Search

Large language models (LLMs) have demonstrated impressive versatility across numerous tasks, yet their generalization capabilities remain poorly understood. To investigate these behaviors, arithmetic tasks serve as important venues. In previous studies, seemingly unrelated mysteries still exist -- (1) models with appropriate positional embeddings can correctly perform longer unseen arithmetic operations such as addition, but their effectiveness varies in more complex tasks like multiplication; (2) models perform well for longer unseen cases in modular addition under specific moduli (e.g., modulo 100) but struggle under very close moduli (e.g., modulo 101), regardless of the positional encoding used. We believe previous studies have been treating the symptoms rather than addressing the root cause -- they have paid excessive attention to improving model components, while overlooking the differences in task properties that may be the real drivers. This is confirmed by our unified theoretical framework for different arithmetic scenarios. For example, unlike multiplication, the digital addition task has the property of translation invariance which naturally aligns with the relative positional encoding, and this combination leads to successful generalization of addition to unseen longer domains. The discrepancy in operations modulo 100 and 101 arises from the base. Modulo 100, unlike 101, is compatible with the decimal system (base 10), such that unseen information in digits beyond the units digit and the tens digit is actually not needed for the task. Extensive experiments with GPT-like models validate our theoretical predictions. These findings deepen our understanding of the generalization mechanisms, and facilitate more data-efficient model training and objective-oriented AI alignment.

Current directed graph embedding methods build upon undirected techniques but often inadequately capture directed edge information, leading to challenges such as: (1) Suboptimal representations for nodes with low in/out-degrees, due to the insufficient neighbor interactions; (2) Limited inductive ability for representing new nodes post-training; (3) Narrow generalizability, as training is overly coupled with specific tasks. In response, we propose DUPLEX, an inductive framework for complex embeddings of directed graphs. It (1) leverages Hermitian adjacency matrix decomposition for comprehensive neighbor integration, (2) employs a dual GAT encoder for directional neighbor modeling, and (3) features two parameter-free decoders to decouple training from particular tasks. DUPLEX outperforms state-of-the-art models, especially for nodes with sparse connectivity, and demonstrates robust inductive capability and adaptability across various tasks. The code is available at https://github.com/alipay/DUPLEX.

Position bias, i.e., users' preference of an item is affected by its placing position, is well studied in the recommender system literature. However, most existing methods ignore the widely coupled ranking bias, which is also related to the placing position of the item. Using both synthetic and industrial datasets, we first show how this widely coexisted ranking bias deteriorates the performance of the existing position bias estimation methods. To mitigate the position bias with the presence of the ranking bias, we propose a novel position bias estimation method, namely gradient interpolation, which fuses two estimation methods using a fusing weight. We further propose an adaptive method to automatically determine the optimal fusing weight. Extensive experiments on both synthetic and industrial datasets demonstrate the superior performance of the proposed methods., Comment: 5 pages, 3 figures

The life trajectories of notable people have been studied to pinpoint the times and places of significant events such as birth, death, education, marriage, competition, work, speeches, scientific discoveries, artistic achievements, and battles. Understanding how these individuals interact with others provides valuable insights for broader research into human dynamics. However, the scarcity of trajectory data in terms of volume, density, and inter-person interactions, limits relevant studies from being comprehensive and interactive. We mine millions of biography pages from Wikipedia and tackle the generalization problem stemming from the variety and heterogeneity of the trajectory descriptions. Our ensemble model COSMOS, which combines the idea of semi-supervised learning and contrastive learning, achieves an F1 score of 85.95%. For this task, we also create a hand-curated dataset, WikiLifeTrajectory, consisting of 8,852 (person, time, location) triplets as ground truth. Besides, we perform an empirical analysis on the trajectories of 8,272 historians to demonstrate the validity of the extracted results. To facilitate the research on trajectory extractions and help the analytical studies to construct grand narratives, we make our code, the million-level extracted trajectories, and the WikiLifeTrajectory dataset publicly available., Comment: Accepted to ICWSM 2025. 15 pages

