论文检索

Jiahao Ji, Tianyu Wang 0028, Yeshu Li, Yusen Huo, Zhilin Zhang 0003, Chuan Yu 0002, Jian Xu 0015, Bo Zheng 0007

Auto-bidding is crucial in facilitating online advertising by automatically providing bids for advertisers. While previous work has made great efforts to model bidding environments for better ad performance, it has limitations in generalizability across environments since these models are typically tailored for specific bidding scenarios. To this end, we approach the scenario-independent principles through a unified function that estimates the achieved effect under specific bids, such as budget consumption, gross merchandise volume (GMV), page views, etc. Then, we propose a bidding foundation model Bid2X to learn this fundamental function from data in various scenarios. Our Bid2X is built over uniform series embeddings that encode heterogeneous data through tailored embedding methods. To capture complex inter-variable and dynamic temporal dependencies in bidding data, we propose two attention mechanisms separately treating embeddings of different variables and embeddings at different times as attention tokens for representation learning. On top of the learned variable and temporal representations, a variable-aware fusion module is used to perform adaptive bidding outcome prediction. To model the unique bidding data distribution, we devise a zero-inflated projection module to incorporate the estimated non-zero probability into its value prediction, which makes up a joint optimization objective containing classification and regression. The objective is proven to converge to the zero-inflated distribution. Our model has been deployed on the ad platform in Taobao, one of the world's largest e-commerce platforms. Offline evaluation on eight large-scale real-world datasets exhibits Bid2X's superiority compared to various baselines and its generality across different scenarios. In real-world applications, Bid2X increased GMV by 4.65% and ROI by 2.44% in online A/B tests, paving the way for the bidding foundation model in computational advertising.

Jiawei Guo, Yang Xiao 0014, Weipeng Huang, Guangyuan Piao

A hierarchical labeling system for mobile applications (apps) benefits a wide range of downstream businesses that integrate the labeling with their proprietary user data, to improve user modeling. Such a label hierarchy can define more granular labels that capture detailed app features beyond the limitations of traditional broad app categories. In this paper, we address the problem of hierarchical multilabel classification for apps by using their textual information such as names and descriptions. We present: 1) HMCN (Hierarchical Multilabel Classification Network) for handling the classification from two perspectives: the first focuses on a multilabel classification without hierarchical constraints, while the second predicts labels sequentially at each hierarchical level considering such constraints; 2) HMCL (Hierarchical Multilabel Contrastive Learning), a scheme that is capable of learning more distinguishable app representations to enhance the performance of HMCN. Empirical results on our Tencent App Store dataset and two public datasets demonstrate that our approach performs well compared with state-of-the-art methods. The approach has been deployed at Tencent and the multilabel classification outputs for apps have helped a downstream task--credit risk management of users--improve its performance by 10.70% with regard to the Kolmogorov-Smirnov metric, for over one year.

Xinyue Feng, Shuxin Zhong, Jinquan Hang, Wenjun Lyu, Yuequn Zhang, Guang Yang 0028, Haotian Wang 0008, Desheng Zhang 0002, Guang Wang 0001

Customer expansion, i.e., growing a business's existing customer base by acquiring new customers, is critical for scaling operations and sustaining the long-term profitability of logistics companies. Although state-of-the-art works model this task as a single-node classification problem under a heterogeneous graph learning framework and achieve good performance, they struggle with extremely positive label sparsity issues in our scenario. Multi-task learning (MTL) offers a promising solution by introducing a correlated, label-rich task to enhance the label-sparse task prediction through knowledge sharing. However, existing MTL methods result in performance degradation because they fail to discriminate task-shared and task-specific structural patterns across tasks. This issue arises from their limited consideration of the inherently complex structure learning process of heterogeneous graph neural networks, which involves the multi-layer aggregation of multi-type relations. To address the challenge, we propose a Structure-Aware Hierarchical Information Sharing Framework (SrucHIS), which explicitly regulates structural information sharing across tasks in logistics customer expansion. SrucHIS breaks down the structure learning phase into multiple stages and introduces sharing mechanisms at each stage, effectively mitigating the influence of task-specific structural patterns during each stage. We evaluate StrucHIS on both private and public datasets, achieving a 51.41% average precision improvement on the private dataset and a 10.52% macro F1 gain on the public dataset. StrucHIS is further deployed at one of the largest logistics companies in China and demonstrates a 41.67% improvement in the success contract-signing rate over existing strategies, generating over 453K new orders within just two months.

