Domain-specific Knowledge Research Articles

Knowledge Graph-Based In-Context Learning for Advanced Fault Diagnosis in Sensor Networks

This paper introduces a novel approach for enhancing fault diagnosis in industrial equipment systems through the application of sensor network-driven knowledge graph-based in-context learning (KG-ICL). By focusing on the critical role of sensor data in detecting and isolating faults, we construct a domain-specific knowledge graph (DSKG) that encapsulates expert knowledge relevant to industrial equipment. Utilizing a long-length entity similarity (LES) measure, we retrieve relevant information from the DSKG. Our method leverages large language models (LLMs) to conduct causal analysis on textual data related to equipment faults derived from sensor networks, thereby significantly enhancing the accuracy and efficiency of fault diagnosis. This paper details a series of experiments that validate the effectiveness of the KG-ICL method in accurately diagnosing fault causes and locations of industrial equipment systems. By leveraging LLMs and structured knowledge, our approach offers a robust tool for condition monitoring and fault management, thereby improving the reliability and efficiency of operations in industrial sectors.

Domain Ontology-Driven Knowledge Graph Generation from Text

A knowledge graph serves as a unified and standardized representation for extracting and representing textual information. In the field of knowledge extraction and representation research, named entity recognition and relation extraction provide effective solutions for knowledge graph generation tasks. However, it is a challenge that lies in extracting domain-specific knowledge from the rich and general textual corpora and generating corresponding domain knowledge graphs to support domain-specific reasoning, question-answering, and decision-making tasks. The hierarchical domain knowledge representation model (i.e. domain ontology) provides a solution for this problem. Therefore, we propose an end-to-end approach based on domain ontology embedding and pre-trained language models for domain knowledge graph generation from text, which incorporates domain node recognition and domain relation extraction phases. We evaluated our domain ontology-driven model on the Wikidata-TekGen dataset and the DBpedia-WebNLG dataset, and the results indicate that our approach based on the pre-trained language models with fewer parameters compared with the baseline models has significantly contributed to the domain knowledge graph generation without prompts.

Facial Detection and Recognition Analysis using Ontology-Driven Machine Learning Model

Facial detection and recognition technologies have achieved remarkable advancements through the integration of ontology-driven machine learning (ML). This study examines how ontology structured framework for representing domain-specific knowledge to enhance the robustness, accuracy, and interpretability of ML models, particularly Convolutional Neural Networks (CNNs), in the field of facial detection and recognition. Acknowledging the limitations of traditional ML methods, such as data inefficiency, inadequate generalization, and insufficient interpretability, this research addresses critical challenges, including semantic discrepancies, variations in environmental conditions, and ethical considerations. By leveraging ontology, the study aims to provide semantic enrichment, contextual adaptation, and data augmentation for ML models. A hybrid methodology is introduced that effectively integrates ontology with CNNs, in conjunction with the Viola-Jones and Eigenfaces algorithms, to improve performance in facial recognition tasks. The study utilizes comprehensive datasets such as CelebA and MegaFace, employing rigorous preprocessing and training processes that incorporate ontology-based feedback mechanisms. The results demonstrated significant improvements in accuracy, robustness, and explainability. Ontology-CNN models outperform traditional approaches in managing variations in facial attributes and environmental conditions. This study concludes that ontology-based frameworks present a promising avenue for developing efficient and ethically responsible facial recognition systems. Future research should focus on exploring the scalability of these models across diverse demographic contexts and further addressing ethical considerations related to privacy and fairness.

Generative Interpretable Visual Design: Using Disentanglement for Visual Conjoint Analysis

This article develops a method to automatically discover and quantify human-interpretable visual characteristics directly from product image data. The method is generative and can create new visual designs spanning the space of visual characteristics. It builds on disentanglement methods in deep learning using variational autoencoders, which aim to discover underlying statistically independent and interpretable visual characteristics of an object. The impossibility theorem in the deep learning literature indicates that supervision with ground truth characteristics would be required to obtain unique disentangled representations. However, these are typically unknown in real-world applications, and are in fact exactly the characteristics that need to be discovered. Extant machine learning methods are unsuitable since they require ground truth labels for each visual characteristic. In contrast, this method postulates the use of readily available product characteristics (such as brand and price) as proxy supervisory signals to enable disentanglement. This method discovers and quantifies human-interpretable and statistically independent characteristics without any specific domain knowledge on the product category. It is applied to a dataset of watches to automatically discover interpretable visual product characteristics, obtain consumer preferences over visual designs, and generate new ideal point designs targeted to specific consumer segments.

