This paper addresses the challenge of out-of-distribution (OOD) generalization in graph machine learning, a field rapidly advancing yet grappling with the discrepancy between source and target data distributions. Traditional graph learning algorithms, based on the assumption of uniform distribution between training and test data, falter in real-world scenarios where this assumption fails, resulting in suboptimal performance. A principal factor contributing to this suboptimal performance is the inherent simplicity bias of neural networks trained through Stochastic Gradient Descent (SGD), which prefer simpler features over more complex yet equally or more predictive ones. This bias leads to a reliance on spurious correlations, adversely affecting OOD performance in various tasks such as image recognition, natural language understanding, and graph classification. Current methodologies, including subgraph-mixup and information bottleneck approaches, have achieved partial success but struggle to overcome simplicity bias, often reinforcing spurious correlations. To tackle this, our study introduces a new learning paradigm for graph OOD issue. We propose DIVE, training a collection of models to focus on all label-predictive subgraphs by encouraging the models to foster divergence on the subgraph mask, which circumvents the limitation of a model solely focusing on the subgraph corresponding to simple structural patterns. Specifically, we employs a regularizer to punish overlap in extracted subgraphs across models, thereby encouraging different models to concentrate on distinct structural patterns. Model selection for robust OOD performance is achieved through validation accuracy. Tested across four datasets from GOOD benchmark and one dataset from DrugOOD benchmark, our approach demonstrates significant improvement over existing methods, effectively addressing the simplicity bias and enhancing generalization in graph machine learning.
论文检索
输入标题、作者或关键词，从 8,216 篇学术成果中精准定位
The recent advancements in Traffic Signal Control (TSC) have highlighted the potential of Reinforcement Learning (RL) as a promising solution to alleviate traffic congestion. Current research in this area primarily concentrates on either online or offline learning strategies, aiming to create optimized policies for specific cities. Nevertheless, the transferability of these policies to new cities is impeded by constraints such as the limited availability of high-quality data and the expensive and risky exploration process. To this end, in this paper, we present an innovative cross-city Traffic Signal Control (TSC) paradigm called CrossLight. Our approach involves meta training using offline data from source cities and adaptively fine-tuning in the target city. This novel methodology aims to address the challenges of transferring TSC policies across different cities effectively. In our proposed approach, we start by acquiring meta-decision pattern knowledge through trajectory dynamics reconstruction via pre-training in source cities. To address disparities in road network topologies between cities, we dynamically construct city topological structures based on the extracted meta-knowledge during the offline meta-training phase. These structures are then used to distill pattern-structure aware representations of decision trajectories from the source cities. To identify effective initial parameters for the learnable components, we employ the Model-Agnostic Meta-Learning (MAML) framework, a popular meta-learning approach. During adaptive fine-tuning in the target city, we introduce a replay buffer that is iteratively updated using online interactions with a rank and filter mechanism. This mechanism, along with a carefully designed exploration strategy, ensures a balance between exploitation and exploration, thereby fostering both the diversity and quality of the trajectories for fine-tuning. Finally, extensive experiments across four cities validate that CrossLight achieves comparable performance in new cities with minimal fine-tuning iterations, surpassing both existing online and offline methods. This success underscores that our CrossLight framework emerges as a groundbreaking and potent paradigm, offering a feasible and effective solution to the intelligent transportation community.
Pre-trained vision-language models like CLIP have shown powerful zero-shot inference ability via image-text matching and prove to be strong few-shot learners in various downstream tasks. However, in real-world scenarios, adapting CLIP to downstream tasks may encounter the following challenges: 1) data may exhibit long-tailed data distributions and might not have abundant samples for all the classes; 2) There might be emerging tasks with new classes that contain no samples at all. To overcome them, we propose a novel framework to achieve efficient and long-tailed generalization, which can be termed as Candle. During the training process, we propose compensating logit-adjusted loss to encourage large margins of prototypes and alleviate imbalance both within the base classes and between the base and new classes. For efficient adaptation, we treat the CLIP model as a black box and leverage the extracted features to obtain visual and textual prototypes for prediction. To make full use of multi-modal information, we also propose cross-modal attention to enrich the features from both modalities. For effective generalization, we introduce virtual prototypes for new classes to make up for their lack of training images. Candle achieves state-of-the-art performance over extensive experiments on 11 diverse datasets while substantially reducing the training time, demonstrating the superiority of our approach. The source code is available at https://github.com/shijxcs/Candle.