Xin Dong, Sen Jia, Ming Rui Wang, Yan Li 0043, Zhenheng Yang, Bingfeng Deng, Hongyu Xiong

Recently, with the emergence of recent Multimodal Large Language Model (MLLM) technology, it has become possible to exploit its video understanding capability on different classification tasks. In practice, we face the difficulty of huge requirements for GPU resource if we need to deploy MLLMs online. In this paper, we propose COEF-VQ, a novel cascaded MLLM framework designed to enhance video quality understanding on the short-video platform while optimizing computational efficiency. Our approach integrates an entropy-based pre-filtering stage, where a lightweight model assesses uncertainty and selectively filters cases before passing them to the more computationally intensive MLLM for final evaluation. By prioritizing high-uncertainty samples for deeper analysis, our framework significantly reduces GPU usage while maintaining the strong classification performance of a full MLLM deployment. To demonstrate the effectiveness of COEF-VQ, we deploy this new framework onto the video management platform (VMP) at the short-video platform, and perform a series of detailed experiments on two in-house tasks related to video quality understanding. We show that COEF-VQ leads to substantial performance gains from the offline evaluation in these two tasks and effectively enhances platform safety with limit resource consumption, significantly reducing inappropriate content video view rate by 9.9% in a online A/B test without affecting engagement. Post-launch monitoring confirmed sustained improvements, validating its real-world impact.

Silin Zhou, Yao Chen 0008, Shuo Shang, Lisi Chen 0001, Bingsheng He, Ryosuke Shibasaki

Trajectory representation learning (TRL) maps trajectories to vector embeddings and facilitates tasks such as trajectory classification and similarity search. State-of-the-art (SOTA) TRL methods transform raw GPS trajectories to grid or road trajectories to capture high-level travel semantics, i.e., regions and roads. However, they lose fine-grained spatial-temporal details as multiple GPS points are grouped into a single grid cell or road segment. To tackle this problem, we propose the BLU rred Encoding method, dubbed BLUE, which gradually reduces the precision of GPS coordinates to create hierarchical patches with multiple levels. The low-level patches are small and preserve fine-grained spatial-temporal details, while the high-level patches are large and capture overall travel patterns. To complement different patch levels with each other, our BLUE is an encoder-decoder model with a pyramid structure. At each patch level, a Transformer is used to learn the trajectory embedding at the current level, while pooling prepares inputs for the higher level in the encoder, and up-resolution provides guidance for the lower level in the decoder. BLUE is trained using the trajectory reconstruction task with the MSE loss. We compare BLUE with 8 SOTA TRL methods for 3 downstream tasks, the results show that BLUE consistently achieves higher accuracy than all baselines, outperforming the best-performing baselines by an average of 30.90%. Our code is available at https://github.com/slzhou-xy/BLUE.

Shuhan Zhong, Weipeng Zhuo, Sizhe Song, Guanyao Li, Zhongyi Yu, S.-H. Gary Chan

Irregular multivariate time series (IMTS) is characterized by the lack of synchronized observations across its different channels. In this paper, we point out that this channel-wise asynchrony can lead to poor channel-wise modeling of existing deep learning methods. To overcome this limitation, we propose MTM, a multi-scale token mixing transformer for the classification of IMTS. We find that the channel-wise asynchrony can be alleviated by down-sampling the time series to coarser timescales, and propose to incorporate a masked concat pooling in MTM that gradually down-samples IMTS to enhance the channel-wise attention modules. Meanwhile, we propose a novel channel-wise token mixing mechanism which proactively chooses important tokens from one channel and mixes them with other channels, to further boost the channel-wise learning of our model. Through extensive experiments on real-world datasets and comparison with state-of-the-art methods, we demonstrate that MTM consistently achieves the best performance on all the benchmarks, with improvements of up to 3.8% in AUPRC for classification.

Ziyu Zheng, Yaming Yang 0002, Ziyu Guan, Wei Zhao 0019, Weigang Lu 0001

Masked Graph Auto-Encoder, a powerful graph self-supervised training paradigm, has recently shown superior performance in graph representation learning. Existing works typically rely on node contextual information to recover the masked information. However, they fail to generalize well to heterophilic graphs where connected nodes may be not similar, because they focus only on capturing the neighborhood information and ignoring the discrepancy information between different nodes, resulting in indistinguishable node representations. In this paper, to address this issue, we propose a Discrepancy-Aware Graph Mask Auto-Encoder (DGMAE). It obtains more distinguishable node representations by reconstructing the discrepancy information of neighboring nodes during the masking process. We conduct extensive experiments on 17 widely-used benchmark datasets. The results show that our DGMAE can effectively preserve the discrepancies of nodes in low-dimensional space. Moreover, DGMAE significantly outperforms state-of-the-art graph self-supervised learning methods on three graph analytic including tasks node classification, node clustering, and graph classification, demonstrating its remarkable superiority. The code of DGMAE is available at https://github.com/zhengziyu77/DGMAE.