Achieving gender equality is a pivotal factor in realizing the UN's Global Goals for Sustainable Development. Gender bias studies work towards this and rely on name-based gender inference tools to assign individual gender labels when gender information is unavailable. However, these tools often inaccurately predict gender for Chinese Pinyin names, leading to potential bias in such studies. With the growing participation of Chinese in international activities, this situation is becoming more severe. Specifically, current tools focus on pronunciation (Pinyin) information, neglecting the fact that the latent connections between Pinyin and Chinese characters (Hanzi) behind convey critical information. As a first effort, we formulate the Pinyin name-gender guessing problem and design a Multi-Task Learning Network assisted by Knowledge Distillation that enables the Pinyin embeddings in the model to possess semantic features of Chinese characters and to learn gender information from Chinese character names. Our open-sourced method surpasses commercial name-gender guessing tools by 9.70\% to 20.08\% relatively, and also outperforms the state-of-the-art algorithms.

Better forecasts may not lead to better decision-making. To address this challenge, decision-oriented learning (DOL) has been proposed as a new branch of machine learning that replaces traditional statistical loss with a decision loss to form an end-to-end model. Applications of DOL in power systems have been developed in recent years. For renewable-rich power systems, uncertainties propagate through sequential tasks, where traditional statistical-based approaches focus on minimizing statistical errors at intermediate stages but may fail to provide optimal decisions at the final stage. This paper first elaborates on the mismatch between more accurate forecasts and more optimal decisions in the power system caused by statistical-based learning (SBL) and explains how DOL resolves this problem. Secondly, this paper extensively reviews DOL techniques and their applications in power systems while highlighting their pros and cons in relation to SBL. Finally, this paper identifies the challenges to adopt DOL in the energy sector and presents future research directions.

Forecast-then-optimize is a widely-used framework for decision-making problems in power systems. Traditionally, statistical losses have been employed to train forecasting models, but recent research demonstrated that improved decision utility in downstream optimization tasks can be achieved by using decision loss as an alternative. However, the implementation of decision loss in power systems faces challenges in 1) accommodating multi-stage decision-making problems where upstream optimality cannot guarantee final optimality; 2) adapting to dynamic environments such as changing parameters and nature of the problem like continuous or discrete optimization tasks. To this end, this paper proposes a novel adaptive decision-objective loss (ADOL) to address the above challenges. Specifically, ADOL first redefines the decision loss as objective utilities rather than objective loss to eliminate the need to manually set the optimal decision, thus ensuring the globally optimal decision. ADOL enables one-off training in a dynamic environment by introducing additional variables. The differentiability and convexity of ADOL provide useful gradients for forecasting model training in conjunction with continuous and discrete optimization tasks. Experiments are conducted for both linear programming-based and mixed integer linear programming-based power system two-stage dispatching cases with changing costs, and the results show that the proposed ADOL is capable of achieving globally optimal decision-making and adaptability to dynamic environments. The method can be extended to other multi-stage tasks in complex systems.

Uplift modeling is widely used in performance marketing to estimate effects of promotion campaigns (e.g., increase of customer retention rate). Since it is impossible to observe outcomes of a recipient in treatment (e.g., receiving a certain promotion) and control (e.g., without promotion) groups simultaneously (i.e., counter-factual), uplift models are mainly trained on instances of treatment and control groups separately to form two models respectively, and uplifts are predicted by the difference of predictions from these two models (i.e., two-model method). When responses are noisy and the treatment effect is fractional, induced individual uplift predictions will be inaccurate, resulting in targeting undesirable customers. Though it is impossible to obtain the ideal ground-truth individual uplifts, known as Individual Treatment Effects (ITEs), alternatively, an average uplift of a group of users, called Average Treatment Effect (ATE), can be observed from experimental deliveries. Upon this, similar to Multiple Instance Learning (MIL) in which each training sample is a bag of instances, our framework sums up individual user uplift predictions for each bag of users as its bag-wise ATE prediction, and regularizes it to its ATE label, thus learning more accurate individual uplifts. Additionally, to amplify the fractional treatment effect, bags are composed of instances with adjacent individual uplift predictions, instead of random instances. Experiments conducted on two datasets show the effectiveness and universality of the proposed framework., Comment: short paper of CIKM22(full version)