Despite the recent progress of molecular representation learning, its effectiveness is assumed on the close-world assumptions that training and testing graphs are from identical distribution. The open-world test dataset is often mixed with out-of-distribution (OOD) samples, where the deployed models will struggle to make accurate predictions. The misleading estimations of molecules' properties in drug screening or design can result in the tremendous waste of wet-lab resources and delay the discovery of novel therapies. Traditional detection methods need to trade off OOD detection and in-distribution (ID) classification performance since they share the same representation learning model. In this work, we propose to detect OOD molecules by adopting an auxiliary diffusion model-based framework, which compares similarities between input molecules and reconstructed graphs. Due to the generative bias towards reconstructing ID training samples, the similarity scores of OOD molecules will be much lower to facilitate detection. Although it is conceptually simple, extending this vanilla framework to practical detection applications is still limited by two significant challenges. First, the popular similarity metrics based on Euclidian distance fail to consider the complex graph structure. Second, the generative model involving iterative denoising steps is notoriously time-consuming especially when it runs on the enormous pool of drugs. To address these challenges, our research pioneers an approach of Prototypical Graph Reconstruction for Molecular OOd Detection, dubbed as PGR-MOOD. Specifically, PGR-MOOD hinges on three innovations: i) An effective metric to comprehensively quantify the matching degree of input and reconstructed molecules according to their discrete edges and continuous node features; ii) A creative graph generator to construct a list of prototypical graphs that are in line with ID distribution but away from OOD one; iii) An efficient and scalable OOD detector to compare the similarity between test samples and pre-constructed prototypical graphs and omit the generative process on every new molecule. Extensive experiments on ten benchmark datasets and six baselines are conducted to demonstrate our superiority: PGR-MOOD achieves more than 8% of average improvement in terms of detection AUC and AUPR accompanied by the reduced cost of testing time and memory consumption. The anonymous code is in: https://github.com/se7esx/PGR-MOOD.
The Combined Algorithm Selection and Hyperparameter Optimization (CASH) problem is pivotal in Automatic Machine Learning (AutoML). Most leading approaches combine Bayesian optimization with post-hoc ensemble building to create advanced AutoML systems. Bayesian optimization (BO) typically focuses on identifying a singular algorithm and its hyperparameters that outperform all other configurations. Recent developments have highlighted an oversight in prior CASH methods: the lack of consideration for diversity among the base learners of the ensemble. This oversight was overcome by explicitly injecting the search for diversity into the traditional CASH problem. However, despite recent developments, BO's limitation lies in its inability to directly optimize ensemble generalization error, offering no theoretical assurance that increased diversity correlates with enhanced ensemble performance. Our research addresses this gap by establishing a theoretical foundation that integrates diversity into the core of BO for direct ensemble learning. We explore a theoretically sound framework that describes the relationship between pair-wise diversity and ensemble performance, which allows our Bayesian optimization framework Optimal Diversity Bayesian Optimization (OptDivBO) to directly and efficiently minimize ensemble generalization error. OptDivBO guarantees an optimal balance between pairwise diversity and individual model performance, setting a new precedent in ensemble learning within CASH. Empirical results on 20 public datasets show that OptDivBO achieves the best average test ranks of 1.57 and 1.4 in classification and regression tasks.
BTTackler: A Diagnosis-based Framework for Efficient Deep Learning Hyperparameter Optimization
PDF ↗Hyperparameter optimization (HPO) is known to be costly in deep learning, especially when leveraging automated approaches. Most of the existing automated HPO methods are accuracy-based, i.e., accuracy metrics are used to guide the trials of different hyperparameter configurations amongst a specific search space. However, many trials may encounter severe training problems, such as vanishing gradients and insufficient convergence, which can hardly be reflected by accuracy metrics in the early stages of the training and often result in poor performance. This leads to an inefficient optimization trajectory because the bad trials occupy considerable computation resources and reduce the probability of finding excellent hyperparameter configurations within a time limitation. In this paper, we propose Bad Trial Tackler (BTTackler), a novel HPO framework that introduces training diagnosis to identify training problems automatically and hence tackles bad trials. BTTackler diagnoses each trial by calculating a set of carefully designed quantified indicators and triggers early termination if any training problems are detected. Evaluations are performed on representative HPO tasks consisting of three classical deep neural networks (DNN) and four widely used HPO methods. To better quantify the effectiveness of an automated HPO method, we propose two new measurements based on accuracy and time consumption. Results show the advantage of BTTackler on two-fold: (1) it reduces 40.33% of time consumption to achieve the same accuracy comparable to baseline methods on average and (2) it conducts 44.5% more top-10 trials than baseline methods on average within a given time budget. We also released an open-source Python library that allows users to easily apply BTTackler to automated HPO processes with minimal code changes\footnotehttps://github.com/thuml/BTTackler.
