TrajCogn: Leveraging LLMs for Cognizing Movement Patterns and Travel Purposes from Trajectories

1 Introduction

A spatio-temporal (ST) trajectory is a sequence of timestamped locations, represented as T = ⟨(l1, t1), (l2, t2), . . . , (ln, tn)⟩. It tracks the movements of an individual or object in a geographical space. With the widespread use of mobile phones, car navigation systems, location-based services, and online map services, ST trajectories are being recorded and collected from various sources [Zheng et al.2008]. They enable a wide range of spatio-temporal data mining tasks and applications, including trajectory prediction [Feng et al.2018, Kong and Wu2018], POI recommendation [Chen et al.2022a, Chen et al.2025], travel time estimation [Wang et al.2018, Shen et al.2025], trajectory similarity measurement [Yao et al.2019, Fang et al.2022a, Li et al.2018a, Yang et al.2021a], and trajectoryuser linking [Zhou et al.2018,Miao et al.2020].

∗Corresponding author.

To enhance the use of ST trajectories in tasks and applications, it is essential to develop a trajectory learning method that 1) effectively captures the information embedded in the trajectory, specifically, movement patterns that describe how the individual or object moves from one location to another and travel purposes that indicate the underlying reason or motivation for the movement; and 2) accurately performs a variety of downstream tasks, reducing the need for designing a separate method for each task and application. methods primarily use self-supervised learning [Dai and Le2015,Devlin et al.2019] to map trajectories into embedding vectors, training models from scratch [Yang et al.2023, Fu and Lee2020, Jiang et al.2023]. However, these models face limitations due to the complexity of trajectory information and constraints in model capacity, as well as the size and quality of available datasets.

On the other hand, versatile models have been highly successful in the domain of natural language processing (NLP), showcasing promising results on various downstream tasks [Radford et al.2019,Devlin et al.2019,Raffel et al.2020, Du et al.2022]. These models, often referred to as large language models (LLMs), benefit mainly from their large capacity, abundant large-scale corpus datasets, and well-thoughtout prompt engineering [Brown et al.2020]. Given the similarities between trajectories and sentences in NLP, LLMs hold promise for enhancing trajectory learning models. Trajectory points share spatio-temporal correlations akin to word contexts, with movement patterns and travel purposes resembling word and sentence semantics. However, adapting LLMs for trajectory modeling presents challenges.

First, LLMs are incapable of processing the raw features in trajectories. LLMs are designed for discrete word sequences, but trajectories involve continuous and discrete spatio-temporal data like GPS coordinates and timestamps, which are difficult for LLMs to process. Second, LLMs are unable to extract the movement patterns and travel purposes directly from trajectories. For example, a moving object might go straight, accelerate, or turn, as shown in Figure 1. These movement patterns can be derived from changes in coordinates, timestamps, and velocities, while travel purposes are linked to the trajectory's origin and destination. The example trajectory which starts near residential buildings and ends near an office building suggests commuting. LLMs, however, focus on word semantics and are not equipped to interpret these spatio-temporal patterns or purposes.

To address these challenges and effectively leverage LLMs to construct a versatile trajectory learning model, we propose a novel approach named Trajectory Cognition (TrajCogn). TrajCogn employs a trajectory prompt to integrate movement patterns and travel purposes, and uses task-p-tuning to adapt to various tasks and make accurate predictions. It includes a trajectory semantic embedder to enable LLMs to process spatio-temporal features and extract movement patterns and travel purposes. Additionally, a cross-reconstruction pretext task based on self-supervised learning is implemented to enhance TrajCogn's ability to learn from trajectory data. Our contributions are summarized as follows:

• We introduce TrajCogn, a model that migrates LLMs to cognize movement patterns and travel purposes from trajectories, accurately performing different downstream tasks despite small dataset limitations.
• We develop a novel trajectory prompt to integrate movement patterns and travel semantics into LLMs, enhancing adaptability to various tasks.
• We propose a trajectory semantic embedder to process spatio-temporal features, allowing LLMs to extract movement patterns and travel semantics explainably.
• We conduct extensive experiments on three real-world trajectory datasets, and the results demonstrate TrajCogn's versatility and strong performance across different tasks.