Liangwei Nathan Zheng, Chang George Dong, Wei Emma Zhang, Lin Yue, Miao Xu 0001, Olaf Maennel, Weitong Chen 0001

Large Language Models (LLMs) have demonstrated impressive performance in time series analysis and seems to understand the time temporal relationship well than traditional transformer-based approaches. However, since LLMs are not designed for time series tasks, simpler models-like linear regressions can often achieve comparable performance with far less complexity. In this study, we perform extensive experiments to assess the effectiveness of applying LLMs to key time series tasks, including forecasting, classification, imputation, and anomaly detection. We compare the performance of LLMs against simpler baseline models, such as single-layer linear models and randomly initialized LLMs. Our results reveal that LLMs offer minimal advantages for these core time series tasks and may even distort the temporal structure of the data. In contrast, simpler models consistently outperform LLMs while requiring far fewer parameters. Furthermore, we analyze existing reprogramming techniques and show, through data manifold analysis, that these methods fail to effectively align time series data with language and display ''pseudo-alignment'' behavior in embedding space. Our findings suggest that the performance of LLM-based methods in time series tasks arises from the intrinsic characteristics and structure of time series data, rather than any meaningful alignment with the language model architecture. We release the code for experiments here: https://github.com/IcurasLW/Official-Repository_Understanding_LLM_for_Time_Series_Analysis.git

Zhiyuan Zhao 0002, Haoxin Liu 0001, Alexander Rodríguez, B. Aditya Prakash

Time-series forecasting is a critical challenge in various domains and has witnessed substantial progress in recent years. Many real-life scenarios, such as public health, economics, and social applications, involve feedback loops where predictive models can trigger actions that influence the outcome they aim to predict, subsequently altering the target variable's distribution. This phenomenon, known as performativity, introduces the potential for 'self-negating' or 'self-fulfilling' predictions. Despite extensive studies on performativity in classification problems across domains, this phenomenon remains largely unexplored in the context of time-series forecasting from a machine-learning perspective. In this paper, we formalize Performative Time-Series Forecasting (PeTS), addressing the challenge for predictions when performativity-induced distribution shifts are possible. We propose a novel solution to PeTS, Feature Performative-Shifting (FPS), which leverages the concept of delayed response to anticipate distribution shifts and subsequently predicts targets. We provide theoretical insights suggesting that FPS potentially leads to reduced generalization error. Extensive experiment results demonstrate that FPS consistently outperforms conventional time-series forecasting and concept drift methods, highlighting its efficacy in handling performativity-induced challenges.

Weijia Zhang 0003, Chenlong Yin, Hao Liu 0026, Hui Xiong 0001

Pre-trained Language Models (PLMs), such as ChatGPT, have significantly advanced the field of natural language processing. This progress has inspired a series of innovative studies that explore the adaptation of PLMs to time series analysis, intending to create a unified foundation model that addresses various time series analytical tasks. However, these efforts predominantly focus on Regularly Sampled Time Series (RSTS), neglecting the unique challenges posed by Irregularly Sampled Time Series (ISTS), which are characterized by uneven sampling intervals and prevalent missing data. To bridge this gap, this work takes the first step in exploring the potential of PLMs for ISTS analysis. We begin by investigating the effect of various methods for representing ISTS, aiming to maximize the efficacy of PLMs in the analysis. Furthermore, we propose a unified PLM-based framework, named ISTS-PLM, to address diverse ISTS analytical tasks. It integrates novel time-aware and variable-aware PLMs tailored to tackle the intractable intra- and inter-time series modeling in ISTS. Finally, extensive experiments on a comprehensive benchmark demonstrate that the ISTS-PLM, utilizing a structured and effective series-based representation for ISTS, consistently achieves state-of-the-art performance across various analytical tasks, such as classification, interpolation, extrapolation, few-shot and zero-shot learning scenarios, spanning scientific domains like healthcare, biomechanics, and climate science.