A point pattern is a dataset of coordinates, typically in 2D or 3D space. Point patterns are ubiquitous in diverse applications including Geographic Information Systems, Astronomy, Ecology, Biology and Medicine. Among the statistics used to quantify point patterns, most are based on Ripley's K-function, which measures the deviation of the observed pattern from a completely random arrangement of points. This approach is useful for constructing null hypothesis tests, but Ripley's K and its variants are less suitable as quantitative effect sizes because their ranges and expected values generally depend on the scale or the size of the region in which the pattern is observed. To address this, we propose a new function that behaves like a correlation coefficient for point patterns: it is tightly bounded by -1 and 1, with a value of -1 corresponding to a maximally dispersed arrangement of points, 0 indicating complete spatial randomness, and 1 representing maximal clustering. These properties are independent of scale and observation window size assuming appropriate edge correction. Evaluating our function on simulated data, we show that it has comparable statistical calibration and power to K-based baselines. We hope that the ease of interpretation of our bounded function will facilitate the analysis of spatial data across multiple fields.
The rapid expansion of Location-Based Social Networks (LBSNs) has highlighted the importance of effective next Point-of-Interest (POI) recommendations, which leverage historical check-in data to predict users' next POIs to visit. Traditional centralized deep neural networks (DNNs) offer impressive POI recommendation performance but face challenges due to privacy concerns and limited timeliness. In response, on-device POI recommendations have been introduced, utilizing federated learning (FL) and decentralized approaches to ensure privacy and recommendation timeliness. However, these methods often suffer from computational strain on devices and struggle to adapt to new users and regions. This paper introduces a novel collaborative learning framework, Diffusion-Based Cloud-Edge-Device Collaborative Learning for Next POI Recommendations (DCPR), leveraging the diffusion model known for its success across various domains. DCPR operates with a cloud-edge-device architecture to offer region-specific and highly personalized POI recommendations while reducing on-device computational burdens. DCPR minimizes on-device computational demands through a unique blend of global and local learning processes. Our evaluation with two real-world datasets demonstrates DCPR's superior performance in recommendation accuracy, efficiency, and adaptability to new users and regions, marking a significant step forward in on-device POI recommendation technology.
CAD assembly modeling, which refers to using CAD software to design new products from a catalog of existing machine components, is important in the industrial field. The graph neural network (GNN) based recommender system for CAD assembly modeling can help designers make decisions and speed up the design process by recommending the next required component based on the existing components in CAD software. These components can be represented as a graph naturally. However, present recommender systems for CAD assembly modeling adopt fixed GNN architectures, which may be sub-optimal for different manufacturers with different data distribution. Therefore, to customize a well-suited recommender system for different manufacturers, we propose a novel neural architecture search (NAS) framework, dubbed CusGNN, which can design data-specific GNN automatically. Specifically, we design a search space from three dimensions (i.e., aggregation, fusion, and readout functions), which contains a wide variety of GNN architectures. Then, we develop an effective differentiable search algorithm to search high-performing GNN from the search space. Experimental results show that the customized GNNs achieve 1.5-5.1% higher top-10 accuracy compared to previous manual designed methods, demonstrating the superiority of the proposed approach. Code and data are available at https://github.com/BUPT-GAMMA/CusGNN.
With the development of foundation models such as large language models, zero-shot transfer learning has become increasingly significant. This is highlighted by the generative capabilities of NLP models like GPT-4, and the retrieval-based approaches of CV models like CLIP, both of which effectively bridge the gap between seen and unseen data. In the realm of graph learning, the continuous emergence of new graphs and the challenges of human labeling also amplify the necessity for zero-shot transfer learning, driving the exploration of approaches that can generalize across diverse graph data without necessitating dataset-specific and label-specific fine-tuning. In this study, we extend such paradigms to Zero-shot transferability in Graphs by introducing ZeroG, a new framework tailored to enable cross-dataset generalization. Addressing the inherent challenges such as feature misalignment, mismatched label spaces, and negative transfer, we leverage a language model to encode both node attributes and class semantics, ensuring consistent feature dimensions across datasets. We also propose a prompt-based subgraph sampling module that enriches the semantic information and structure information of extracted subgraphs using prompting nodes and neighborhood aggregation, respectively. We further adopt a lightweight fine-tuning strategy that reduces the risk of overfitting and maintains the zero-shot learning efficacy of the language model. The results underscore the effectiveness of our model in achieving significant cross-dataset zero-shot transferability, opening pathways for the development of graph foundation models.