2 RelatedWorks

Trajectory Learning Models aim to extract information from trajectories and perform various related tasks. Compared to task-specific prediction models [Feng et al.2018, Yao et al.2019, Fang et al.2022a, Wang et al.2018, Chen et al.2022b, Li et al.2018b], which are end-to-end trained for one specific task, trajectory learning models are versatile and useful in modern intelligent transportation applications that usually involve multiple tasks.

Most existing efforts adhere to the self-supervised learning approach. Earlier research commonly used RNNs to reconstruct discrete locations [Li et al.2018a,Liu et al.2020,Fu and Lee2020] or continuous movement features [Yao et al.2017] of trajectories based on auto-encoding [Hinton and Salakhutdinov2006] and variational auto-encoders [Kingma and Welling2014]. Additionally, methods like CTLE [Lin et al.2021] and Toast [Chen et al.2021], based on transformers [Vaswani et al.2017] and Masked Language Model (MLM) tasks [Devlin et al.2019], treat trajectory points as tokens in a sentence. Furthermore, contrastive learning methods such as PIM [Yang et al.2021b], TrajCL [Chang et al.2023], and MMTEC [Lin et al.2023] implicitly model the travel purpose of a trajectory. More recently, methods combining multiple approaches have been developed. START [Jiang et al.2023] leverages both MLM tasks and SimCLR [Chen et al.2020], while LightPath [Yang et al.2023] incorporates a reconstruction task and a contrastivestyle rational reasoning task.

Since these methods are self-supervised and trained from scratch, their performance heavily relies on the size and quality of the training datasets, which often have limitations. Despite the achievements of existing methods, further efforts are needed to enhance the performance of trajectory learning models.

Cross-domain Application of LLMs. The versatility and superior performance of large language models (LLMs) in the NLP domain have led to efforts to adopt LLMs in other fields to enhance performance. GPT4TS [Zhou et al.2023] uses LLMs by freezing the self-attention feed-forward layers. Time-LLM [Jin et al.2023] introduces a reprogramming framework. LM4VisualEncoding [Pang et al.2023] incorporates a frozen transformer block from an LLM as a general-purpose visual encoder layer. UrbanGPT [Li et al.2024a] employs spatio-temporal encoding combined with LLM instruction-tuning for generalized zero-shot prediction. FlashST [Li et al.2024b] introduces an innovative prompttuning framework designed to adapt pre-trained models for spatio-temporal prediction tasks.

Although these studies provide valuable insights, their methods cannot be directly applied to trajectory learning. Trajectory data has unique spatio-temporal features that require tailored approaches and considerations.

3 Preliminaries

3.1 Definition

Definition 1 (Road Network). A road network is represented as a directed graph G = (V, E). V is a set of |V| vertices, and each vertex vi ∈ V represents an intersection between road segments or the end of a segment. E is a set of |E| segments, where each segment si ∈ E represents a road segment linking two vertices.

Definition 2 (Trajectory). A trajectory T is a sequence of timestamped locations, represented as T = ⟨(l1, t1), (l2, t2), · · · , (ln, tn)⟩. Here, each location li is represented by its latitude and longitude coordinates, i.e., li = (llat i , llng i ). The timestamp ti indicates when li is visited. To simplify, we denote the i-th trajectory point (li, ti) as τi.

Definition 3 (Point of Interest, POI). A POI is a particular location that individuals may find valuable or intriguing. It is denoted as p = (l, n, a), where l represents its coordinates, n indicates its name, and a refers to its address.

3.2 Problem Statement

Trajectory Learning. The objective is to develop a trajectory learning model fΘ with a set of learnable parameters Θ. This model takes a trajectory T as input and extracts information from it. Subsequently, this model can adapt to various downstream tasks by accurately predicting the required outputs y for the task at hand, denoted as ˆ y = fΘ(T ). For example, in travel time estimation, y and ˆ y represent the ground truth and the estimated travel time, respectively.

3.3 Pre-trained Language Model

In this work, a Large Language Model (LLM) refers to a Transformer-based language model pre-trained on corpus datasets. It consists of four essential functions. Formally,

where ◦ represents the composition of functions. Specifically, a LLM consists of a tokenizer (Tok) to break down text into discrete tokens, a word token embedding layer (WTE) that converts the tokens into numerical vectors to capture their linguistic features, a transformer block (TransBlk) that further processes the vectors to capture their contextual relationships, and a prediction head (LMHead) that is respondible for making specific predictions, such as generating the next word in a sequence. In a LLM, the dimension of the word token embedding is denoted as d.

