Retrieval-Augmented Generation (RAG) systems can amplify political bias from underlying web corpora. To empirically demonstrate this amplification, we first analyze 16,254 documents from the C4 dataset and 24,300 LLM-generated responses, revealing significant left-leaning and supportive stance bias that can propagate strongly from retrieval to generation. To mitigate this amplification of political bias, we propose FAIR-RAG, an end-to-end framework integrating (1) multi-LLM persona-based annotation, (2) a vector database with political-stance metadata, and (3) a multi-stage fairness engine designed for each of the three stages in RAG systems. FAIR-RAG achieves Attention Weighted Rank Fairness of 97.51 (82.1% improvement) and Perspective Balance of 51.01/82.37 (average 5.6% improvement over state-of-the-art) while maintaining high output quality (Context Precision: 0.974/0.975, Faithfulness: 0.994/0.996). Ablation studies confirm that all three components must operate collaboratively for optimal bias mitigation. This work provides a foundational framework for developing trustworthy and equitable AI information systems. All source code and experimental scripts are publicly available at: https://github.com/bigbases/FAIR-RAG.

The dynamic nature of streaming data often introduces distribution shifts that challenge typical text classification models. This paper proposes an online learning framework tailored for streaming text classification under distribution shifts. First, we decompose a neural network-based text classification model into distinct modules and analyze the varying impact of updating these modules under different types of shifts. Based on this insight, we define three novel indicators to efficiently measure the extent of distribution shifts without evaluating the entire model. These indicators enable the development of predictive models that dynamically optimize module update strategies, balancing learning efficiency and accuracy in real-time. To the best of our knowledge, this is the first approach to systematically adapt model updates according to a trade-off between efficiency and accuracy in online text classification. Extensive experiments on real-world streaming datasets demonstrate the effectiveness of our method, which consistently outperforms both static update strategies and state-of-the-art online text classification models. Our code and data are publicly available athttps://github.com/bigbases/modular-online-adaptation.

Spike Neural Networks offer energy efficiency and are promising candidates for ultra-low-power inference on neuromorphic hardware. While extensively studied in computer vision, their application in Natural Language Processing remains limited and underexplored. Three significant challenges of the existing work are as follows: (1) spike firing functions are sensitive to initial conditions, (2) spike timings are stochastic even for identical token inputs, preventing the stable preservation of contextual information, and (3) the analysis of spike occurrences on learning effectiveness is limited. To improve learning efficiency and stability, we propose Dynamic Spike Predictor (DSP) that adaptively regulates spike generation. DSP predicts a scale-adjusted input current at each time step to regulate spike activity, maintaining stable gradient flow, with only about 0.2% additional parameters to the backbone SNNs. We validate its effectiveness through comprehensive experiments on three NLI benchmarks (CB, RTE, and SICK), addressing research questions on the learning performance, robustness, and extensibility of DSP. The code is available at https://github.com/bigbases/Spike-Predictor.

Search engine personalization, while enhancing user satisfaction, can lead to information disparities. Previous studies on this topic face limitations, such as the absence of context-aware data collection, superficial URL-level analysis, and human-dependent annotations. We propose FAIR-SE, a Framework for Analyzing Information dispaRities in Search Engines that addresses these challenges through AWS Lambda-based concurrent data collection and LLM-generated persona-based content analysis. We collected search results across four user contexts (Search History, Geo-location, Language Preference, and Access Environment) and analyzed them through four analytical perspectives (Political Leaning, Topic-specific Stance, Subjectivity, and Bias). Experiments conducted on two globally prominent search engines across nine controversial topics demonstrate the efficacy of FAIR-SE regarding benchmark accuracy, persona consistency, and ability to reflect real-world discourse patterns across diverse topics. Our statistical analysis identifies distinct search engine characteristics and demonstrates significant information disparities in our case studies examining regional disparities in search results. Our code and datasets are publicly available at: https://github.com/bigbases/FAIR-SE.

Federated learning (FL) is a privacy-preserving machine learning algorithm that enables multiple clients to collaborate. To respond to non-independent and identically distributed (non-IID) environments between clients, personalized FL (PFL) has been actively investigated. The typical PFL model consists of two parts: 1) the head (i.e., classifier) for the final classification and 2) the body (i.e., feature extractor) for extracting representations from local datasets. The head is maintained separately in each client for personalization; the body is aggregated for generalization. FedSDP introduces a bridge layer, called a personalized layer, between the head and the body to preserve individual, non-shared local prototypes for each client. A personalized layer decouples the body and head, strengthening the generalization and personalization, respectively. Based on this architecture, this study proposes a new PFL framework, Federated Self-Derived Prototypes (FedSDP), to dynamically balance personalization and generalization. To this end, we introduce two dynamic adjustments for generating self-derived prototypes: 1) global-local similarity weight (GL-Sim Weight) and 2) personalization early stopping indicator (P-Stop Indicator). GL-Sim Weight based on the similarity between the global and local prototypes is utilized to adjust the degree of personalization of each local model. PStop Indicator is calculated based on the changed degree of local parameters in each client, determining the early stopping for personalization in the client and further concentrating on generalization. Our comprehensive experiments demonstrate that FedSDP outperforms existing state-of-the-art FL frameworks, showing superior effectiveness in non-IID settings. Our code and data are available at https://github.com/bigbases/FedSDP.