Educational recommendation seeks to suggest knowledge concepts that match a learner's ability, thus facilitating a personalized learning experience. In recent years, reinforcement learning (RL) methods have achieved considerable results by taking the encoding of the learner's exercise log as the state and employing an RL-based agent to make suitable recommendations. However, these approaches suffer from handling the diverse and dynamic learner's knowledge states. In this paper, we introduce the privileged feature distillation technique and propose the P rivileged K nowledge S tate D istillation (PKSD ) framework, allowing the RL agent to leverage the "actual'' knowledge state as privileged information in the state encoding to help tailor recommendations to meet individual needs. Concretely, our PKSD takes the privileged knowledge states together with the representations of the exercise log for the state representations during training. And through distillation, we transfer the ability to adapt to learners to aknowledge state adapter. During inference, theknowledge state adapter would serve as the estimated privileged knowledge states instead of the real one since it is not accessible. Considering that there are strong connections among the knowledge concepts in education, we further propose to collaborate the graph structure learning for concepts into our PKSD framework. This new approach is termed GEPKSD (Graph-Enhanced PKSD). As our method is model-agnostic, we evaluate PKSD and GEPKSD by integrating them with five different RL bases on four public simulators, respectively. Our results verify that PKSD can consistently improve the recommendation performance with various RL methods, and our GEPKSD could further enhance the effectiveness of PKSD in all the simulations.
Multitask learning is a widely used paradigm for training models on diverse tasks, with applications ranging from graph neural networks to language model fine-tuning. Since tasks may interfere with each other, a key notion for modeling their relationships is task affinity. This includes pairwise task affinity, computed among pairs of tasks, and higher-order affinity, computed among subsets of tasks. Naively computing either of them requires repeatedly training on data pooled from various task combinations, which is computationally intensive. We present a new algorithm Grad-TAG that can estimate task affinities without this repeated training. The key idea of Grad-TAG is to train a "base" model for all tasks and then use a linearization technique to estimate the loss of any other model with a specific task combination. The linearization works by computing a gradient-based first-order approximation of the loss, using low-dimensional projections of gradients as features in a logistic regression trained to predict labels for the specific task combination. We show theoretically that the linearized model can provably approximate the loss when the gradient-based approximation is accurate, and also empirically verify that on several large models. Then, given the estimated task affinity matrix, we design a semi-definite program for clustering to group similar tasks that maximize the average density of clusters. We evaluate Grad-TAG's performance across seven datasets, including multi-label classification on graphs, and instruction fine-tuning of language models. Our results show that our task affinity estimates are within 2.7% distance of the true affinities while needing only 3% of FLOPs compared to full training. On our largest graph with 21M edges and 500 labeling tasks, our algorithm delivers an estimate accurate to within 5% of the true affinities, while using only 112.3 GPU hours. Our results show that Grad-TAG achieves excellent performance and runtime tradeoffs compared to existing approaches.
SLADE: Detecting Dynamic Anomalies in Edge Streams without Labels via Self-Supervised Learning
PDF ↗To detect anomalies in real-world graphs, such as social, email, and financial networks, various approaches have been developed. While they typically assume static input graphs, most real-world graphs grow over time, naturally represented as edge streams. In this context, we aim to achieve three goals: (a) instantly detecting anomalies as they occur, (b) adapting to dynamically changing states, and (c) handling the scarcity of dynamic anomaly labels. In this paper, we propose SLADE (Self-supervised Learning for Anomaly Detection in Edge Streams) for rapid detection of dynamic anomalies in edge streams, without relying on labels. SLADE detects the shifts of nodes into abnormal states by observing deviations in their interaction patterns over time. To this end, it trains a deep neural network to perform two self-supervised tasks: (a) minimizing drift in node representations and (b) generating long-term interaction patterns from short-term ones. Failure in these tasks for a node signals its deviation from the norm. Notably, the neural network and tasks are carefully designed so that all required operations can be performed in constant time (w.r.t. the graph size) in response to each new edge in the input stream. In dynamic anomaly detection across four real-world datasets, SLADE outperforms nine competing methods, even those leveraging label supervision. Our code and datasets are available at https://github.com/jhsk777/SLADE.