4 Methodology

4.1 Overview

Figure 2 shows the overall framework of TrajCogn. It is implemented in the following four steps:

1. Trajectory and POI Feature Extraction: Given a trajectory T , we perform map-matching and calculate highorder features such as velocity, acceleration, and direction to expand its features, denoted as ˜ T . We also extract the address and name features of POIs near the trajectory's origin and destination.
2. Trajectory Prompt Construction: We integrate the extracted features into one sequence, called the trajectory prompt. This prompt also includes a task-p-tuning mechanism-based suffix to enable adaptation to various tasks.
3. Trajectory Prompt Embedding: We map the trajectory prompt into a sequence of d-dimensional embeddings with a trajectory semantic embedder. This embedder is designed to enable LLMs to process spatio-temporal features and effectively extract movement patterns and travel purposes with explainability.
4. Model Training and Task Adaptation: We process the embedding sequence with a LLM Encoder for Trajectory (LET). The last point of the output sequence of LET is used for performing downstream tasks. The learnable parameters in the model are refined by integrating a crossreconstruction pretext task and further optimized with a dedicated objective function for each specific downstream task.

The following sections provide a detailed explanation of the steps in TrajCogn.

4.2 Trajectory Prompt

As illustrated in Figure 1, movement patterns in a trajectory can be represented by positions on the road network and variations in spatio-temporal features. Travel purposes can be inferred from the functionalities of locations near the OD points, and the address and name features of a POI indicate its functionalities.

To incorporate the movement patterns and travel purposes of a trajectory, we first extract spatio-temporal and POI features from the trajectory, as shown in Figure 2(a). To integrate these features into LLMs, we introduce a Trajectory Prompt, as illustrated in Figure 2(b). This prompt fuses natural language and the extracted features into a sequence. Furthermore, to adapt the model to different downstream tasks, we introduce a task-p-tuning mechanism, which provides a specific suffix for each task.

Trajectory and POI Feature Extraction

Given a trajectory T = ⟨(l1, t1), (l2, t2), . . . , (ln, tn)⟩ and a road network G, we apply the Leuven Map Matching (LMM) algorithm to map each trajectory point τi onto the road network, denoted as LMM(τi, G) = (li, si, ti), where si is the road segment containing li. We then calculate the velocity vi, acceleration ai, and direction θi of each point based on differences between consecutive points. The trajectory point ˜ τi = (li, si, ti, vi, ai, θi) is formed with these spatio-temporal features, setting the velocity and acceleration of the last point ˜ τn to 0. The final trajectory ˜ T = ⟨ ˜ τ1, ˜ τ2, . . . , ˜ τn⟩ is obtained. For POI feature extraction, we first identify the origin l1 and destination ln of the trajectory T . Using the Ball Tree algorithm, we retrieve the closest NPOI POIs to l1, denoted as PO = {p(1) o , . . . , p(NPOI) o }, ordered by distance from the origin. The same process is applied to retrieve POIs around ln, denoted as PD. For each POI p ∈ PO ∪PD, we extract its address p.a and name p.n features, both represented as word lists.

Trajectory Prompt Construction

The trajectory prompt is made up of four parts. First, the ⟨Head Part⟩ introduces the context by stating "The trajectory happened on {day-in-week} at {hour} o'clock." Second, the ⟨POI Part⟩ provides details about the POIs near the origin and destination, saying "starts near: {p(1) o , p(2) o , . . . , p(NPOI) o }, ends near: {p(1) d , p(2) d , . . . , p(NPOI) d }." Third, the ⟨Trajectory Part⟩ includes the extracted features of the trajectory points, described as "passes through {˜ τ1, ˜ τ2, . . . , ˜ τn}."

Finally, the ⟨Suffix Prompt⟩ is constructed using the taskp-tuning mechanism, which combines hard and soft components [Han et al.2022]. The hard component contains taskspecific words, while the soft component is a task-specific token represented as [Token]. For example, for travel time estimation (TTE), the suffix prompt would read "the total travel time is [TTEToken]." For destination prediction (DP), it would be "the destination is [DPToken]." For trajectory classification, it would be "the user connected to this trajectory is [TCToken]."

4.3 Trajectory Semantic Embedder

In order to equip LLMs with the ability to process the spatiotemporal features in the trajectory prompt, we propose the Trajectory Semantic Embedder, demonstrated in Figure 2(c).