Weakly supervised learning in thymoma histopathology classification: an interpretable approach

IntroductionThymoma classification is challenging due to its diverse morphology. Accurate classification is crucial for diagnosis, but current methods often struggle with complex tumor subtypes. This study presents an AI-assisted diagnostic model that combines weakly supervised learning with a divide-and-conquer multi-instance learning (MIL) approach to improve classification accuracy and interpretability.MethodsWe applied the model to 222 thymoma slides, simplifying the five-class classification into binary and ternary steps. The model features an attention-based mechanism that generates heatmaps, enabling visual interpretation of decisions. These heatmaps align with clinically validated morphological differences between thymoma subtypes. Additionally, we embedded domain-specific pathological knowledge into the interpretability framework.ResultsThe model achieved a classification AUC of 0.9172. The generated heatmaps accurately reflected the morphological distinctions among thymoma subtypes, as confirmed by pathologists. The model's transparency allows pathologists to visually verify AI decisions, enhancing diagnostic reliability.DiscussionThis model offers a significant advancement in thymoma classification, combining high accuracy with interpretability. By integrating weakly supervised learning, MIL, and attention mechanisms, it provides an interpretable AI framework that is applicable in clinical settings. The model reduces the diagnostic burden on pathologists and has the potential to improve patient outcomes by making AI tools more transparent and clinically relevant.

Tuning Vision-Language Models With Multiple Prototypes Clustering.

Benefiting from advances in large-scale pre-training, foundation models, have demonstrated remarkable capability in the fields of natural language processing, computer vision, among others. However, to achieve expert-level performance in specific applications, such models often need to be fine-tuned with domain-specific knowledge. In this paper, we focus on enabling vision-language models to unleash more potential for visual understanding tasks under few-shot tuning. Specifically, we propose a novel adapter, dubbed as lusterAdapter, which is based on trainable multiple prototypes clustering algorithm, for tuning the CLIP model. It can not only alleviate the concern of catastrophic forgetting of foundation models by introducing anchors to inherit common knowledge, but also improve the utilization efficiency of few annotated samples via bringing in clustering and domain priors, thereby improving the performance of few-shot tuning. We have conducted extensive experiments on 11 common classification benchmarks. The results show our method significantly surpasses the original CLIP and achieves state-of-the-art (SOTA) performance under all benchmarks and settings. For example, under the 16-shot setting, our method exhibits a remarkable improvement over the original CLIP by 19.6%, and also surpasses TIP-Adapter and GraphAdapter by 2.7% and 2.2%, respectively, in terms of average accuracy across the 11 benchmarks.

Foundation model of ECG diagnosis: Diagnostics and explanations of any form and rhythm on ECG.

We propose a knowledge-enhanced electrocardiogram (ECG) diagnosis foundation model (KED) that utilizes large language models to incorporate domain-specific knowledge of ECG signals. This model is trained on 800,000 ECGs from nearly 160,000 unique patients. Despite being trained on single-center data, KED demonstrates exceptional zero-shot diagnosis performance across various regions, including different locales in China, the United States, and other regions. This performance spans across all age groups for various conditions such as morphological abnormalities, rhythm abnormalities, conduction blocks, hypertrophy, myocardial ischemia, and infarction. Moreover, KED exhibits robust performance on diseases it has not encountered during its training. When compared to three experienced cardiologists on real clinical datasets, the model achieves comparable performance in zero-shot diagnosis of seven common clinical ECG types. We concentrate on the zero-shot diagnostic capability and the generalization performance of the proposed ECG foundation model, particularly in the context of external multi-center data and previously unseen disease.

