This study investigates the feasibility and performance of federated learning (FL) for multi-label ICD code classification using clinical notes from the MIMIC-IV dataset. Unlike previous approaches that rely on centralized training or fine-tuned large language models, we propose a lightweight and scalable pipeline combining frozen text embeddings with simple multilayer perceptron (MLP) classifiers. This design offers a privacy-preserving and deployment-efficient alternative for clinical NLP applications, particularly suited to distributed healthcare settings. Extensive experiments across both centralized and federated configurations were conducted, testing six publicly available embedding models from Massive Text Embedding Benchmark leaderboard and three MLP classifier architectures under two medical coding (ICD-9 and ICD-10). Additionally, ablation studies over ten random stratified splits assess performance stability. Results show that embedding quality substantially outweighs classifier complexity in determining predictive performance, and that federated learning can closely match centralized results in idealized conditions. While the models are orders of magnitude smaller than state-of-the-art architectures and achieved competitive micro and macro F1 scores, limitations remain including the lack of end-to-end training and the simplified FL assumptions. Nevertheless, this work demonstrates a viable way toward scalable, privacy-conscious medical coding systems and offers a step toward for future research into federated, domain-adaptive clinical AI.
Identifying vulnerabilities in source code is crucial, especially in critical software components. Existing methods such as static analysis, dynamic analysis, formal verification, and recently Large Language Models are widely used to detect security flaws. This paper introduces CASTLE (CWE Automated Security Testing and Low-Level Evaluation), a benchmarking framework for evaluating the vulnerability detection capabilities of different methods. We assess 13 static analysis tools, 10 LLMs, and 2 formal verification tools using a hand-crafted dataset of 250 micro-benchmark programs covering 25 common CWEs. We propose the CASTLE Score, a novel evaluation metric to ensure fair comparison. Our results reveal key differences: ESBMC (a formal verification tool) minimizes false positives but struggles with vulnerabilities beyond model checking, such as weak cryptography or SQL injection. Static analyzers suffer from high false positives, increasing manual validation efforts for developers. LLMs perform exceptionally well in the CASTLE dataset when identifying vulnerabilities in small code snippets. However, their accuracy declines, and hallucinations increase as the code size grows. These results suggest that LLMs could play a pivotal role in future security solutions, particularly within code completion frameworks, where they can provide real-time guidance to prevent vulnerabilities. The dataset is accessible at https://github.com/CASTLE-Benchmark.
Despite the popularity and widespread use of semi-structured data formats such as JSON, end-to-end supervised learning applied directly to such data remains underexplored. We present ORIGAMI (Object RepresentatIon via Generative Autoregressive ModellIng), a transformer-based architecture that directly processes nested key/value pairs while preserving their hierarchical semantics. Our key technical contributions include: (1) a structure-preserving tokenizer, (2) a novel key/value position encoding scheme, and (3) a grammar-constrained training and inference framework that ensures valid outputs and accelerates training convergence. These enhancements enable efficient end-to-end modeling of semi-structured data. By reformulating classification as next-token prediction, ORIGAMI naturally handles both single-label and multi-label tasks without architectural modifications. Empirical evaluation across diverse domains demonstrates ORIGAMI's effectiveness: On standard tabular benchmarks converted to JSON, ORIGAMI remains competitive with classical and state-of-the-art approaches. On native JSON datasets, we outperform baselines on multi-label classification and specialized models such as convolutional and graph neural networks on a code classification task. Through extensive ablation studies, we validate the impact of each architectural component and establish ORIGAMI as a robust framework for end-to-end learning on semi-structured data.
Addressing the complexity of accurately classifying International Classification of Diseases (ICD) codes from medical discharge summaries is challenging due to the intricate nature of medical documentation. This paper explores the use of Large Language Models (LLM), specifically the LLAMA architecture, to enhance ICD code classification through two methodologies: direct application as a classifier and as a generator of enriched text representations within a Multi-Filter Residual Convolutional Neural Network (MultiResCNN) framework. We evaluate these methods by comparing them against state-of-the-art approaches, revealing LLAMA's potential to significantly improve classification outcomes by providing deep contextual insights into medical texts.
