User Trajectory Prediction Unifying Global and Local Temporal Information

arXiv:2508.02161 · cs.LG · Submitted 2025-08-04 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "User Trajectory Prediction Unifying Global and Local Temporal Information".

Jane: The gist: A trajectory prediction model based on multilayer perceptron (MLP), multi-scale convolutional neural network (MSCNN),

Tom: First, who's behind it and why it matters.

Title and authors: Tom: We've been talking about the MMCTP model and how it blends global and local information to improve trajectory predictions, but let's look at what the paper actually claims about this work.

Jane: Right. The title itself, "User Trajectory Prediction Unifying Global and Local Temporal Information," sets up exactly what they are trying to achieve here.

Lu: Essentially, they are tackling the challenge that trajectory data has two distinct layers: broad patterns and fine-grained local interactions happening at different speeds.

Tom: They propose the MMCTP model, which uses the MLP for global information extraction and an MSCNN for local information extraction.

Meng: So it’s a hybrid architecture, combining those two specific neural network types to handle those two different types of temporal data.

Jane: And then they use cross-attention to merge these two streams together in a way that lets the model weigh the global and local context appropriately.

Lu: The core idea is that sequential approaches often interfere with each other, but this model avoids that interference by modeling them separately and then fusing them later.

Tom: And the experimental results are quite compelling; they show this MMCTP significantly lowers prediction errors compared to a baseline model called ModernTCN.

Jane: Specifically, they report reducing mean squared error by five point zero four percent and mean absolute error by four point three five percent over twelve steps of prediction, which is a concrete improvement metric.

Lu: That reduction in error is significant when you consider the complexity of predicting human movement over time series data where things are nonlinear and noisy.

Tom: And they’ve also done some ablation studies to show that taking away either the global or local part actually hurts performance, confirming that both components are essential.

Meng: That confirms their hypothesis—you need both for a good result, not just one or the other.

Jane: So this isn't just about adding more layers; it’s about fundamentally changing how we structure the information flow to capture different temporal aspects of movement.

Lu: It shows that modeling multi-scale characteristics is a vital step for accurately predicting complex behaviors in time series data, which is a big concept.

Tom: So, as you listen to this show, think about how trajectory prediction systems are moving from simple models to these more nuanced hybrid approaches.

The paper's summary: Jane: Now let’s get into the technical details of what the MMCTP model actually does under the hood.

Tom: The paper describes two distinct extraction modules first: the global temporal information extraction module using an MLP and a local temporal information extraction module using an MSCNN.

Lu: For that MLP part, they process the input data through RevIN normalization to handle distribution shifts, and they inject positional embeddings with sine and cosine functions for spatial location info.

Meng: That means they’re carefully preparing the input before feeding it into the MLP to get those overall patterns.

Jane: Then, in parallel, the MSCNN module models interactions within a local temporal range by using convolutional kernels of various sizes.

Tom: The idea behind using diverse kernel sizes is specifically to capture behavioral patterns at different temporal resolutions simultaneously, which is key for multi-scale modeling.

Lu: And finally, they use the cross-attention mechanism to fuse the outputs from these two streams—the global information and the local information matrix I.

Jane: The attention calculation involves projecting the local information into a query matrix Q and the global information into key K and value V, using scaled dot product similarity scores.

Tom: They use multiple attention heads for this fusion, which allows each head to focus on different aspects of the input data independently before they are concatenated and transformed by a weight matrix WZ.

Meng: It’s a sophisticated way to combine the information; it’s not just dumping one output into another.

Jane: And they also mention that the overall time complexity of this model can be simplified to O(q), which is pretty efficient for real-time use.

Lu: That O(q) complexity suggests that the computational cost scales very well, which is important when you’re deploying these kinds of models.

Tom: So, to recap, it’s a layered approach where you extract global and local features separately and then intelligently fuse them with attention.

The paper's improvements: Tom: Beyond just the accuracy gains, the paper highlights several key improvements they suggest for this type of trajectory prediction model.

Jane: One major improvement is the explicit modeling of multi-scale characteristics, which helps it adapt to different temporal resolutions in user behavior.

Lu: The MSCNN’s use of varied kernel sizes directly addresses this by allowing it to capture those diverse behavioral patterns that single-scale models miss.

Meng: And they also show that incorporating MSCNN generally performs better than using fixed-size convolutional kernels, so their choice of flexible kernels was effective.

Tom: Plus, the cross-attention mechanism itself is presented as a powerful tool for fusing global and local context in a way that suppresses redundant information and noise.

Jane: This dynamic focus is what allows the model to suppress noise while still capturing all the necessary contextual signals from both scales of data.

Lu: The theoretical analysis also confirms that you cannot get low generalization error without including both components; removing either part degrades performance.

Tom: So, they’re making a strong argument for this combined approach as the best way to reduce the overall prediction error across different time scales.