Large language models (LLMs) have achieved remarkable proficiency on solving diverse problems. However, their generalization ability is not always satisfying and the generalization problem is common for generative transformer models in general. Researchers take basic mathematical tasks like n-digit addition or multiplication as important perspectives for investigating their generalization behaviors. It is observed that when training models on n-digit operations (e.g., additions) in which both input operands are n-digit in length, models generalize successfully on unseen n-digit inputs (in-distribution (ID) generalization), but fail miserably on longer, unseen cases (out-of-distribution (OOD) generalization). We bring this unexplained performance drop into attention and ask whether there is systematic OOD generalization. Towards understanding LLMs, we train various smaller language models which may share the same underlying mechanism. We discover that the strong ID generalization stems from structured representations, while behind the unsatisfying OOD performance, the models still exhibit clear learned algebraic structures. Specifically, these models map unseen OOD inputs to outputs with learned equivalence relations in the ID domain, which we call the equivalence generalization. These findings deepen our knowledge regarding the generalizability of generative models including LLMs, and provide insights into potential avenues for improvement., Comment: Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI-24)

Celebrities' whereabouts are of pervasive importance. For instance, where politicians go, how often they visit, and who they meet, come with profound geopolitical and economic implications. Although news articles contain travel information of celebrities, it is not possible to perform large-scale and network-wise analysis due to the lack of automatic itinerary detection tools. To design such tools, we have to overcome difficulties from the heterogeneity among news articles: 1)One single article can be noisy, with irrelevant people and locations, especially when the articles are long. 2)Though it may be helpful if we consider multiple articles together to determine a particular trip, the key semantics are still scattered across different articles intertwined with various noises, making it hard to aggregate them effectively. 3)Over 20% of the articles refer to the celebrities' trips indirectly, instead of using the exact celebrity names or location names, leading to large portions of trips escaping regular detecting algorithms. We model text content across articles related to each candidate location as a graph to better associate essential information and cancel out the noises. Besides, we design a special pooling layer based on attention mechanism and node similarity, reducing irrelevant information from longer articles. To make up the missing information resulted from indirect mentions, we construct knowledge sub-graphs for named entities (person, organization, facility, etc.). Specifically, we dynamically update embeddings of event entities like the G7 summit from news descriptions since the properties (date and location) of the event change each time, which is not captured by the pre-trained event representations. The proposed CeleTrip jointly trains these modules, which outperforms all baseline models and achieves 82.53% in the F1 metric., Comment: Accepted to ICWSM 2024, 12 pages

Some scientists are more likely to explore unfamiliar research topics while others tend to exploit existing ones. In previous work, correlations have been found between scientists' topic choices and their career performances. However, literature has yet to untangle the intricate interplay between scientific impact and research topic choices, where scientific exploration and exploitation intertwine. Here we study two metrics that gauge how frequently scientists switch topic areas and how large those jumps are, and discover that 'cautious explorers' who switch topics frequently but do so to 'close' domains have notably better future performance and can be identified at a remarkably early career stage. Cautious explorers who balance exploration and exploitation in their first four career years have up to 19% more citations per future paper. Our results suggest that the proposed metrics depict the scholarly traits of scientists throughout their careers and provide fresh insight, especially for nurturing junior scientists., Comment: 16 pages of main text and 94 pages of supplementary information. v2: Added page number and fixed typo in author list

The objective-based forecasting considers the asymmetric and non-linear impacts of forecasting errors on decision objectives, thus improving the effectiveness of its downstream decision-making process. However, existing objective-based forecasting methods are risk-neutral and not suitable for tasks like power system inertia management and unit commitment, of which decision-makers are usually biased toward risk aversion in practice. To tackle this problem, this paper proposes a generic risk-aware objective-based forecasting method. It enables decision-makers to customize their forecasting with different risk preferences. The equivalence between the proposed method and optimization under uncertainty (stochastic/robust optimization) is established for the first time. Case studies are carried out on a Great Britain 2030 power system with system operational data from National Grid. The results show that the proposed model with deterministic optimization can approximate the performance of stochastic programming or robust optimization at only a fraction of their computational cost.