Shuo Zhang 0015, Jing Wang 0060, Shiqin Nie, Jinghang Yue, Weikang Zhu, Youfang Lin

Incomplete time series classification is both practically valuable and challenging as missing values in time series data are prevalent in real-world scenarios. Current approaches suffer from two major limitations. First, they overemphasize the consistency of data reconstruction during missing value imputation while neglecting the task-effectiveness of the imputed results for the classification. Second, they fail to systematically establish a synergistic optimization mechanism between data imputation and feature representation. To address these challenges, we propose a Hierarchical Conditional Information Bottleneck (HCIB) framework, which achieves incomplete time series classification through end-to-end joint optimization. Specifically, at the data imputation level, we re-examine the dual effects of missing data: the loss of critical information (Loss) versus the gain in interference suppression (Gain), elucidating this duality through bias-variance trade-off theory. Building on this analysis, we propose a task-information sufficiency criterion and extend the information bottleneck theory into a task-driven imputation framework by incorporating label information as a conditional constraint. At the feature representation level, we construct a hierarchical information bottleneck architecture to learn compressed yet informative temporal representations from the task-oriented imputed data. Furthermore, we derive the optimizable objective function for HCIB and design specialized neural network architectures for time series. Comprehensive experiments on multivariate and univariate time series datasets across multiple domains consistently demonstrate that the proposed method achieves significant improvements in classification performance compared to SOTA approaches.

Junru Zhang 0001, Lang Feng 0002, Xu Guo 0002, Han Yu 0001, Yabo Dong, Duanqing Xu

Single-domain generalization (SDG) in time series classification (TSC) poses significant challenges for current time-series domain generalization methods due to the extremely limited data available from only one source domain. In this study, we propose Segment-dErived Expansion of Domains (SEED), a diffusion-based method that effectively expands domain diversity for SDG. We reveal that individual instances exhibit intrinsic temporal shifts over time, which provides a principled foundation for creating multiple pseudo domains by segmenting each instance into distinct parts. To do so, SEED extracts two complementary representations from each time-series segment: 1) a segment-specific representation that captures diverse distributional variations, and 2) a segment-invariant representation that preserves class semantics. SEED formulates these representations as pseudo-domain prompts to guide a diffusion model in generating diverse yet semantically consistent time-series data. Additionally, SEED introduces a novel prompt-fused sampling for diffusion, enabling flexible recombination of segment-specific features to continuously expand the pseudo-domain space. We provide both theoretical analysis and extensive empirical evaluations on four widely used TSC benchmarks to validate its ability in reducing generalization error and improving model's performances in SDG. In our experiments, SEED significantly improves classification accuracy by 7.68% on average compared to the strong baselines.

Sukwon Yun, Xin Liu 0020, Yunhak Oh, Junseok Lee 0002, Tianlong Chen 0001, Tsuyoshi Murata, Chanyoung Park 0001

In real-world graphs, we often encounter missing feature situations where a few or the majority of node features, e.g., sensitive information, are missed. In such scenarios, directly utilizing Graph Neural Networks (GNNs) would yield sub-optimal results in downstream tasks such as node classification. Despite the emergence of a few GNN-based methods attempting to mitigate its missing situation, when only a few features are available, they rather perform worse than traditional structure-based models. To this end, we propose a novel framework that further illuminates the potential of classical Label Propagation (Oldie), taking advantage of Feature Propagation, especially when only a partial feature is available. Now called by GOODIE, it takes a hybrid approach to obtain embeddings from the Label Propagation branch and Feature Propagation branch. To do so, we first design a GNN-based decoder that enables the Label Propagation branch to output hidden embeddings that align with those of the FP branch. Then, GOODIE automatically captures the significance of structure and feature information thanks to the newly designed Structure-Feature Attention. Followed by a novel Pseudo-Label contrastive learning that differentiates the contribution of each positive pair within pseudo-labels originating from the LP branch, GOODIE outputs the final prediction for the unlabeled nodes. Through extensive experiments, we demonstrate that our proposed model, GOODIE, outperforms the existing state-of-the-art methods not only when only a few features are available but also in abundantly available situations. Source code of GOODIE is available at: https://github.com/SukwonYun/GOODIE.