Conclusion: Jane: So to wrap things up on "User Trajectory Prediction Unifying Global and Local Temporal Information," we see that combining global and local information extraction is necessary for achieving low generalization error.

Tom: The main conclusion is that this MMCTP model outperforms others in multi-step trajectory prediction across various datasets with different sampling intervals.

Meng: It seems like the authors have successfully designed a system that balances prediction accuracy with an acceptable inference time, which is a practical win for deployment.

Lu: It pushes the idea forward by showing that this structure isn't just theoretical; it’s effective in practice for complex motion forecasting problems.

Jane: The implication is that for anyone working on these kinds of problems, looking at both large-scale trends and immediate local dynamics is crucial for making robust predictions.

Tom: So, we’re done with this paper. We’ve seen how the MMCTP model uses global MLP and local MSCNN fused by cross-attention to tackle multi-scale trajectory prediction challenges.

Lu: It’s a solid piece of research that shows how careful architectural design can effectively manage complex temporal information streams.

Meng: For practical application, it means we can expect better predictions in real-world scenarios where the environment and behavior are constantly changing.

Jane: We look forward to seeing what new models come along next that tackle these kinds of challenges with similar ideas or different angles.

Wei Hao, Bin Chong, Ronghua Ji, Chen Hou

cs.LG

Submitted: 2025-08-04

Updated: 2026-10-09

Comments: This manuscript has not passed peer review

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 78/100

The gist: The gist: A trajectory prediction model based on multilayer perceptron (MLP), multi-scale convolutional neural network (MSCNN), and cross-attention (CA) is proposed to reduce forecasting error by

Key concepts

Global Temporal Information Extraction: This part of the model uses an MLP with RevIN normalization and positional/temporal embeddings to capture overall long-term patterns in the trajectory data. It focuses on understanding the broad trends across the entire time sequence, helping to prevent errors caused by ignoring overall context.
Local Temporal Information Extraction: The MSCNN module models interactions within short, local temporal windows using convolutional kernels of various sizes. This allows the model to capture fine-grained behavioral patterns and immediate fluctuations in the trajectory data that occur over specific, small time intervals.
Cross-Attention (CA) Mechanism: This fusion module uses a cross-attention mechanism where local information acts as the query and global information acts as keys/values. This allows the model to intelligently weigh which global trends are most relevant for predicting specific local movements, creating a unified prediction.
Information Fusion: The core idea is merging the distinct temporal signals from both modules. By using cross-attention, the model learns how to optimally blend the high-level context (global) with the detailed, immediate dynamics (local) to produce a more accurate and robust final prediction.

Terminology

Adaptability — 1× in this paper · explained in 1 episode
Multilayer Perceptron (MLP) — 1× in this paper · explained in 2 episodes

Summary

The gist: A trajectory prediction model based on multilayer perceptron (MLP), multi-scale convolutional neural network (MSCNN), and cross-attention (CA) is proposed to reduce forecasting error by fusing global and local temporal information while maintaining acceptable inference time

Model Architecture

The proposed model, MMCTP, is designed around three main components: a global temporal information extraction module, a local temporal information extraction module, and an information fusion module. Specifically, the MLP is used to extract the global temporal information of each feature. In parallel, MSCNN is employed to extract the local temporal information by modeling interactions among features within a local temporal range.

Global Temporal Information Extraction

The MLP component is utilized to capture overall patterns, and it employs several specific techniques for input processing. The input data is first processed through RevIN normalization to mitigate the impact of data distribution shift. This process involves normalization along the temporal dimension for each sample and each variable, defined by the formula in equation (4). Furthermore, positional embedding injects spatial location information into the sequence using sine and cosine functions and temporal embedding represents time-related context. These embeddings are combined through element-wise addition before being fed into an MLP to extract global temporal information. The architecture also incorporates a skip connection in the MLP, which enhances the model’s understanding of multilevel information.

Local Temporal Information Extraction

The MSCNN module is responsible for modeling interactions among features within a local temporal range. To capture behavioral patterns at different temporal resolutions, convolutional kernels of diverse sizes are employed in MSCNN. The process involves concatenating the most recent segment of the input sequence with zero-padded positions representing future predictions before employing MSCNN. The dimensions are rearranged to facilitate convolutional kernels sliding along the temporal dimension for local temporal information extraction.

Information Fusion Module

The cross-attention (CA) mechanism is applied to fuse the global and local temporal information. In this module, the local temporal information matrix I is projected into a query matrix Q, while the global temporal information matrix S is projected into a key matrix K and a value matrix V. The attention mechanism calculates the similarity score by the scaled dot product of Q and K using a scaling factor of √1/dk to prevent gradient vanishing. Multiple attention heads are used, where each head focuses on distinct aspects of the input data. The outputs from all subspaces are concatenated and subsequently linearly transformed using a weight matrix WZ to unify the input and output dimensions.