Achieving gender equality is an important pillar for humankind's sustainable future. Pioneering data-driven gender bias research is based on large-scale public records such as scientific papers, patents, and company registrations, covering female researchers, inventors and entrepreneurs, and so on. Since gender information is often missing in relevant datasets, studies rely on tools to infer genders from names. However, available open-sourced Chinese gender-guessing tools are not yet suitable for scientific purposes, which may be partially responsible for female Chinese being underrepresented in mainstream gender bias research and affect their universality. Specifically, these tools focus on character-level information while overlooking the fact that the combinations of Chinese characters in multi-character names, as well as the components and pronunciations of characters, convey important messages. As a first effort, we design a Chinese Heterogeneous Graph Attention (CHGAT) model to capture the heterogeneity in component relationships and incorporate the pronunciations of characters. Our model largely surpasses current tools and also outperforms the state-of-the-art algorithm. Last but not least, the most popular Chinese name-gender dataset is single-character based with far less female coverage from an unreliable source, naturally hindering relevant studies. We open-source a more balanced multi-character dataset from an official source together with our code, hoping to help future research promoting gender equality., Comment: 8 pages, 4 figures

Restructuring the regional industrial structure (RIS) has the potential to halt economic recession and achieve revitalization. Understanding the current status and dynamics of RIS will greatly assist in studying and evaluating the current industrial structure. Previous studies have focused on qualitative and quantitative research to rationalize RIS from a macroscopic perspective. Although recent studies have traced information at the industrial enterprise level to complement existing research from a micro perspective, the ambiguity of the underlying variables contributing to the industrial sector and its composition, the dynamic nature, and the large number of multivariant features of RIS records have obscured a deep and fine-grained understanding of RIS. To this end, we propose an interactive visualization system, RISeer, which is based on interpretable machine learning models and enhanced visualizations designed to identify the evolutionary patterns of the RIS and facilitate inter-regional inspection and comparison. Two case studies confirm the effectiveness of our approach, and feedback from experts indicates that RISeer helps them to gain a fine-grained understanding of the dynamics and evolution of the RIS., Comment: IEEE Transactions on Visualization and Computer Graphics (Proc. IEEE VIS 2022)

Fine-tuning pretrained models is a common practice in domain generalization (DG) tasks. However, fine-tuning is usually computationally expensive due to the ever-growing size of pretrained models. More importantly, it may cause over-fitting on source domain and compromise their generalization ability as shown in recent works. Generally, pretrained models possess some level of generalization ability and can achieve decent performance regarding specific domains and samples. However, the generalization performance of pretrained models could vary significantly over different test domains even samples, which raises challenges for us to best leverage pretrained models in DG tasks. In this paper, we propose a novel domain generalization paradigm to better leverage various pretrained models, named specialized ensemble learning for domain generalization (SEDGE). It first trains a linear label space adapter upon fixed pretrained models, which transforms the outputs of the pretrained model to the label space of the target domain. Then, an ensemble network aware of model specialty is proposed to dynamically dispatch proper pretrained models to predict each test sample. Experimental studies on several benchmarks show that SEDGE achieves significant performance improvements comparing to strong baselines including state-of-the-art method in DG tasks and reduces the trainable parameters by ~99% and the training time by ~99.5%.

Event extraction is a classic task in natural language processing with wide use in handling large amount of yet rapidly growing financial, legal, medical, and government documents which often contain multiple events with their elements scattered and mixed across the documents, making the problem much more difficult. Though the underlying relations between event elements to be extracted provide helpful contextual information, they are somehow overlooked in prior studies. We showcase the enhancement to this task brought by utilizing the knowledge graph that captures entity relations and their attributes. We propose a first event extraction framework that embeds a knowledge graph through a Graph Neural Network and integrates the embedding with regular features, all at document-level. Specifically, for extracting events from Chinese financial announcements, our method outperforms the state-of-the-art method by 5.3% in F1-score.