A plug-and-play knowledge-enhanced module for medical reports generation

Medical reports generation from patient–doctor conversations aims to capture the salient contents from dialogues in the form of medical reports, to assist the healthcare archiving and follow-ups. Unlike general summarization tasks, summaries in specialized domains, such as medical report generation, require stronger support from domain-specific knowledge to enhance the internal logic specific to their field. However, existing methods either model the context sequentially without incorporating knowledge graphs, resulting in unclear representation of domain-specific logical structures, or introduce the expert medical knowledge using homogeneous graphs, which fail to effectively fuse information from different semantic spaces. To address this, we propose to incorporate the knowledge into the step of context modeling by introducing a plug-and-play heterogeneous graph encoder. Furthermore, a fact-aware module is introduced to help our model in integrating key context information filtered by external knowledge during decoding. Compared with previous works, our method makes full use of knowledge, and the proposed modules demonstrate easy adaption on existing frameworks. Experimental results on two public medical dialogue summarization datasets indicate that our method significantly outperforms a range of baselines while being smaller and capable of achieving the state-of-the-art performance11Code is available at https://github.com/salvatoreferragamo/Med-HGE..

Pathfinder: A Semantic Framework for Literature Review and Knowledge Discovery in Astronomy

The exponential growth of astronomical literature poses significant challenges for researchers navigating and synthesizing general insights or even domain-specific knowledge. We present pathfinder, a machine learning framework designed to enable literature review and knowledge discovery in astronomy, focusing on semantic searching with natural language instead of syntactic searches with keywords. Utilizing state-of-the-art large language models (LLMs) and a corpus of 385,166 peer-reviewed papers from the Astrophysics Data System, pathfinder offers an innovative approach to scientific inquiry and literature exploration. Our framework couples advanced retrieval techniques with LLM-based synthesis to search astronomical literature by semantic context as a complement to currently existing methods that use keywords or citation graphs. It addresses complexities of jargon, named entities, and temporal aspects through time-based and citation-based weighting schemes. We demonstrate the tool’s versatility through case studies, showcasing its application in various research scenarios. The system’s performance is evaluated using custom benchmarks, including single-paper and multipaper tasks. Beyond literature review, pathfinder offers unique capabilities for reformatting answers in ways that are accessible to various audiences (e.g., in a different language or as simplified text), visualizing research landscapes, and tracking the impact of observatories and methodologies. This tool represents a significant advancement in applying artificial intelligence to astronomical research, aiding researchers at all career stages in navigating modern astronomy literature.

Extracting key insights from earnings call transcript via information-theoretic contrastive learning

Earnings conference calls provide critical insights into a company’s financial health, future outlook, and strategic direction. Traditionally, analysts manually analyze these lengthy transcripts to extract key information, a process that is both time-consuming and prone to bias and error. To address this, text mining tools, particularly extractive summarization, are increasingly being used to automatically extract key insights, aiming to standardize the analysis process and improve efficiency. Extractive summarization automates the selection of the most informative sentences, offering a promising solution for transcript analysis. However, existing extractive summarization techniques face several challenges, such as the lack of labeled training data, difficulties in incorporating domain-specific knowledge, and inefficiencies in handling large-scale datasets. In this work, we introduce ECT-SKIE, an information-theoretic, self-supervised approach for extracting key insights from earnings call transcripts. We leverage variational information bottleneck theory to extract insights in parallel, significantly accelerating the process. In addition, we propose a structure-aware contrastive learning strategy that enables model training without the need for labeled data. We further develop a novel container-based key sentence extractor to alleviate sentence redundancy. Using a large-scale dataset of U.S. market earnings call transcripts, we evaluate our method against nine representative baselines across three downstream tasks. Experimental results show that ECT-SKIE can consistently extract high-quality key sentences. The code is publicly available at: https://github.com/MongoTap/ECT-SKIE.

Human-CoMo: A Combination of Cognitive Knowledge Engineering and Data Modelling for Human-Cyber-Physical Systems in Intelligent Manufacturing

Many cyber-physical systems face the challenge of appropriately integrating domain-specific human expert knowledge into the cyber part to create a shared sphere of knowledge and intelligent interactions between humans and the semi-autonomous technical system. Cognitive engineering contributes methods and insights into higher-order cognition that help to embed human knowledge in an appropriate way. The original research introduces a novel transdisciplinary framework called Human-CoMo, which demonstrates a systematic modelling process, different human perspectives, and the integration of expert knowledge at multiple hierarchical levels. Fundamental principles inspired by human cognition, such as conceptual chunking and knowledge precision, are characterised. Furthermore, it is shown how knowledge hierarchies can be methodically reflected in appropriate data analysis and modelling levels for small and big data applications including artificial intelligence approaches. Combined knowledge- and data-based modelling approaches offer more flexibility to integrate the strengths of humans and technology in a complementary way. The cognitive foundations and their computational reflections are outlined for the technical example process electroplating from the field of materials and surface engineering. The possibilities and limitations of integrating human knowledge through formalisation and implications for future forms of human-machine interaction are discussed.

