Ming Zhang

Wei Ju, Xiao Luo, Meng Qu et al.

This paper studies semi-supervised graph classification, a crucial task with a wide range of applications in social network analysis and bioinformatics. Recent works typically adopt graph neural networks to learn graph-level representations for classification, failing to explicitly leverage features derived from graph topology (e.g., paths). Moreover, when labeled data is scarce, these methods are far from satisfactory due to their insufficient topology exploration of unlabeled data. We address the challenge by proposing a novel semi-supervised framework called Twin Graph Neural Network (TGNN). To explore graph structural information from complementary views, our TGNN has a message passing module and a graph kernel module. To fully utilize unlabeled data, for each module, we calculate the similarity of each unlabeled graph to other labeled graphs in the memory bank and our consistency loss encourages consistency between two similarity distributions in different embedding spaces. The two twin modules collaborate with each other by exchanging instance similarity knowledge to fully explore the structure information of both labeled and unlabeled data. We evaluate our TGNN on various public datasets and show that it achieves strong performance.

Siyu Yi, Zhengyang Mao, Wei Ju et al. · pku

Graph classification, aiming at learning the graph-level representations for effective class assignments, has received outstanding achievements, which heavily relies on high-quality datasets that have balanced class distribution. In fact, most real-world graph data naturally presents a long-tailed form, where the head classes occupy much more samples than the tail classes, it thus is essential to study the graph-level classification over long-tailed data while still remaining largely unexplored. However, most existing long-tailed learning methods in visions fail to jointly optimize the representation learning and classifier training, as well as neglect the mining of the hard-to-classify classes. Directly applying existing methods to graphs may lead to sub-optimal performance, since the model trained on graphs would be more sensitive to the long-tailed distribution due to the complex topological characteristics. Hence, in this paper, we propose a novel long-tailed graph-level classification framework via Collaborative Multi-expert Learning (CoMe) to tackle the problem. To equilibrate the contributions of head and tail classes, we first develop balanced contrastive learning from the view of representation learning, and then design an individual-expert classifier training based on hard class mining. In addition, we execute gated fusion and disentangled knowledge distillation among the multiple experts to promote the collaboration in a multi-expert framework. Comprehensive experiments are performed on seven widely-used benchmark datasets to demonstrate the superiority of our method CoMe over state-of-the-art baselines.

Jingqi Tong, Yurong Mou, Hangcheng Li et al.

"Thinking with Text" and "Thinking with Images" paradigm significantly improve the reasoning ability of large language models (LLMs) and Vision Language Models (VLMs). However, these paradigms have inherent limitations. (1) Images capture only single moments and fail to represent dynamic processes or continuous changes, and (2) The separation of text and vision as distinct modalities, hindering unified multimodal understanding and generation. To overcome these limitations, we introduce "Thinking with Video", a new paradigm that leverages video generation models, such as Sora-2, to bridge visual and textual reasoning in a unified temporal framework. To support this exploration, we developed the Video Thinking Benchmark (VideoThinkBench). VideoThinkBench encompasses two task categories: (1) vision-centric tasks (e.g., Eyeballing Puzzles), and (2) text-centric tasks (e.g., subsets of GSM8K, MMMU). Our evaluation establishes Sora-2 as a capable reasoner. On vision-centric tasks, Sora-2 is generally comparable to state-of-the-art (SOTA) VLMs, and even surpasses VLMs on several tasks, such as Eyeballing Games. On text-centric tasks, Sora-2 achieves 92% accuracy on MATH, and 75.53% accuracy on MMMU. Furthermore, we systematically analyse the source of these abilities. We also find that self-consistency and in-context learning can improve Sora-2's performance. In summary, our findings demonstrate that the video generation model is the potential unified multimodal understanding and generation model, positions "thinking with video" as a unified multimodal reasoning paradigm.

Junyu Luo, Bohan Wu, Xiao Luo et al. · pku

Post-training of Large Language Models (LLMs) is crucial for unlocking their task generalization potential and domain-specific capabilities. However, the current LLM post-training paradigm faces significant data challenges, including the high costs of manual annotation and diminishing marginal returns on data scales. Therefore, achieving data-efficient post-training has become a key research question. In this paper, we present the first systematic survey of data-efficient LLM post-training from a data-centric perspective. We propose a taxonomy of data-efficient LLM post-training methods, covering data selection, data quality enhancement, synthetic data generation, data distillation and compression, and self-evolving data ecosystems. We summarize representative approaches in each category and outline future research directions. By examining the challenges in data-efficient LLM post-training, we highlight open problems and propose potential research avenues. We hope our work inspires further exploration into maximizing the potential of data utilization in large-scale model training. Paper List: https://github.com/luo-junyu/Awesome-Data-Efficient-LLM

Yue Zhang, Ming Zhang, Haipeng Yuan et al.