Penghui Yang 0001, Chen-Chen Zong, Sheng-Jun Huang, Lei Feng 0006, Bo An 0001

Traditional knowledge distillation focuses on aligning the student's predicted probabilities with both ground-truth labels and the teacher's predicted probabilities. However, the transition to predicted probabilities from logits would obscure certain indispensable information. To address this issue, it is intuitive to additionally introduce a logit-level loss function as a supplement to the widely used probability-level loss function, for exploiting the latent information of logits. Unfortunately, we empirically find that the amalgamation of the newly introduced logit-level loss and the previous probability-level loss will lead to performance degeneration, even trailing behind the performance of employing either loss in isolation. We attribute this phenomenon to the collapse of the classification head, which is verified by our theoretical analysis based on the neural collapse theory. Specifically, the gradients of the two loss functions exhibit contradictions in the linear classifier yet display no such conflict within the backbone. Drawing from the theoretical analysis, we propose a novel method called dual-head knowledge distillation, which partitions the linear classifier into two classification heads responsible for different losses, thereby preserving the beneficial effects of both losses on the backbone while eliminating adverse influences on the classification head. Extensive experiments validate that our method can effectively exploit the information inside the logits and achieve superior performance against state-of-the-art counterparts

Kaixin Xu, Yangyang Wu, Shijun Wu, Xiaoye Miao, Guoqing Chao, Mengying Zhu, Meng Xi 0002, Xinkui Zhao

Incomplete multi-view partial multi-label classification (IMvPMLC), which tackles the combined challenges of incompleteness in both multi-view and multi-label problems, has drawn considerable attention. Existing IMvPMLC methods have made progress but still face several challenges: (i) They mainly focus on the consistency of representations across multiple views but overlook the relationships among instances, leading to suboptimal representations. (ii) They primarily utilize only the available labels for supervised learning, ignoring the missing label distribution and limiting their ability to capture label correlations. In this paper, we propose a novel model named Dual Structure-guided Contrastive Network (DSCN) for IMvPMLC. Specifically, we introduce a similarity-guided instance-level contrastive learning mechanism to achieve multi-view consistent and discriminative representations across instances by leveraging instance structures, while a multi-view attention-based fusion strategy dynamically facilitates the fusion of multi-view representations to derive a robust consensus representation. Then, we design a multi-view shared classifier integrated with a correlation-guided label-level contrastive learning mechanism to enhance predictions by leveraging complementary information across multiple views and capturing label structures, effectively exploiting missing label distribution. Extensive experiments on five benchmark datasets demonstrate that, DSCN yields a more than 13% accuracy, compared with the state-of-the-art approaches. The code and datasets are available at https://anonymous.4open.science/r/DSCN-D471.

Xuansheng Wu, Wenhao Yu 0002, Xiaoming Zhai, Ninghao Liu 0001

Modern text classification methods heavily rely on contextual embeddings from large language models (LLMs). Compared to human-engineered features, these embeddings provide automatic and effective representations for classification model training. However, they also introduce a challenge: we lose the ability to manually remove unintended features, such as sensitive or task-irrelevant features, to guarantee regulatory compliance or improve the generalizability of classification models. This limitation arises because LLM embeddings are opaque and difficult to interpret. In this paper, we propose a novel framework to identify and regularize unintended features in the LLM latent space. Specifically, we first pre-train a sparse autoencoder (SAE) to extract interpretable features from LLM latent spaces. To ensure the SAE can capture task-specific features, we further fine-tune it on task-specific datasets. In training the classification model, we propose a simple and effective regularizer, by minimizing the similarity between the classifier weights and the identified unintended feature, to remove the impact of these unintended features on classification. We evaluate the proposed framework on three real-world tasks, including toxic chat detection, reward modeling, and disease diagnosis. Results show that the proposed self-regularization framework can improve the classifier's generalizability by regularizing those features that are not semantically correlated to the task. This work pioneers controllable text classification on LLM latent spaces by leveraging interpreted features to address generalizability, fairness, and privacy challenges. The code and data are publicly available at https://github.com/JacksonWuxs/Controllable_LLM_Classifier.

