluate DeepeST performance with state-of-the-

art approaches from the ﬁeld of machine learning

classiﬁcation − XGBoost and Random Forest − and

Deep Learning − BITULER and TULVAE. DeepeST

achieves more expressive values of accuracy, preci-

sion, recall, and f1-score in all experiments. As future

directions, ﬁrst, we aim at analyzing more features in

our models (as the risk facilities (Chainey and Rat-

cliffe, 2013) for criminal activities and weather fea-

tures to Gowalla and Brightkite datasets). Moreover,

to provide operational advantages, optimization, and

innovation, e.g − incorporate modules to trajectory

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

670

Table 4: Results to Brightkite.

Method Best set of the grid search

Accuracy Precision Recall F1-Score

mean std mean std mean std mean std

DeepeST-LSTM un:400, es:100, dp:0.5 0,9591 0,0018 0,9644 0,0048 0,9514 0,0025 0,9512 0,0033

DeepeST-BILSTM un:100, es:100, dp:0.5 0,9632 0,0030 0,9658 0,0038 0,9563 0,0044 0,9557 0,0041

BiTULER un:100, es:100, dp:0.5 0,9372 0,0052 0,9417 0,0055 0,9211 0,0068 0,9234 0,0066

TULVAE un:100, es:300, dp:0.5, z:300 0,9452 0,0044 0,9439 0,0062 0,9325 0,0048 0,9308 0,0054

Random Forest ne:200, md:30, mss:5, msl:1, mf:auto, bs:False 0,8717 0,0043 0,8744 0,0090 0,8431 0,0060 0,8440 0,0072

XGBoost ne:2000, md:5, gm:0, ss:0.8, cst:0.5, l1:1, l2:100 0,8769 0,0047 0,8717 0,0059 0,8481 0,0063 0,8483 0,0065

Table 5: Results to Gowalla.

Method Best set of the grid search

Accuracy Precision Recall F1-Score

mean std mean std mean std mean std

DeepeST-LSTM un:100, es:400, dp:0.5 0,9760 0,0039 0,9821 0,0027 0,9744 0,0042 0,9750 0,0039

DeepeST-BILSTM un:200, es:100, dp:0.5 0,9739 0,0038 0,9798 0,0034 0,9727 0,0040 0,9723 0,0044

BiTULER un:300, es:400, dp:0.5 0,9122 0,0050 0,9274 0,0070 0,9101 0,0060 0,9078 0,0072

TULVAE un:100, es:300, dp:0.5, z:300 0,9159 0,0085 0,9338 0,0121 0,9111 0,0096 0,9105 0,0109

Random Forest ne:400, md:30, mss:2, msl:2, mf:sqrt, bs:False 0,7047 0,0088 0,7020 0,0139 0,6841 0,0093 0,6631 0,0109

XGBoost ne:2000, md:10, gm:0, ss:0.8, cst:0.5, l1:1, l2:100 0,6545 0,0112 0,6393 0,0197 0,6327 0,0134 0,6143 0,0152

Table 6: Results to Criminal Dataset.

Method Best set of the grid search

Accuracy Precision Recall F1-Score

mean std mean std mean std mean std

DeepeST-LSTM un:100, es:400, dp:0.5 0,9188 0,0010 0,8792 0,0075 0,8283 0,0043 0,8504 0,0040

DeepeST-BILSTM un:200, es:100, dp:0.5 0,9203 0,0013 0,8826 0,0071 0,8365 0,0052 0,8564 0,0026

Random Forest ne:400, md:30, mss:2, msl:2, mf:sqrt, bs:False 0,7917 0,0002 0,6806 0,0017 0,5515 0,0010 0,5910 0,0012

XGBoost ne:2000, md:10, gm:0, ss:0.8, cst:0.5, l1:1, l2:100 0,8121 0,0008 0,6671 0,0017 0,5765 0,0015 0,6084 0,0011

preprocessing and features creation. Finally, we aim

at studying how to improve accuracy by means of

other deep learning techniques, like attention mech-

anisms.

REFERENCES

Breiman, L. (2001). Random forests. Machine Learning,

45(1):5–32.

Chainey, S. and Ratcliffe, J. (2013). GIS and crime map-

ping. John Wiley & Sons.

Chen, T. and Guestrin, C. (2016). Xgboost: A scalable tree

boosting system. In Proceedings of the ACM SIGKDD

International Conference on Knowledge Discovery

and Data Mining, volume 13-17-Augu, pages 785–

794.

Chollet, F. (2018). Deep Learning mit Python und Keras:

Das Praxis-Handbuch vom Entwickler der Keras-

Bibliothek. MITP-Verlags GmbH & Co. KG.

Fang, S.-H., Liao, H.-H., Fei, Y.-X., Chen, K.-H., Huang,

J.-W., Lu, Y.-D., and Tsao, Y. (2016). Transportation

modes classiﬁcation using sensors on smartphones.

Sensors, 16(8):1324.

Gao, Q., Zhou, F., Zhang, K., Trajcevski, G., Luo, X., and

Zhang, F. (2017). Identifying human mobility via

trajectory embeddings. In IJCAI, volume 17, pages

1689–1695.

Hochreiter, S. and Schmidhuber, J. (1997). Long Short-

Term Memory. Neural Computation, 9(8):1735–1780.

Lee, J.-G., Han, J., Li, X., and Gonzalez, H. (2008).

Traclass: trajectory classiﬁcation using hierarchical

region-based and trajectory-based clustering. Pro-

ceedings of the VLDB Endowment, 1(1):1081–1094.

Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013).

Efﬁcient estimation of word representations in vector

space. arXiv preprint arXiv:1301.3781.

Patel, D. (2013). Incorporating duration and region associ-

ation information in trajectory classiﬁcation. Journal

of Location Based Services, 7(4):246–271.

Patterson, D. J., Liao, L., Fox, D., and Kautz, H. (2003). In-

ferring high-level behavior from low-level sensors. In

International Conference on Ubiquitous Computing,

pages 73–89. Springer.

Schuster, M. and Paliwal, K. K. (1997). Bidirectional re-

current neural networks. IEEE Transactions on Signal

Processing, 45(11):2673–2681.

Zheng, Y. (2015). Trajectory Data Mining. ACM Transac-

tions on Intelligent Systems and Technology, 6(3):1–

41.

Zheng, Y., Liu, L., Wang, L., and Xie, X. (2008). Learning

transportation mode from raw gps data for geographic

applications on the web. In Proceedings of the 17th

international conference on World Wide Web, pages

247–256. ACM.

Zhou, F., Gao, Q., Trajcevski, G., Zhang, K., Zhong, T., and

Zhang, F. (2018). Trajectory-user linking via varia-

tional autoencoder. In IJCAI, pages 3212–3218.

Zhou, F., Yin, R., Trajcevski, G., Zhang, K., Wu, J., and

Khokhar, A. (2019). Improving human mobility iden-

tiﬁcation with trajectory augmentation. pages 1–31.

Springer.

Using Deep Learning for Trajectory Classiﬁcation

671