Spatio-temporal Feature Embedding

The spatio-temporal features in the trajectory prompt are embedded into a d-dimensional space. For discrete road segments si, an index-fetching embedding module EE ∈ R|E|×d is used. Timestamps ti are embedded using modules Edw ∈ R7×d and Eh ∈ R24×d for day-in-week and hour features, respectively. Continuous features of trajectory points are embedded using a one-dimensional convolution, inspired by previous studies [Wang et al.2018, Liang et al.2022], to model movement patterns. The continuous embedding vector of τi is obtained as follows:

where k denotes the kernel size, τ con =

i , llat i , vi, ai, θi, ti) presents the continuous features.

i (llat

Finally, the embedding vector ei of the i-th trajectory point τi is derived as follows:

Movement Pattern Semantic Projection

To improve the model's understanding and interpretability of movement patterns, each embedding vector ei is projected onto a semantic-rich textual space, as shown in Figure 3. This space is defined by a set of descriptive words M and a set of vitual words A. The descriptive words are listed in Table 1 and the virtual words are initialized randomly and trained end-to-end. These words form anchor words, whose embeddings are concatenated into Ean. Using multi-head attention [Vaswani et al.2017], where ei serves as the query and Ean acts as both the key and value, ei is projected onto this space, resulting in ˜ zi. The final embedding vector zi is obtained by adding a residual connection through a two-layer MLP:

The sequence of these embeddings for the trajectory is denoted as ZT = ⟨z1, z2, . . . , zn⟩.

POI Feature Embedding

The travel purpose is inferred by analyzing the functionalities of POIs near the origin and destination points. To model these functionalities, we derive embeddings for POIs based on their address and name features. For the closest POIs to the origin or destination, p(1) o or p(1) d , embeddings are obtained by concatenating their address and name:

where ∥ denotes list concatenation. For other POIs, only the name is used for embedding: ETok(p) = WTE ◦ Tok(p.n).

Sequence of Trajectory Prompt Embeddings

After obtaining embeddings of spatio-temporal and POI features, the remaining textual components in the trajectory prompt are embedded using WTE ◦ Tok. Then, we concatenate the embeddings into a sequence in the same order as their raw features appear in the prompt. For example, the embeddings of the trajectory part are obtained as follows:

The embeddings of the ⟨Head Part⟩, ⟨POI Part⟩, and ⟨Suffix Prompt⟩ are denoted as Zh, Zp, and Zs, respectively. Finally, the sequence of trajectory prompt embeddings is gathered as follows:

4.4 LLMEncoder for Trajectory

We use the transformer block TransBlk from an LLM as the backbone for the proposed LLM Encoder for Trajectory (LET). To better adapt the pre-trained TransBlk to trajectory learning, we employ the Low Rank Adaptation (LoRA) algorithm [Hu et al.2022], adding extra parameters to TransBlk.

Construction of LET

As illustrated in Figure 2(d), all parameters in the TransBlk are kept fixed, while we introduce a new learnable parameter matrix in every self-attention block by applying LoRA algorithm. The proposed LET can be expressed as follows:

LET takes the embedding sequence Z from Equation 7 as input, and outputs a sequence of hidden vectors H = LET(Z),H ∈ RL×d , where L represents the length of Z.

Adaptation to Downstream Tasks

LET adapts to different downstream tasks using the task-ptuning mechanism described in Section 4.2. Specifically, the hidden vector corresponding to the task-specific token, i.e., the L-th hidden vector htask ∈ Rd in H, can be utilized to perform downstream tasks.

In this study, we present Travel Time Estimation (TTE), Destination Prediction (DP), Similar Trajectory Search (STS) and Trajectory Classification (TC) tasks for evaluation, as shown in Figure 2(e).

The TTE task aims to estimate travel time using spatial features and departure time, excluding time-related features. A two-layer MLP is used to predict travel time: ˆ yTTE = MLPTTE(htask).

The DP task aims to predict the destination road segment, excluding the last 5 trajectory points and nearby POIs to prevent data leakage. A two-layer MLP predicts the segment: ˆ yDP = argmaxs(ˆp), ˆp = Softmax(MLP(htask)).