Recently, the evaluation of Large Language Models has emerged as a popular area of research. The three crucial questions for LLM evaluation are ``what, where, and how to evaluate''. However, the existing research mainly focuses on the first two questions, which are basically what tasks to give the LLM during testing and what kind of knowledge it should deal with. As for the third question, which is about what standards to use, the types of evaluators, how to score, and how to rank, there hasn't been much discussion. In this paper, we analyze evaluation methods by comparing various criteria with both manual and automatic evaluation, utilizing onsite, crowd-sourcing, public annotators and GPT-4, with different scoring methods and ranking systems. We propose a new dataset, LLMEval and conduct evaluations on 20 LLMs. A total of 2,186 individuals participated, leading to the generation of 243,337 manual annotations and 57,511 automatic evaluation results. We perform comparisons and analyses of different settings and conduct 10 conclusions that can provide some insights for evaluating LLM in the future. The dataset and the results are publicly available at https://github.com/llmeval .

Yu Pan, Ye Yuan, Yichun Yin et al.

The rapid progress of Transformers in artificial intelligence has come at the cost of increased resource consumption and greenhouse gas emissions due to growing model sizes. Prior work suggests using pretrained small models to improve training efficiency, but this approach may not be suitable for new model structures. On the other hand, training from scratch can be slow, and progressively stacking layers often fails to achieve significant acceleration. To address these challenges, we propose a novel method called Apollo, which prep\textbf{a}res lessons for ex\textbf{p}anding \textbf{o}perations by \textbf{l}earning high-\textbf{l}ayer functi\textbf{o}nality during training of low layers. Our approach involves low-value-prioritized sampling (LVPS) to train different depths and weight sharing to facilitate efficient expansion. We also introduce an interpolation method for stable model depth extension. Experiments demonstrate that Apollo achieves state-of-the-art acceleration ratios, even rivaling methods using pretrained models, making it a universal and efficient solution for training deep models while reducing time, financial, and environmental costs.

Changhao Jiang, Ming Zhang, Yifei Cao et al.

The GPT-4 technical report suggests that downstream performance can be predicted from pre-training signals, but offers little methodological detail on how to quantify this. This work address this gap by modeling knowledge retention, the capacity of a pre-trained language model to memorize factual information from its corpus, and introduce a principled method to estimate it prior to training. We propose Size-dependent Mutual Information (SMI), an information-theoretic predictor that integrates knowledge frequency, knowledge specificity, and model size to forecast closed-book question answering (QA) accuracy. SMI is validated through large-scale document retrieval over the disclosed pre-training corpora of 21 public and 3 custom models, combined with a robust multi-template QA evaluation. Experiments show that SMI significantly outperforms repetition-based baselines and achieves $R^2$ > 0.7 in predicting QA accuracy for models above 1B parameters, without additional training. The analysis further reveals diminishing returns from scaling data and model size and provides evidence for an intrinsic upper bound on knowledge retention achievable by pre-training alone, motivating retrieval and other augmentation strategies.

Yikai Zhu, Jianhao Shen, Ming Zhang

In the task of factoid question answering over knowledge base, many questions have more than one plausible interpretation. Previous works on SimpleQuestions assume only one interpretation as the ground truth for each question, so they lack the ability to answer ambiguous questions correctly. In this paper, we present a new way to utilize the dataset that takes into account the existence of ambiguous questions. Then we introduce a simple and effective model which combines local knowledge subgraph with attention mechanism. Our experimental results show that our approach achieves outstanding performance in this task.

Luchen Liu, Haoxian Wu, Zichang Wang et al.

Sepsis is a life-threatening condition that seriously endangers millions of people over the world. Hopefully, with the widespread availability of electronic health records (EHR), predictive models that can effectively deal with clinical sequential data increase the possibility to predict sepsis and take early preventive treatment. However, the early prediction is challenging because patients' sequential data in EHR contains temporal interactions of multiple clinical events. And capturing temporal interactions in the long event sequence is hard for traditional LSTM. Rather than directly applying the LSTM model to the event sequences, our proposed model firstly aggregates heterogeneous clinical events in a short period and then captures temporal interactions of the aggregated representations with LSTM. Our proposed Heterogeneous Event Aggregation can not only shorten the length of clinical event sequence but also help to retain temporal interactions of both categorical and numerical features of clinical events in the multiple heads of the aggregation representations. In the PhysioNet/Computing in Cardiology Challenge 2019, with the team named PKU_DLIB, our proposed model, in high efficiency, achieved utility score (0.321) in the full test set.