We share observations and challenges from an ongoing effort to implement Explainable AI (XAI) in a domain-specific workflow for cybersecurity analysts. Specifically, we briefly describe a preliminary case study on the use of XAI for source code classification, where accurate assessment and timeliness are paramount. We find that the outputs of state-of-the-art saliency explanation techniques (e.g., SHAP or LIME) are lost in translation when interpreted by people with little AI expertise, despite these techniques being marketed for non-technical users. Moreover, we find that popular XAI techniques offer fewer insights for real-time human-AI workflows when they are post hoc and too localized in their explanations. Instead, we observe that cyber analysts need higher-level, easy-to-digest explanations that can offer as little disruption as possible to their workflows. We outline unaddressed gaps in practical and effective XAI, then touch on how emerging technologies like Large Language Models (LLMs) could mitigate these existing obstacles.
Researchers have investigated the potential of leveraging pre-trained language models, such as CodeBERT, to enhance source code-related tasks. Previous methodologies have relied on CodeBERT's '[CLS]' token as the embedding representation of input sequences for task performance, necessitating additional neural network layers to enhance feature representation, which in turn increases computational expenses. These approaches have also failed to fully leverage the comprehensive knowledge inherent within the source code and its associated text, potentially limiting classification efficacy. We propose CodeClassPrompt, a text classification technique that harnesses prompt learning to extract rich knowledge associated with input sequences from pre-trained models, thereby eliminating the need for additional layers and lowering computational costs. By applying an attention mechanism, we synthesize multi-layered knowledge into task-specific features, enhancing classification accuracy. Our comprehensive experimentation across four distinct source code-related tasks reveals that CodeClassPrompt achieves competitive performance while significantly reducing computational overhead.
Transformers, renowned for their self-attention mechanism, have achieved state-of-the-art performance across various tasks in natural language processing, computer vision, time-series modeling, etc. However, one of the challenges with deep Transformer models is the oversmoothing problem, where representations across layers converge to indistinguishable values, leading to significant performance degradation. We interpret the original self-attention as a simple graph filter and redesign it from a graph signal processing (GSP) perspective. We propose a graph-filter-based self-attention (GFSA) to learn a general yet effective one, whose complexity, however, is slightly larger than that of the original self-attention mechanism. We demonstrate that GFSA improves the performance of Transformers in various fields, including computer vision, natural language processing, graph-level tasks, speech recognition, and code classification.
Categorizing source codes accurately and efficiently is a challenging problem in real-world programming education platform management. In recent years, model-based approaches utilizing abstract syntax trees (ASTs) have been widely applied to code classification tasks. We introduce an approach named the Sparse Attention-based neural network for Code Classification (SACC) in this paper. The approach involves two main steps: In the first step, source code undergoes syntax parsing and preprocessing. The generated abstract syntax tree is split into sequences of subtrees and then encoded using a recursive neural network to obtain a high-dimensional representation. This step simultaneously considers both the logical structure and lexical level information contained within the code. In the second step, the encoded sequences of subtrees are fed into a Transformer model that incorporates sparse attention mechanisms for the purpose of classification. This method efficiently reduces the computational cost of the self-attention mechanisms, thus improving the training speed while preserving effectiveness. Our work introduces a carefully designed sparse attention pattern that is specifically designed to meet the unique needs of code classification tasks. This design helps reduce the influence of redundant information and enhances the overall performance of the model. Finally, we also deal with problems in previous related research, which include issues like incomplete classification labels and a small dataset size. We annotated the CodeNet dataset with algorithm-related labeling categories, which contains a significantly large amount of data. Extensive comparative experimental results demonstrate the effectiveness and efficiency of SACC for the code classification tasks.