Graph representation learning (GRL) aims to map a graph into a low-dimensional vector space while preserving graph topology and node properties. This study proposes a novel GRL model, Multi-Level GRL (simply, ML-GRL), that recursively partitions input graphs by selecting the most appropriate community detection algorithm at each graph or partitioned subgraph. To preserve the relationship between subgraphs, ML-GRL incorporates global graphs that effectively maintain the overall topology. ML-GRL employs a prediction model, which is pre-trained using graph-based features and covers a wide range of graph distributions, to estimate GRL accuracy of each community detection algorithm without partitioning graphs or subgraphs and evaluating them. ML-GRL improves learning accuracy by selecting the most effective community detection algorithm while enhancing learning efficiency from parallel processing of partitioned subgraphs. Through extensive experiments with two different tasks, we demonstrate ML-GRL's superiority over the six representative GRL models in terms of both learning accuracy and efficiency. Specifically, ML-GRL not only improves the accuracy of existing GRL models by 3.68 ~ 47.59% for link prediction and 1.75 ~ 40.90% for node classification but also significantly reduces their running time by 9.63 ~ 62.71% and 7.14 ~ 82.14%, respectively. Our source code is available at https://github.com/pnpy6elp/Multi_Level_GRL.

Continual learning aims to progressively acquire new knowledge while retaining previously learned information, addressing the challenge of catastrophic forgetting. This paper introduces a novel continual learning method, called SPrint, which is the first research effort to devise the principles of self-paced learning for solving continual learning problems. SPrint dynamically adapts the complexity of samples for both new and previous tasks in response to the model’s current learning capacity. It employs a self-paced loss function for sampling new tasks and a forgetting occurrence for sampling previous tasks with replay memory. Through extensive empirical evaluation, we demonstrate that SPrint consistently outperforms state-of-the-art methods in various continual learning benchmarks. Our source code is publicly available at https://github.com/bigbases/SPrint.

This study presents an adaptive online learning method for text classification under distribution shifts. We formulate a typical neural network-based text classification model as multiple logical modules. By leveraging the characteristics of the modules, we introduce three novel indicators to effectively measure the degree of dynamic distribution shifts without evaluating the model. To enhance online learning, we tactically trade off between learning efficiency and accuracy based on distribution shifts measured in real time. To the best of our knowledge, this is the first effort to adapt the model to the preference of learning efficiency or accuracy for online text classification. Extensive experiments on real-world streaming text datasets show that our method outperforms the best static update strategy and state-of-the-art online text classification models. Our code and data are available at https://github.com/bigbases/online-learning-text.

In this study, we propose a novel graph representation learning (GRL) model, called Two-Level GRL with Subgraph-as-a-Node (SaaN 2L-GRL in short), that partitions input graphs into smaller subgraphs for effective and scalable GRL in two levels: 1) local GRL and 2) global GRL. To realize the two-level GRL in an efficient manner, we propose an abstracted graph, called Subgraph-as-a-Node Graph (SaaN in short), to effectively maintain the high-level graph topology while significantly reducing the size of the graph. By applying the SaaN graph to both local and global GRL, SaaN 2L-GRL can effectively preserve the overall structure of the entire graph while precisely representing the nodes within each subgraph. Through time complexity analysis, we confirm that SaaN 2L-GRL significantly reduces the learning time of existing GRL models by using the SaaN graph for global GRL, instead of using the original graph, and processing local GRL on subgraphs in parallel. Our extensive experiments show that SaaN 2L-GRL outperforms existing GRL models in both accuracy and efficiency. In addition, we show the effectiveness of SaaN 2L-GRL using diverse kinds of graph partitioning methods, including five community detection algorithms and representative edge- and vertex-cut algorithms.

In this paper, we propose a novel graph representation learning (GRL) model that aims to improve both representation accuracy and learning efficiency. We design a Two-Level GRL architecture based on the graph partitioning: 1) local GRL on nodes within each partitioned subgraph and 2) global GRL on subgraphs. By partitioning the graph through community detection, we enable elaborate node learning in the same community. Based on Two-Level GRL, we introduce an abstracted graph, Community-as-a-Node Graph(CaaN), to effectively maintain the high-level structure with a significantly reduced graph. By applying the CaaN graph to local and global GRL, we propose Two-Level GRL with Community-as-a-Node (CaaN 2L) that effectively maintains the global structure of the entire graph while accurately representing the nodes in each community. A salient point of the proposed model is that it can be applied to any existing GRL model by adopting it as the base model for local and global GRL. Through extensive experiments employing seven popular GRL models, we show that our model outperforms them in both accuracy and efficiency.

Recently, parallel search engines have been implemented based on scalable distributed file systems such as Google File System. However, we claim that building a massively-parallel search engine using a parallel DBMS can be an attractive alternative since it supports a higher-level (i.e., SQL-level) interface than that of a distributed file system for easy and less error-prone application development while providing scalability. Regarding higher-level functionality, we can draw a parallel with the traditional O/S file system vs. DBMS. In this paper, we propose a new approach of building a massively-parallel search engine using a DB-IR tightly-integrated parallel DBMS. To estimate the performance, we propose a hybrid (i.e., analytic and experimental) performance model for the parallel search engine. We argue that the model can accurately estimate the performance of a massively-parallel (e.g., 300-node) search engine using the experimental results obtained from a small-scale (e.g., 5-node) one. We show that the estimation error between the model and the actual experiment is less than 2.13% by observing that the bulk of the query processing time is spent at the slave (vs. at the master and network) and by estimating the time spent at the slave based on actual measurement. Using our model, we demonstrate a commercial-level scalability and performance of our architecture. Our proposed system ODYS is capable of handling 1 billion queries per day (81 queries/sec) for 30 billion Web pages by using only 43,472 nodes with an average query response time of 194 ms. By using twice as many (86,944) nodes, ODYS can provide an average query response time of 148 ms. These results show that building a massively-parallel search engine using a parallel DBMS is a viable approach with advantages of supporting the high-level (i.e., DBMS-level), SQL-like programming interface.