The STS task aims to find the most similar trajectory using cosine similarity on htask. We construct ground truth by selecting 1,000 test trajectories, using odd-numbered points as queries T q and even-numbered points as targets T t . We exclude the 10 closest trajectories and randomly select 5,000 others as negatives. Distances are calculated by downsampling to a uniform length and using mean square error, as per [Fang et al.2022b].

The TC task aims to classify trajectories by driver, using a two-layer MLP to predict the driver: ˆ yTC = argmaxd(ˆp), ˆp = Softmax(MLP(htask))

4.5 Model Training

We propose a cross-reconstruction pretext task to train the learnable parameters in the model, helping it adapt to trajectories. Before performing a specific task, the model can be further fine-tuned with supervision for that task.

Cross-reconstrution Pretext Task

The proposed pretext task involves reconstructing each trajectory point given ⟨Head Part⟩ and ⟨POI Part⟩, and reconstructing each POI given ⟨Head Part⟩ and ⟨Trajectory Part⟩.

As shown in Figure 4, for trajectory points, LET processes embeddings up to the i - 1 step as Htraj,i-1 = LET(Zh∥Zp∥Zt,:i-1), and predicting features with a twolayer MLP. The loss Ltraj combines the cross-entropy loss of the predicted segments and the MSE loss of the predicted continuous features. For POIs, LET processes embeddings similarly as HPOI,i-1 = LET(Zh∥Zt∥Zp,:i-1), predicting features with the LLM's LMHead. The loss LPOI is the crossentropy of predicted features. The total pretext loss is

Teacher-forcing is used to enhance training efficiency.

Task-specific Fine-tuning

When performing a specific task, the proposed model can be fine-tuned with the task's supervision to further improve prediction accuracy.

For the TTE task, the loss function is defined with mean square error (MSE) loss. For the DP and TC task, the loss function is defined with the cross-entropy loss. The STS task does not require fine-tuning, using the hidden state htask from the pretext task directly.

Table 3: Performance of variants of TrajCogn.

al.2023] incorporates a time-aware encoder and GAT, trained with MLM and SimCLR-based contrastive tasks. LightPath [Yang et al.2023] uses a sparse path encoder with path reconstruction and cross-view contrastive tasks.

5.3 Settings

For each dataset, we divide the trajectories into training, validation, and testing sets in an 8:1:1 ratio, ordered by departure time. Pre-training for the cross-reconstruction task and embedding methods lasts 20 epochs, with early stopping for downstream predictors based on validation performance.

All models are implemented using PyTorch [Paszke et al.2019]. We choose GPT2 [Radford et al.2019] as the foundation LLM to develop our model and obtain addresses and names of POIs using Amap APIs 4 . Our source codes are available at https://github.com/Zeru19/TrajCogn. Key hyperparameters for TrajCogn are NA = 15, K = 5, r = 8, and NPOI = 3, optimized based on Acc@1 and Recall for destination prediction on Chengdu's validation set. The Adam optimizer is used with a learning rate of 1e-4 for our method and 0.001 for others. Experiments run on Ubuntu 22.04 with Intel(R) CPUs and nVidia(R) TITAN RTX GPUs, repeated 5 times to report mean and deviation of metrics.

5.4 Performance Comparison

Overall Performance

Table 2 present a comprehensive comparison of the performance of all task-adaptable trajectory learning methods

4 https://lbs.amap.com/

5 Experiments

5.1 Datasets

In our experiments, we use three real-world datasets called Chengdu, Xi'an and Porto. Chengdu and Xi'an datasets were released by Didi 1 and consist of GPS trajectories recorded by taxis in Chengdu and Xi'an, China. Porto is an open-source dataset released for a Kaggle competition 2 . Trajectories with fewer than 6 points were excluded. Road networks from OpenStreetMap 3 were used for map-matching. An overview of the dataset statistics is shown in Appendix A.

5.2 ComparisonMethods

We compare the proposed method with several state-of-theart trajectory learning methods. Traj2vec [Yao et al.2017] calculates features with sliding windows and trains with an auto-regressive pretext task. T2vec [Li et al.2018a] pretrains the model using a denoising auto-encoder to reconstruct trajecotories. TremBR [Fu and Lee2020] constructs an RNN-based seq2seq model for recovering road segments and times. CTLE [Lin et al.2021] utilizes a bi-directional Transformer with MLM tasks for location and hour predictions. Toast [Chen et al.2021] employs a context-aware node2vec model with MLM and sequence discrimination tasks. TrajCL [Chang et al.2023] introduces a dual-feature, attentionbased encoder trained with InfoNCE loss. START [Jiang et across four tasks and two datasets. The TTE task was excluded for Porto due to potential data leakage from equalinterval sampling, and trajectory classification was limited to Porto due to insufficient trajectories per driver in Chengdu. Our proposed method consistently outperforms the others and performs well across tasks, providing evidence that it is an advanced task-adaptable trajectory learning method.