Theoretical Analysis and Results

The theoretical analysis demonstrates that removing either the local or the global information extraction component results in degraded performance. Specifically, if only local information is extracted while global information is ignored, the model can capture short-term fluctuations but lacks awareness of the overall trend, making it more susceptible to perturbations and increasing variance. Conversely, if local information extraction is removed and only global long-term information is extracted, the model fails to capture local fluctuations, leading to insufficient fitting and increased bias. Experimental results show that MMCTP reduces mean squared error (MSE) by 5.04% and mean absolute error (MAE) by 4.35% compared with ModernTCN in 12-step prediction, while maintaining similar inference time. Ablation experiments confirm that removing the MLP module increases MSE by 1.04% and MAE by 0.43% on the dataset with a 15s sampling interval. Furthermore, incorporating MSCNN into the model mostly outperforms those obtained using fixed-size convolutional kernels. The overall time complexity of the model can be simplified to O(q).

Conclusion

The combination of global and local information extraction in the model is necessary to achieve a lower generalization error. The paper concludes that the proposed model outperforms others in multi-step trajectory prediction across datasets with different sampling intervals. Future work will focus on applying the MMCTP model to edge computing service migration and adaptively adjusting the kernel sizes in MSCNN to more flexibly capture trajectory dynamics across multiple temporal scales. The overall inference time trends observed on the three datasets are generally consistent. The model is designed to reduce prediction errors within an acceptable inference time. The results achieved by incorporating MSCNN into the model mostly outperform those obtained using fixed-size convolutional kernels. The overall effectiveness of incorporating MSCNN to capture these sparse local dynamics diminishes, resulting in limited performance improvement. The MMCTP’s inference time is similar to ModernTCN and is positioned between the values observed in Informer and TSMixer. The model predicts multiple steps simultaneously, with inference speed minimally affected by the prediction length. The visualization display of prediction results shows that the longitude and latitude remain relatively stable, while the altitude shows more pronounced variation. In contrast, Seq2Seq LSTM suffers from error accumulation in multi-step prediction because the prediction at the current time step depends on the result of the previous time step.

References

[1] C. Hou and Q. Zhao, “Optimal task-offloading control for edge computing system with tasks offloaded and computed in sequence,” IEEE Trans. Autom. Sci. Eng., vol. 20, no. 2, pp. 1378–1392, 2022

[4] Z. Zhao et al, “Reinforced-LSTM trajectory prediction-driven dynamic service migration: A case study,” IEEE Trans. Netw. Sci. Eng., vol. 9, no. 4, pp. 2786–2802, 2022

[5] L. Huang, Y. Ma, S. Wang, and Y. Liu, “An attention-based spatiotemporal LSTM network for next POI recommendation,” IEEE Trans. Serv. Comput., vol. 14, no. 6, pp. 1585–1597, 2019

[8] Y. Zheng, L. Zhang, X. Xie, and W.-Y. Ma, “Mining interesting locations and travel sequences from GPS trajectories,” in Proc. 18th Int. Conf. World Wide Web, 2009, pp. 791–800

[9] S. Zernetsch, S. Kohnen, M. Goldhammer, K. Doll, and B. Sick, “Trajectory prediction of cyclists using a physical model and an artificial neural network,” in Proc. 2016 IEEE Intell. Vehicles Symp., 2016, pp. 833–838

[10] C. Hou, “Activation control of multiple piecewise linear neural networks,” IEEE Trans. Autom. Sci. Eng., vol. 22, pp. 4895–4907, 2024

[11] L. Astuti, C.-H. Chiu, Y.-C. Lin, and M.-C. Lin, “Social-aware trajectory prediction using goal-directed attention networks with egocentric vision,” PeerJ Comput. Sci., vol.

Improvements for AI systems

This model can accurately predict user trajectory sequences by jointly leveraging global and local temporal information to reduce prediction errors within an acceptable inference time, as demonstrated by its ability to reduce mean squared error (MSE) by 5.04% and mean absolute error (MAE) by 4.35% compared with ModernTCN in 12-step prediction.
The system can handle diverse behavioral patterns across different time scales because the Multi-Scale Convolutional Neural Network (MSCNN) employs Convolutional kernels with different sizes to capture temporal information at multiple resolutions, enhancing the model’s adaptability to different behavioral patterns.
The prediction accuracy is enhanced by fusing global and local temporal information using a Cross-Attention (CA) mechanism, which allows the model to dynamically focus on global temporal patterns relevant to recent local temporal dynamics, while suppressing redundant information and noise.
The system can be optimized for resource-constrained edge computing scenarios because its time complexity is analyzed as O(q), showing that the overall asymptotic time complexity is similar to a single layer, which is relatively low.
The model's generalization error analysis confirms the necessity of both components by stating that removing either component results in degraded performance, concluding that the combination of global and local information extraction in the model is necessary to achieve a lower generalization error.