Knowledge distillation (KD) has emerged as a promising technique for addressing the computational challenges associated with deploying large-scale recommender systems. KD transfers the knowledge of a massive teacher system to a compact student model, to reduce the huge computational burdens for inference while retaining high accuracy. The existing KD studies primarily focus on one-time distillation in static environments, leaving a substantial gap in their applicability to real-world scenarios dealing with continuously incoming users, items, and their interactions. In this work, we delve into a systematic approach to operating the teacher-student KD in a non-stationary data stream. Our goal is to enable efficient deployment through a compact student, which preserves the high performance of the massive teacher, while effectively adapting to continuously incoming data. We propose Continual Collaborative Distillation (CCD) framework, where both the teacher and the student continually and collaboratively evolve along the data stream. CCD facilitates the student in effectively adapting to new data, while also enabling the teacher to fully leverage accumulated knowledge. We validate the effectiveness of CCD through extensive quantitative, ablative, and exploratory experiments on two real-world datasets. We expect this research direction to contribute to narrowing the gap between existing KD studies and practical applications, thereby enhancing the applicability of KD in real-world systems.
Gandalf: Learning Label-label Correlations in Extreme Multi-label Classification via Label Features
PDF ↗Extreme Multi-label Text Classification (XMC) involves learning a classifier that can assign an input with a subset of most relevant labels from millions of label choices. Recent works in this domain have increasingly focused on a symmetric problem setting where both input instances and label features are short-text in nature. Short-text XMC with label features has found numerous applications in areas such as query-to-ad-phrase matching in search ads, title-based product recommendation, prediction of related searches. In this paper, we propose Gandalf, a novel approach which makes use of a label co-occurrence graph to leverage label features as additional data points to supplement the training distribution. By exploiting the characteristics of the short-text XMC problem, it leverages the label features to construct valid training instances, and uses the label graph for generating the corresponding soft-label targets, hence effectively capturing the label-label correlations. Surprisingly, models trained on these new training instances, although being less than half of the original dataset, can outperform models trained on the original dataset, particularly on the PSP@k metric for tail labels. With this insight, we aim to train existing XMC algorithms on both, the original and new training instances, leading to an average 5% relative improvements for 6 state-of-the-art algorithms across 4 benchmark datasets consisting of up to 1.3M labels. Gandalf can be applied in a plug-and-play manner to various methods and thus forwards the state-of-the-art in the domain, without incurring any additional computational overheads. Code has been open-sourced at www.github.com/xmc-aalto/InceptionXML.
Addressing Prediction Delays in Time Series Forecasting: A Continuous GRU Approach with Derivative Regularization
PDF ↗Time series forecasting has been an essential field in many different application areas, including economic analysis, meteorology, and so forth. The majority of time series forecasting models are trained using the mean squared error (MSE). However, this training based on MSE causes a limitation known as prediction delay. The prediction delay, which implies the ground-truth precedes the prediction, can cause serious problems in a variety of fields, e.g., finance and weather forecasting --- as a matter of fact, predictions succeeding ground-truth observations are not practically meaningful although their MSEs can be low. This paper proposes a new perspective on traditional time series forecasting tasks and introduces a new solution to mitigate the prediction delay. We introduce a continuous-time gated recurrent unit (GRU) based on the neural ordinary differential equation (NODE) which can supervise explicit time-derivatives. We generalize the GRU architecture in a continuous-time manner and minimize the prediction delay through our time-derivative regularization. Our method outperforms in metrics such as MSE, Dynamic Time Warping (DTW) and Time Distortion Index (TDI). In addition, we demonstrate the low prediction delay of our method in a variety of datasets.
Approaches to recommendation are typically evaluated in one of two ways: (1) via a (simulated) online experiment, often seen as the gold standard, or (2) via some offline evaluation procedure, where the goal is to approximate the outcome of an online experiment. Several offline evaluation metrics have been adopted in the literature, inspired by ranking metrics prevalent in the field of Information Retrieval. (Normalised) Discounted Cumulative Gain (nDCG) is one such metric that has seen widespread adoption in empirical studies, and higher (n)DCG values have been used to present new methods as the state-of-the-art in top-n recommendation for many years. Our work takes a critical look at this approach, and investigates when we can expect such metrics to approximate the gold standard outcome of an online experiment. We formally present the assumptions that are necessary to consider DCG an unbiased estimator of online reward and provide a derivation for this metric from first principles, highlighting where we deviate from its traditional uses in IR. Importantly, we show that normalising the metric renders it inconsistent, in that even when DCG is unbiased, ranking competing methods by their normalised DCG can invert their relative order. Through a correlation analysis between off- and on-line experiments conducted on a large-scale recommendation platform, we show that our unbiased DCG estimates strongly correlate with online reward, even when some of the metric's inherent assumptions are violated. This statement no longer holds for its normalised variant, suggesting that nDCG's practical utility may be limited.
