Abstract:Spatial-temporal prediction is a fundamental problem for constructing smart city, which is useful for tasks such as traffic control, taxi dispatching, and environmental policy making. Due to data collection mechanism, it is common to see data collection with unbalanced spatial distributions. For example, some cities may release taxi data for multiple years while others only release a few days of data; some regions may have constant water quality data monitored by sensors whereas some regions only have a small collection of water samples. In this paper, we tackle the problem of spatial-temporal prediction for the cities with only a short period of data collection. We aim to utilize the long-period data from other cities via transfer learning. Different from previous studies that transfer knowledge from one single source city to a target city, we are the first to leverage information from multiple cities to increase the stability of transfer. Specifically, our proposed model is designed as a spatial-temporal network with a meta-learning paradigm. The meta-learning paradigm learns a well-generalized initialization of the spatial-temporal network, which can be effectively adapted to target cities. In addition, a pattern-based spatial-temporal memory is designed to distill long-term temporal information (i.e., periodicity). We conduct extensive experiments on two tasks: traffic (taxi and bike) prediction and water quality prediction. The experiments demonstrate the effectiveness of our proposed model over several competitive baseline models.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: **How to improve the accuracy of spatio - temporal prediction by transferring knowledge from multiple source cities when the amount of data in the target city is limited**. Specifically, the paper focuses on how to use the long - term data of other cities to help those cities with only short - term data to make more accurate traffic flow and water quality predictions in the construction of smart cities. ### Problem Background In the process of building a smart city, spatio - temporal prediction is a fundamental problem and is very important for tasks such as traffic control, taxi dispatching, and environmental policy - making. However, due to different data collection mechanisms, the spatio - temporal data distributions in different cities are often unbalanced. For example, some cities may have released taxi data for many years, while other cities have only released data for a few days; some areas may have continuously monitored water quality data, while other areas have only a small amount of water sample data. ### Core Problem of the Paper The paper aims to solve the spatio - temporal prediction problem of those cities with only short - term data (i.e., target cities). To improve the prediction accuracy of these cities, the author proposes a new method to extract useful knowledge from other cities with long - term data (i.e., source cities) through transfer learning. Different from the previous methods that only transfer knowledge from a single source city, this paper attempts to extract information from multiple source cities for the first time to increase the stability and effectiveness of knowledge transfer. ### Overview of the Solution To solve the above problems, the author proposes a framework named MetaST, which combines the Spatial - Temporal Network (ST - net) and the Meta - Learning paradigm. Specifically: 1. **Spatial - Temporal Network (ST - net)**: It is used to capture spatio - temporal features and correlations and is composed of a local Convolutional Neural Network (CNN) and a Long - Short - Term Memory Network (LSTM). 2. **Meta - Learning Paradigm**: By using the initialization parameters learned from multiple source cities, it enables ST - net to quickly adapt to the data of the target city. 3. **Pattern - based Spatial - Temporal Memory (ST - mem)**: It is used to extract and store long - term spatio - temporal patterns (such as periodicity) and use these patterns in the target city to improve prediction. ### Experimental Verification The author conducted extensive experiments on two tasks: traffic prediction (taxi and bicycle flow prediction) and water quality prediction. The experimental results show that the MetaST framework performs well on multiple benchmark models, especially in the target cities with limited data, and its performance is better than other methods. ### Summary The main contributions of the paper include: - For the first time, it studies the problem of transferring knowledge from multiple source cities to improve the spatio - temporal prediction of the target city. - It proposes the MetaST framework, which combines the spatio - temporal network and the meta - learning paradigm and introduces a global spatio - temporal memory module. - It verifies the effectiveness of MetaST on multiple real - world spatio - temporal datasets. In this way, the paper provides an effective solution to help those cities with insufficient data to better perform spatio - temporal prediction, thereby supporting various application scenarios in the construction of smart cities.

Learning from Multiple Cities: A Meta-Learning Approach for Spatial-Temporal Prediction

Network-scale traffic prediction via knowledge transfer and regional MFD analysis

STG-Meta: Spatial-Temporal Graph Meta-Learning for Traffic Forecasting.

Meta-STMF: Meta-Learning Based Spatial Temporal Prediction Model Fusion Approach

Cross-City Transfer Learning for Deep Spatio-Temporal Prediction

Air Quality Prediction Model Based on Spatiotemporal Data Analysis and Metalearning

Meta-Transfer Learning Empowered Temporal Graph Networks for Cross-City Real Estate Appraisal

Traffic Prediction with Transfer Learning: A Mutual Information-based Approach

SSMT: Few-Shot Traffic Forecasting with Single Source Meta-Transfer

Transfer Learning for Urban Computing : A Case Study for Optimal Retail Store Placement

One Size Fits All: A Unified Traffic Predictor for Capturing the Essential Spatial–Temporal Dependency

Fuzzy Hybrid Framework with Dynamic Weights for Short‐term Traffic Flow Prediction by Mining Spatio‐temporal Correlations

STDA-Meta: A Meta-Learning Framework for Few-Shot Traffic Prediction

Meta-Learning-Based Spatial-Temporal Adaption for Coldstart Air Pollution Prediction

Meta-Learning over Time for Destination Prediction Tasks

Exploring the Generalizability of Spatio-Temporal Traffic Prediction: Meta-Modeling and an Analytic Framework

MetaStore: A Task-adaptative Meta-learning Model for Optimal Store Placement with Multi-city Knowledge Transfer

Semantic-Fused Multi-Granularity Cross-City Traffic Prediction

Deep Learning of Spatiotemporal Patterns for Urban Mobility Prediction Using Big Data

CityTrans: Domain-Adversarial Training with Knowledge Transfer for Spatio-Temporal Prediction Across Cities

AutoSTL: Automated Spatio-Temporal Multi-Task Learning.