AI-Driven Crack Detection for Remanufacturing Cylinder Heads Using Deep Learning and Engineering-Informed Data Augmentation

Detecting cracks in cylinder heads traditionally relies on manual inspection, which is time-consuming and susceptible to human error. As an alternative, automated object detection utilizing computer vision and machine learning models has been explored. However, these methods often face challenges due to a lack of sufficiently annotated training data, limited image diversity, and the inherently small size of cracks. Addressing these constraints, this paper introduces a novel automated crack-detection method that enhances data availability through a synthetic data generation technique. Unlike general data augmentation practices, our method involves copying cracks from one location to another, guided by both random and informed engineering decisions about likely crack formations due to cyclic thermomechanical loads. The innovative aspect of our approach lies in the integration of domain-specific engineering knowledge into the synthetic generation process, which substantially improves detection accuracy. We evaluate our method’s effectiveness using two metrics: the F2 score, which emphasizes recall to prioritize detecting all potential cracks, and mean average precision (MAP), a standard measure in object detection. Experimental results demonstrate that, without engineering insights, our method increases the F2 score from 0.40 to 0.65, while maintaining a stable MAP. Incorporating detailed engineering knowledge further enhances the F2 score to 0.70 and improves MAP to 0.57, representing increases of 63% and 43%, respectively. These results confirm that our approach not only mitigates the limitations of traditional data augmentation but also significantly advances the reliability and precision of crack detection in industrial settings.

AECR: Automatic attack technique intelligence extraction based on fine-tuned large language model

Cyber Threat Intelligence (CTI) reports contain resourceful intelligence on cyber-attack campaigns, which provides great help for security analysts to infer attack trends and enhance their defenses. However, due to the diversity of report content and writing styles, current intelligence extraction is mostly based on time-consuming manual efforts. Moreover, existing automatic methods generally neglect the importance of background knowledge and produce inexact extraction results. These problems prevent the effective utilization and sharing of intelligence from CTI reports. In this paper, we primarily focus on the automatic extraction of attack technique (AT) intelligence, which reveals patterns of attack behaviors and hardly changes over time. We propose a novel automatic AT extraction pipeline for CTI reports (AECR). AECR explores the feasibility of extracting AT intelligence based on a fined-tuned large language model (LLM). Particularly, we endow the selected LLM with enhanced domain-specific knowledge to improve its comprehension of AT-relevant content and alleviate the hallucination problem. Experimental results demonstrate that AECR outperforms state-of-the-art methods by a wide margin with a reasonable time cost. Specifically, we improve the accuracy, precision, recall, and F1-score by 108%, 37.2%, 22.4%, and 67.5% respectively. To the best of our knowledge, AECR is the first to perform AT extraction based on fine-tuned LLM.

Evaluation of a novel large language model (LLM)-powered chatbot for oral boards scenarios

Abstract Purpose While previous studies have demonstrated that generative artificial intelligence (AI) can pass medical licensing exams, AI’s role as an examiner in complex, interactive assessments remains unknown. AI-powered chatbots could serve as educational tools to simulate oral examination dialogues. Here, we present initial validity evidence for an AI-powered chatbot designed for general surgery residents to prepare for the American Board of Surgery (ABS) Certifying Exam (CE). Methods We developed a chatbot using GPT-4 to simulate oral board scenarios. Scenarios were completed by general surgery residents from six different institutions. Two experienced surgeons evaluated the chatbot across five domains: inappropriate content, missing content, likelihood of harm, extent of harm, and hallucinations. We measured inter-rater reliability to determine evaluation consistency. Results Seventeen residents completed a total of 20 scenarios. Commonly tested topics included small bowel obstruction (30%), diverticulitis (20%), and breast disease (15%). Based on two independent reviewers, evaluation revealed 11–25% of chatbot simulations had no errors and an additional 11%–35% contained errors of minimal clinical significance. The chatbot limitations included incorrect management advice and critical omissions of information. Conclusions This study demonstrates the potential of an AI-powered chatbot in enhancing surgical education through oral board simulations. Despite challenges in accuracy and safety, the chatbot offers a novel approach to medical education, underscoring the need for further refinement and standardized evaluation frameworks. Incorporating domain-specific knowledge and expert insights is crucial for improving the efficacy of AI tools in medical education.