Tissue phenotyping is a fundamental computational pathology (CPath) task in learning objective characterizations of histopathologic biomarkers in anatomic pathology. However, whole-slide imaging (WSI) poses a complex computer vision problem in which the large-scale image resolutions of WSIs and the enormous diversity of morphological phenotypes preclude large-scale data annotation. Current efforts have proposed using pretrained image encoders with either transfer learning from natural image datasets or self-supervised pretraining on publicly-available histopathology datasets, but have not been extensively developed and evaluated across diverse tissue types at scale. We introduce UNI, a general-purpose self-supervised model for pathology, pretrained using over 100 million tissue patches from over 100,000 diagnostic haematoxylin and eosin-stained WSIs across 20 major tissue types, and evaluated on 33 representative CPath clinical tasks in CPath of varying diagnostic difficulties. In addition to outperforming previous state-of-the-art models, we demonstrate new modeling capabilities in CPath such as resolution-agnostic tissue classification, slide classification using few-shot class prototypes, and disease subtyping generalization in classifying up to 108 cancer types in the OncoTree code classification system. UNI advances unsupervised representation learning at scale in CPath in terms of both pretraining data and downstream evaluation, enabling data-efficient AI models that can generalize and transfer to a gamut of diagnostically-challenging tasks and clinical workflows in anatomic pathology.
As a part of a research project concerning software maintainability assessment in collaboration with the development team, we wanted to explore dissensions between developers and the confounding effect of size. To this end, this study replicated and extended a recent study from Schnappinger et al. with the public part of its dataset and the metrics extracted from the graph-based tool Javanalyser. The entire processing pipeline was automated, from metrics extraction to the training of machine learning models. The study was extended by predicting the continuous maintainability to take account of dissensions. Then, all experimental shots were duplicated to evaluate the overall influence of the class size. In the end, the original study was successfully replicated. Moreover, good performance was achieved on the continuous maintainability prediction. Finally, the class size was not sufficient for fine-grained maintainability prediction. This study shows the necessity to explore the nature of what is measured by code metrics, and is also the first step in the construction of a maintainability model.
Determining industry and product/service codes for a company is an important real-world task and is typically very expensive as it involves manual curation of data about the companies. Building an AI agent that can predict these codes automatically can significantly help reduce costs, and eliminate human biases and errors. However, unavailability of labeled datasets as well as the need for high precision results within the financial domain makes this a challenging problem. In this work, we propose a hierarchical multi-class industry code classifier with a targeted multi-label product/service code classifier leveraging advances in unsupervised representation learning techniques. We demonstrate how a high quality industry and product/service code classification system can be built using extremely limited labeled dataset. We evaluate our approach on a dataset of more than 20,000 companies and achieved a classification accuracy of more than 92\%. Additionally, we also compared our approach with a dataset of 350 manually labeled product/service codes provided by Subject Matter Experts (SMEs) and obtained an accuracy of more than 96\% resulting in real-life adoption within the financial domain.
Semantic understanding of programs has attracted great attention in the community. Inspired by recent successes of large language models (LLMs) in natural language understanding, tremendous progress has been made by treating programming language as another sort of natural language and training LLMs on corpora of program code. However, programs are essentially different from texts after all, in a sense that they are normally heavily structured and syntax-strict. In particular, programs and their basic units (i.e., functions and subroutines) are designed to demonstrate a variety of behaviors and/or provide possible outputs, given different inputs. The relationship between inputs and possible outputs/behaviors represents the functions/subroutines and profiles the program as a whole. Therefore, we propose to incorporate such a relationship into learning, for achieving a deeper semantic understanding of programs. To obtain inputs that are representative enough to trigger the execution of most part of the code, we resort to fuzz testing and propose fuzz tuning to boost the performance of program understanding and code representation learning, given a pre-trained LLM. The effectiveness of the proposed method is verified on two program understanding tasks including code clone detection and code classification, and it outperforms current state-of-the-arts by large margins. Code is available at https://github.com/rabbitjy/FuzzTuning.