Luchen Liu, Haoran Li, Zhiting Hu et al.

Clinical outcome prediction based on the Electronic Health Record (EHR) plays a crucial role in improving the quality of healthcare. Conventional deep sequential models fail to capture the rich temporal patterns encoded in the longand irregular clinical event sequences. We make the observation that clinical events at a long time scale exhibit strongtemporal patterns, while events within a short time period tend to be disordered co-occurrence. We thus propose differentiated mechanisms to model clinical events at different time scales. Our model learns hierarchical representationsof event sequences, to adaptively distinguish between short-range and long-range events, and accurately capture coretemporal dependencies. Experimental results on real clinical data show that our model greatly improves over previous state-of-the-art models, achieving AUC scores of 0.94 and 0.90 for predicting death and ICU admission respectively, Our model also successfully identifies important events for different clinical outcome prediction tasks

Luchen Liu, Jianhao Shen, Ming Zhang et al.

The availability of a large amount of electronic health records (EHR) provides huge opportunities to improve health care service by mining these data. One important application is clinical endpoint prediction, which aims to predict whether a disease, a symptom or an abnormal lab test will happen in the future according to patients' history records. This paper develops deep learning techniques for clinical endpoint prediction, which are effective in many practical applications. However, the problem is very challenging since patients' history records contain multiple heterogeneous temporal events such as lab tests, diagnosis, and drug administrations. The visiting patterns of different types of events vary significantly, and there exist complex nonlinear relationships between different events. In this paper, we propose a novel model for learning the joint representation of heterogeneous temporal events. The model adds a new gate to control the visiting rates of different events which effectively models the irregular patterns of different events and their nonlinear correlations. Experiment results with real-world clinical data on the tasks of predicting death and abnormal lab tests prove the effectiveness of our proposed approach over competitive baselines.

Yichun Yin, Furu Wei, Li Dong et al.

In this paper, we develop a novel approach to aspect term extraction based on unsupervised learning of distributed representations of words and dependency paths. The basic idea is to connect two words (w1 and w2) with the dependency path (r) between them in the embedding space. Specifically, our method optimizes the objective w1 + r = w2 in the low-dimensional space, where the multi-hop dependency paths are treated as a sequence of grammatical relations and modeled by a recurrent neural network. Then, we design the embedding features that consider linear context and dependency context information, for the conditional random field (CRF) based aspect term extraction. Experimental results on the SemEval datasets show that, (1) with only embedding features, we can achieve state-of-the-art results; (2) our embedding method which incorporates the syntactic information among words yields better performance than other representative ones in aspect term extraction.

Jian Tang, Ming Zhang, Qiaozhu Mei

It has always been a burden to the users of statistical topic models to predetermine the right number of topics, which is a key parameter of most topic models. Conventionally, automatic selection of this parameter is done through either statistical model selection (e.g., cross-validation, AIC, or BIC) or Bayesian nonparametric models (e.g., hierarchical Dirichlet process). These methods either rely on repeated runs of the inference algorithm to search through a large range of parameter values which does not suit the mining of big data, or replace this parameter with alternative parameters that are less intuitive and still hard to be determined. In this paper, we explore to "eliminate" this parameter from a new perspective. We first present a nonparametric treatment of the PLSA model named nonparametric probabilistic latent semantic analysis (nPLSA). The inference procedure of nPLSA allows for the exploration and comparison of different numbers of topics within a single execution, yet remains as simple as that of PLSA. This is achieved by substituting the parameter of the number of topics with an alternative parameter that is the minimal goodness of fit of a document. We show that the new parameter can be further eliminated by two parameter-free treatments: either by monitoring the diversity among the discovered topics or by a weak supervision from users in the form of an exemplar topic. The parameter-free topic model finds the appropriate number of topics when the diversity among the discovered topics is maximized, or when the granularity of the discovered topics matches the exemplar topic. Experiments on both synthetic and real data prove that the parameter-free topic model extracts topics with a comparable quality comparing to classical topic models with "manual transmission". The quality of the topics outperforms those extracted through classical Bayesian nonparametric models.