Trusted commonsense knowledge enhanced depression detection based on three-way decision

Depression detection on social media aims to identify depressive tendencies within textual posts, providing timely intervention by the early detection of mental health issues. In predominant approaches, the Pre-trained Language Models(PLMs) are trained solely on public datasets, falling short of vertical scenarios due to insufficient domain-specific and commonsense knowledge. In addition, ambiguous commonsense knowledge could be misleading to PLMs and results in false judgments. Therefore, it poses significant challenges to select commonsense knowledge that is trusted. To address this, we propose CoKE, a model that incorporates trusted commonsense knowledge based on three-way decision theory to enhance depression detection. CoKE comprises three key modules: trusted screening, knowledge generation, and knowledge fusion. First, we utilize psychiatric clinical scales and three-way decision theory to screen out the uncertain domain from the massive user posts. Then, an adaptive framework is applied to generate and refine trusted commonsense knowledge that can explain the true semantics of posts in the uncertain domain. Finally, a dynamic integration of posts with highly trusted knowledge is achieved through a gating mechanism, resulting in embeddings enhanced by trusted commonsense knowledge that are more effective in determining depressive tendencies. We evaluate our model on two prominent datasets, eRisk2017 and eRisk2018, demonstrating its superiority over previous state-of-the-art baseline models.

A Hybrid Deep Learning Framework for OFDM with Index Modulation Under Uncertain Channel Conditions

Index modulation (IM) is considered a promising approach for fifth-generation wireless systems due to its spectral efficiency and reduced complexity compared to conventional modulation techniques. However, IM faces difficulties in environments with unpredictable channel conditions, particularly in accurately detecting index values and dynamically adjusting index assignments. Deep learning (DL) offers a potential solution by improving detection performance and resilience through the learning of intricate patterns in varying channel conditions. In this paper, we introduce a robust detection method based on a hybrid DL (HDL) model designed specifically for orthogonal frequency-division multiplexing with IM (OFDM-IM) in challenging channel environments. Our proposed HDL detector leverages a one-dimensional convolutional neural network (1D-CNN) for feature extraction, followed by a bidirectional long short-term memory (Bi-LSTM) network to capture temporal dependencies. Before feeding data into the network, the channel matrix and received signals are preprocessed using domain-specific knowledge. We evaluate the bit error rate (BER) performance of the proposed model using different optimizers and equalizers, then compare it with other models. Moreover, we evaluate the throughput and spectral efficiency across varying SNR levels. Simulation results demonstrate that the proposed hybrid detector surpasses traditional and other DL-based detectors in terms of performance, underscoring its effectiveness for OFDM-IM under uncertain channel conditions.

Leveraging the corporate university model to develop and deliver customized course content

Purpose These experiences can provide two significant findings for researchers and other organizations seeking to develop similar training programs. Design/methodology/approach This paper documents how a financial institution developed a customized bank accounting course for its employees through the well-researched and frequently used “Corporate University Model.” It did so through leveraging the educational expertise of a local university professor, and the subject matter expertise of bank managers. The organization followed up on this success by developing a customized risk management course. As before, the bank utilized the Corporate University Model. However, the content was more specialized than that of the previous course since it related to the unique operations of the organization. Therefore, the bank found it necessary to make some adjustments to its development and delivery approach. Findings First, even though corporate universities are often depicted as complex separate entities within companies, the Corporate University Model can be successfully utilized to develop and deliver individual courses that are managed by individual departments. Second, the bank’s risk management course demonstrated how a professional instructor could be used in conjunction with a company’s subject matter experts to effectively deliver course content. This approach is very effective with course content that is unique to the organization’s specific operations. Research limitations/implications While literature indicates that corporate universities often involve the creation of a separate entity, which can be difficult to configure due to their complexity (Lui Abel and Li, 2012), the leveraging of the CU model to construct and deliver single courses has not been explored. Additionally, the existing models do not discuss how roles outlined in the CU model (Patrucco et al., 2017) may change due to content complexity. From a limitation standpoint, it would be helpful to apply this approach to additional courses and organizations. Additionally, data could be collected to evaluate course effectiveness by learning goal. Practical implications This case study shows that the CU model has value for individual courses that continue to reside and be managed in the departments that developed them. Additionally, it shows that even though a professional instructor may not have all requisite knowledge in more specific knowledge domains, this can be solved by involving SMEs in content delivery. This still allows the benefit of a professional instructor who is skilled in course delivery and development. Originality/value These findings provide insights that have not been described in other papers that detail the application of the Corporate University Model. This helps to expand the manner in which companies can attain its benefits.