The use of modern Natural Language Processing (NLP) techniques has shown to be beneficial for software engineering tasks, such as vulnerability detection and type inference. However, training deep NLP models requires significant computational resources. This paper explores techniques that aim at achieving the best usage of resources and available information in these models. We propose a generic approach, EarlyBIRD, to build composite representations of code from the early layers of a pre-trained transformer model. We empirically investigate the viability of this approach on the CodeBERT model by comparing the performance of 12 strategies for creating composite representations with the standard practice of only using the last encoder layer. Our evaluation on four datasets shows that several early layer combinations yield better performance on defect detection, and some combinations improve multi-class classification. More specifically, we obtain a +2 average improvement of detection accuracy on Devign with only 3 out of 12 layers of CodeBERT and a 3.3x speed-up of fine-tuning. These findings show that early layers can be used to obtain better results using the same resources, as well as to reduce resource usage during fine-tuning and inference.
Code classification is a difficult issue in program understanding and automatic coding. Due to the elusive syntax and complicated semantics in programs, most existing studies use techniques based on abstract syntax tree (AST) and graph neural network (GNN) to create code representations for code classification. These techniques utilize the structure and semantic information of the code, but they only take into account pairwise associations and neglect the high-order correlations that already exist between nodes in the AST, which may result in the loss of code structural information. On the other hand, while a general hypergraph can encode high-order data correlations, it is homogeneous and undirected which will result in a lack of semantic and structural information such as node types, edge types, and directions between child nodes and parent nodes when modeling AST. In this study, we propose to represent AST as a heterogeneous directed hypergraph (HDHG) and process the graph by heterogeneous directed hypergraph neural network (HDHGN) for code classification. Our method improves code understanding and can represent high-order data correlations beyond paired interactions. We assess heterogeneous directed hypergraph neural network (HDHGN) on public datasets of Python and Java programs. Our method outperforms previous AST-based and GNN-based methods, which demonstrates the capability of our model.
The application of deep learning techniques in software engineering becomes increasingly popular. One key problem is developing high-quality and easy-to-use source code representations for code-related tasks. The research community has acquired impressive results in recent years. However, due to the deployment difficulties and performance bottlenecks, seldom these approaches are applied to the industry. In this paper, we present xASTNN, an eXtreme Abstract Syntax Tree (AST)-based Neural Network for source code representation, aiming to push this technique to industrial practice. The proposed xASTNN has three advantages. First, xASTNN is completely based on widely-used ASTs and does not require complicated data pre-processing, making it applicable to various programming languages and practical scenarios. Second, three closely-related designs are proposed to guarantee the effectiveness of xASTNN, including statement subtree sequence for code naturalness, gated recursive unit for syntactical information, and gated recurrent unit for sequential information. Third, a dynamic batching algorithm is introduced to significantly reduce the time complexity of xASTNN. Two code comprehension downstream tasks, code classification and code clone detection, are adopted for evaluation. The results demonstrate that our xASTNN can improve the state-of-the-art while being faster than the baselines.
Recent studies have demonstrated remarkable advancements in source code learning, which applies deep neural networks (DNNs) to tackle various software engineering tasks. Similar to other DNN-based domains, source code learning also requires massive high-quality training data to achieve the success of these applications. Data augmentation, a technique used to produce additional training data, is widely adopted in other domains (e.g. computer vision). However, the existing practice of data augmentation in source code learning is limited to simple syntax-preserved methods, such as code refactoring. In this paper, considering that source code can also be represented as text data, we take an early step to investigate the effectiveness of data augmentation methods originally designed for natural language texts in the context of source code learning. To this end, we focus on code classification tasks and conduct a comprehensive empirical study across four critical code problems and four DNN architectures to assess the effectiveness of 25 data augmentation methods. Our results reveal specific data augmentation methods that yield more accurate and robust models for source code learning. Additionally, we discover that the data augmentation methods remain beneficial even when they slightly break source code syntax.