Xuan Wu, Yizheng Zhao

Neuro-symbolic computing aims to integrate neural learning with symbolic reasoning to address the fundamental challenge of symbol grounding. While neural networks excel at pattern recognition, they struggle to maintain logical consistency. Conversely, symbolic systems provide formal reasoning capabilities but lack mechanisms for handling perceptual uncertainty. This paper introduces EmALC , a novel neuro-symbolic framework that bridges neural perception with symbolic logic through differentiable fuzzy semantics. Our approach addresses a key limitation of existing methods: while previous neuro-symbolic approaches like Logic Tensor Networks employ first-order fuzzy logic, where key reasoning problems are undecidable, EmALC ensures decidable reasoning by leveraging a fuzzy variant of ALC -- a decidable fragment of first-order logic. Unlike previous approaches that often compromise logical soundness for learning capability, EmALC maintains provable semantic consistency through a hierarchical loss function while mitigating reasoning shortcuts via rule-based revision strategies. Experimental evaluation demonstrates EmALC's effectiveness: on ontology revision tasks, it achieves 100% success rate in correcting masked groundings while preserving semantic integrity; on semantic image interpretation tasks, it improves object classification F1-scores by up to 5.56% through ontology-guided knowledge revision.

Xiangchao Wen, Zhen Liu 0006, Yuxin You

Graph adversarial attacks refer to a class of adversarial attack methods targeting graph-structured data, such as social networks, knowledge graphs, and molecular structures. The objective of these attacks is to introduce subtle yet carefully designed perturbations to graph data (e.g., nodes, edges, or features) to cause graph-based machine learning models, such as graph neural networks (GNNs), to produce incorrect predictions or classification results. The core idea behind graph adversarial attacks is to exploit the vulnerabilities of the model, maximizing the degradation of its performance with minimal perturbations. However, due to the incompleteness of attack strategies, graph adversarial attacks often exhibit redundancy, meaning that some perturbations are ineffective. These ineffective perturbations consume the attack budget and increase the likelihood of detection. To address this issue, this paper proposes Budget-Reduced Attack Filtering (BRAF in short) frameworks capable of eliminating ineffective perturbations produced by existing graph adversarial attack methods, thereby optimizing the attack budget. Extensive experiments validate the effectiveness and scalability of the proposed filtering frameworks. The code is available at https://github.com/Xiangchao-Wen/Filter-Graph-Attack.

Yang Wang 0102, Jiaqi Zhu 0001, Miaomiao Li 0007, Jiang Liu, Yilin Li 0003, Yi Yang 0060, Jiafan Li, Hongan Wang

Recently, rule-based time series classification models are widely used in safety-critical scenarios demanding strong interpretability, through providing explicit and rigorous rationales. However, due to the discrepancy between discrete rules and continuous neural networks, the generated rules are not completely consistent with the actual decision-making process, rendering users hesitant to trust the model. Additionally, existing methods learn the role (weight) of each time point independently, which is not in line with human understanding and the consecutive nature of temporal properties. In this paper, we propose a novel neuro-symbolic model named TemporalRule, aiming to automatically learn Signal Temporal Logic (STL) rules for interpretable time series classification. Our model directly optimizes the neural networks representing discrete rules via gradient grafting, producing reliable rules that can exactly determine the classification results. Notably, a temporal logical layer is designed to simulate expressive temporal operators, including one 2-predicate operator (Until) and two 2-level composite operators (EA and AE), through adaptively learning the time bounds of involved intervals. That creates consistent weights for consecutive time points, enhancing the intuitiveness of derived rules. Extensive experiments on diverse real-world datasets in safety-critical domains show that TemporalRule achieves superior and stable classification accuracy compared to state-of-the-art rule-based approaches, and the learned rules can precisely and concisely explicate the classification results, leading to a trustable model.

Mengyu Wang, Tiejun Ma, Shay B. Cohen

Stock selection, which aims to predict stock prices and identify the most profitable ones, is a crucial task in finance. While existing methods primarily focus on developing model structures and building graphs for improved selection, pre-training strategies remain underexplored in this domain. Current stock series pre-training follows methods from other areas without adapting to the unique characteristics of financial data, particularly overlooking stock-specific contextual information and the non-stationary nature of stock prices. Consequently, the latent statistical features inherent in stock data are underutilized. In this paper, we propose three novel pre-training tasks tailored to stock data characteristics: stock code classification, stock sector classification, and moving average prediction. We develop the Stock Specialized Pre-trained Transformer (SSPT) based on a two-layer transformer architecture. Extensive experimental results validate the effectiveness of our pre-training methods and provide detailed guidance on their application. Evaluations on five stock datasets, including four markets and two time periods, demonstrate that SSPT consistently outperforms the market and existing methods in terms of both cumulative investment return ratio and Sharpe ratio. Additionally, our experiments on simulated data investigate the underlying mechanisms of our methods, providing insights into understanding price series. Our code is publicly available at: https://github.com/astudentuser/Pre-training-Time-Series-Models-with-Stock-Data-Customization.