1 https://gaia.didichuxing.com/

2 https://www.kaggle.com/c/pkdd-15-predict-taxi-servicetrajectory-i

3 https://www.openstreetmap.org/

Methods like Traj2vec, T2vec, and TremBR use RNNbased frameworks but lack crucial spatio-temporal features, leading to poor downstream performance. CTLE and Toast use bi-directional Transformers with MLM tasks [Devlin et al.2019] but miss essential continuous features and travel purposes, affecting their STS task performance. TrajCL, START, and LightPath use contrastive learning, aiding STS performance, but overlook POI functionalities and struggle with movement pattern extraction, resulting in unsatisfactory TTE and DP results. Our method utilizes LLMs for adaptable trajectory learning, effectively extracting movement patterns and incorporating POI functionalities and travel purposes through a trajectory prompt. These strengths lead to superior performance across tasks.

Scalability

To compare the scalability of the proposed model against START, one of the state-of-the-art models, we refine our model using varying proportions of the training data: 100%, 60%, and 20% for the destination prediction task on the Chengdu dataset. We use the START model as a reference point with an identical learning rate of 5 × 10-4 for comparison. The results are presented in Figure 5. It can be seen that our model demonstrates faster progress and achieves superior performance with less data compared to START. This shows that our model can be adapted to downstream tasks with lightweight finetuning.

5.5 Model Analysis

Effectiveness of Components

To evaluate components implemented in TrajCogn, we compared the performance of the complete model with the following variants. w/o PT excludes the cross-reconstruction pretext task, training directly on downstream tasks. w/o POI removes the ⟨POI Part⟩ from the trajectory prompt. w/o Conv replaces the convolution operator in the trajectory semantic embedder with a fully connected layer. w/o PSP excludes the pattern semantic projector, using only the trajectory point embedding ei; and w/o M omits the movement pattern vocabulary, relying solely on virtual anchor words.

We measured the performance of these variants on the Chengdu dataset, and the results are presented in Table 3. w/o PT shows performance degradation, proving the contribution of the cross-reconstruction pretext task to TrajCogn. The worse performance witnessed by w/o POI demonstrates the effectiveness of integrating POI information. w/o Conv, w/o PSP, and w/o M all have worse performance compared to full, showing that the removed components all contribute to TrajCogn's performance.

Attention Map Visualization

To demonstrate our model's ability to extract interpretable movement patterns, we visualize attention scores in the pattern semantic projector, as shown in Figure 6. Each subfigure shows the trajectory with marked points and subtrajectories on the left, and their attention maps on the right. We observed that trajectory points' movement patterns correspond to specific anchor words. Terms like "turn," "slow," and "steady" reveal the semantics of these patterns. For instance, turns increase attention for "turn," while "slow" and "stationary" suggest slowmovement. However, these words do not always fully capture true movement semantics, highlighting the need for accurate labeled data for better alignment.

Additional Model Analysis

Further analysis of TrajCogn is detailed in the Supplementary Material. Appendix B covers its performance on the Xi'an dataset. Appendix C examines hyper-parameter effectiveness. Appendix D explores TrajCogn's efficiency. Appendix E discusses the impact of anchor words, while Appendix F looks at additional features. Appendix G analyzes the foundation LLM's effectiveness.

6 Conclusion

We propose TrajCogn, a novel trajectory learning model that leverages LLMs to model trajectories and accurately perform various trajectory-related tasks. TrajCogn introduces a trajectory prompt that combines movement patterns and travel purposes, enabling task adaptability. It also includes a trajectory semantic embedder for processing spatio-temporal features, allowing for effective and explainable extraction of movement patterns and travel purposes. Experiments on real-world datasets confirm TrajCogn's superior performance.

Ethical Statement

There are no ethical issues.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (No. 62272033).

Contribution Statement

Zeyu Zhou and Yan Lin contributed equally to this research.