We study the stochastic Budgeted Multi-Armed Bandit (MAB) problem, where a player chooses from K arms with unknown expected rewards and costs. The goal is to maximize the total reward under a budget constraint. A player thus seeks to choose the arm with the highest reward-cost ratio as often as possible. Current approaches for this problem have several issues, which we illustrate. To overcome them, we propose a new upper confidence bound (UCB) sampling policy, ømega-UCB, that uses asymmetric confidence intervals. These intervals scale with the distance between the sample mean and the bounds of a random variable, yielding a more accurate and tight estimation of the reward-cost ratio compared to our competitors. We show that our approach has sublinear instance-dependent regret in general and logarithmic regret for parameter ρ ≥ 1, and that it outperforms existing policies consistently in synthetic and real settings.
The burgeoning volume of graph data presents significant computational challenges in training graph neural networks (GNNs), critically impeding their efficiency in various applications. To tackle this challenge, graph condensation (GC) has emerged as a promising acceleration solution, focusing on the synthesis of a compact yet representative graph for efficiently training GNNs while retaining performance. Despite the potential to promote scalable use of GNNs, existing GC methods are limited to aligning the condensed graph with merely the observed static graph distribution. This limitation significantly restricts the generalization capacity of condensed graphs, particularly in adapting to dynamic distribution changes. In real-world scenarios, however, graphs are dynamic and constantly evolving, with new nodes and edges being continually integrated. Consequently, due to the limited generalization capacity of condensed graphs, applications that employ GC for efficient GNN training end up with sub-optimal GNNs when confronted with evolving graph structures and distributions in dynamic real-world situations. To overcome this issue, we propose open-world graph condensation (OpenGC), a robust GC framework that integrates structure-aware distribution shift to simulate evolving graph patterns and exploit the temporal environments for invariance condensation. This approach is designed to extract temporal invariant patterns from the original graph, thereby enhancing the generalization capabilities of the condensed graph and, subsequently, the GNNs trained on it. Furthermore, to support the periodic re-condensation and expedite condensed graph updating in life-long graph learning, OpenGC reconstructs the sophisticated optimization scheme with kernel ridge regression and non-parametric graph convolution, significantly accelerating the condensation process while ensuring the exact solutions. Extensive experiments on both real-world and synthetic evolving graphs demonstrate that OpenGC outperforms state-of-the-art (SOTA) GC methods in adapting to dynamic changes in open-world graph environments.
Mobile traffic prediction plays a crucial role in enabling efficient network management and service provisioning. Traditional prediction approaches treat different mobile application services (such as Uber, Facebook, Twitter, etc) as isolated entities, neglecting potential correlation among them. Moreover, such isolated prediction methods necessitate the uploading of historical traffic data from all regions to forecast city-wide traffic, resulting in consuming substantial bandwidth resources and risking prediction failure in the event of data loss in specific regions. To address these challenges, we propose a novel Cross-service Attention-based Spatial-Temporal Graph Convolutional Network (CsASTGCN) for precise and communication-efficient multi-service mobile traffic prediction. Our methodology allows each mobile service to transmit the traffic data of only a fraction of regions for city-wide traffic prediction of all mobile services, which reduces the resource consumption caused by data transmission. Specifically, the sparse traffic data are initially transmitted to the cloud server and the masked graph autoencoder is utilized to roughly reconstruct the traffic volume for regions with missing data. Subsequently, a cross-service attention-based predictor is designed to calculate the data correlation among different mobile services within the same region. Considering the constantly emerging mobile services, we incorporate a novel model-based adaptive transfer learning scheme to extract valuable knowledge from the existing models and expedite the training of a new model for a new service without training from scratch, thereby enhancing the scalability of our framework. Extensive experiments conducted on a large-scale real-world mobile traffic dataset demonstrate that our model greatly outperforms the existing schemes, enhancing both the communication-efficiency and robustness of large-scale multi-service traffic prediction.