Graph neural network (GNN)-based graph learning has been popular in natural language and programming language processing, particularly in text and source code classification. Typically, GNNs are constructed by incorporating alternating layers which learn transformations of graph node features, along with graph pooling layers that use graph pooling operators (e.g., Max-pooling) to effectively reduce the number of nodes while preserving the semantic information of the graph. Recently, to enhance GNNs in graph learning tasks, Manifold-Mixup, a data augmentation technique that produces synthetic graph data by linearly mixing a pair of graph data and their labels, has been widely adopted. However, the performance of Manifold-Mixup can be highly affected by graph pooling operators, and there have not been many studies that are dedicated to uncovering such affection. To bridge this gap, we take an early step to explore how graph pooling operators affect the performance of Mixup-based graph learning. To that end, we conduct a comprehensive empirical study by applying Manifold-Mixup to a formal characterization of graph pooling based on 11 graph pooling operations (9 hybrid pooling operators, 2 non-hybrid pooling operators). The experimental results on both natural language datasets (Gossipcop, Politifact) and programming language datasets (JAVA250, Python800) demonstrate that hybrid pooling operators are more effective for Manifold-Mixup than the standard Max-pooling and the state-of-the-art graph multiset transformer (GMT) pooling, in terms of producing more accurate and robust GNN models.
Inspired by the great success of Deep Neural Networks (DNNs) in natural language processing (NLP), DNNs have been increasingly applied in source code analysis and attracted significant attention from the software engineering community. Due to its data-driven nature, a DNN model requires massive and high-quality labeled training data to achieve expert-level performance. Collecting such data is often not hard, but the labeling process is notoriously laborious. The task of DNN-based code analysis even worsens the situation because source code labeling also demands sophisticated expertise. Data augmentation has been a popular approach to supplement training data in domains such as computer vision and NLP. However, existing data augmentation approaches in code analysis adopt simple methods, such as data transformation and adversarial example generation, thus bringing limited performance superiority. In this paper, we propose a data augmentation approach MIXCODE that aims to effectively supplement valid training data, inspired by the recent advance named Mixup in computer vision. Specifically, we first utilize multiple code refactoring methods to generate transformed code that holds consistent labels with the original data. Then, we adapt the Mixup technique to mix the original code with the transformed code to augment the training data. We evaluate MIXCODE on two programming languages (Java and Python), two code tasks (problem classification and bug detection), four benchmark datasets (JAVA250, Python800, CodRep1, and Refactory), and seven model architectures (including two pre-trained models CodeBERT and GraphCodeBERT). Experimental results demonstrate that MIXCODE outperforms the baseline data augmentation approach by up to 6.24% in accuracy and 26.06% in robustness.
Deep learning models have been successfully applied to a variety of software engineering tasks, such as code classification, summarisation, and bug and vulnerability detection. In order to apply deep learning to these tasks, source code needs to be represented in a format that is suitable for input into the deep learning model. Most approaches to representing source code, such as tokens, abstract syntax trees (ASTs), data flow graphs (DFGs), and control flow graphs (CFGs) only focus on the code itself and do not take into account additional context that could be useful for deep learning models. In this paper, we argue that it is beneficial for deep learning models to have access to additional contextual information about the code being analysed. We present preliminary evidence that encoding context from the call hierarchy along with information from the code itself can improve the performance of a state-of-the-art deep learning model for two software engineering tasks. We outline our research agenda for adding further contextual information to source code representations for deep learning.
Distribution shift has been a longstanding challenge for the reliable deployment of deep learning (DL) models due to unexpected accuracy degradation. Although DL has been becoming a driving force for large-scale source code analysis in the big code era, limited progress has been made on distribution shift analysis and benchmarking for source code tasks. To fill this gap, this paper initiates to propose CodeS, a distribution shift benchmark dataset, for source code learning. Specifically, CodeS supports two programming languages (Java and Python) and five shift types (task, programmer, time-stamp, token, and concrete syntax tree). Extensive experiments based on CodeS reveal that 1) out-of-distribution detectors from other domains (e.g., computer vision) do not generalize to source code, 2) all code classification models suffer from distribution shifts, 3) representation-based shifts have a higher impact on the model than others, and 4) pre-trained bimodal models are relatively more resistant to distribution shifts.