OptiFeat: enhancing feature selection, a hybrid approach combining subject matter expertise and recursive feature elimination method

Optimizing the performance of Java Virtual Machines (JVMs) (Sahin et al. in Proc IEEE Int Congr Big Data BigData Congr 410–417, 2016) is crucial for achieving efficient execution of Java applications. Feature selection plays a pivotal role in identifying the most relevant parameters for fine-tuning JVMs, thereby enhancing their overall efficiency. This paper presents a novel hybrid approach that integrates both subject matter expertise and Recursive Feature Elimination (RFE) (Yin et al. in J Big Data 10(1):15, 2023) model to refine feature selection for JVM fine-tuning using machine learning techniques. Traditional feature selection methods often lack the ability to incorporate domain-specific knowledge, resulting in suboptimal selections (Khaire and Dhanalakshmi in J King Saud Univ Comput Inf Sci 34(4):1060–1073, 2022). In contrast, the hybrid approach leverages the expertise of JVM administrators or developers to guide the feature selection process. By integrating domain knowledge into the feature selection pipeline, ensure the inclusion of crucial JVM parameters that may not be captured by automated techniques alone. Furthermore, employed the RFE model, a powerful recursive feature elimination algorithm, to iteratively identify and eliminate irrelevant features from the initial feature set. This iterative process enhances the efficiency of feature selection by systematically pruning less influential parameters, thereby improving the overall performance of the JVM. To validate the effectiveness of the hybrid approach, conducted experiments using real-world JVM datasets and compare the performance of the method against existing feature selection techniques. The results demonstrate that the approach not only achieves superior performance in terms of JVM fine-tuning but also provides insights into the significance of domain expertise in optimizing JVM performance (Menéndez and Bartlett in http://arxiv.org/abs/2310.16510, 2023). It contributes to the field of JVM optimization by proposing a novel hybrid approach that combines subject matter expertise with machine learning-based feature selection techniques. By leveraging both domain knowledge and automated algorithms, the approach offers a comprehensive solution for enhancing feature selection in JVM fine-tuning, ultimately leading to improved performance and efficiency in Java application execution.

Improving the Classification of Defect Levels in Main Electrical Equipment Through Advanced Prompting on a Large Language Model

The stable operation of the power system is closely related to the national economy and people’s livelihoods. Therefore, the timely detection, qualitative assessment, and handling of major equipment defects are crucial. The classification of defect levels in main electrical equipment is a fundamental task in this process, which is often manually completed, supplemented by knowledge bases or expert systems. However, this approach is time-consuming, labor-intensive, involves challenging human–machine interaction, and relies on expert experience. Conversational large language models, such as ChatGPT, ERNIE Bot, ChatGLM, have garnered widespread recognition in various domains. However, these models may have errors in the reasoning process, resulting in biased or even erroneous outputs, which is referred to as the “hallucinations”. The hallucinations’ problem of large language models poses challenges in specific fields. To mitigate the hallucinations in large language models, researchers often seek to incorporate domain-specific knowledge into these models through methods like fine-tuning or prompt learning. In order to enhance model performance while minimizing computational costs, this study adopts the prompt learning approach. Specifically, we propose a large language model prompt learning framework based on knowledge graphs, aiming to provide the large language model with reasoning support by leveraging specific information stored within the knowledge graph and receive explainable reasoning result. Experimental results demonstrate that our module achieved superior results on the power defect dataset compared to the non-prompt method.

Diverse paths, stable differences – role of prior knowledge in learning biology over undergraduate years

ABSTRACT Undergraduate students enrol at university with diverse levels of prior knowledge and intuitive conceptions about essential biological topics. This longitudinal study explored the development of a conceptual understanding about three topics and how domain-specific prior knowledge affects this development. Undergraduate students (N = 50) of biological and environmental sciences participated in three measurement points – baseline (2019), follow-up 1 (2020), and follow-up 2 (2021) – where the same questionnaire with open-ended tasks was given. A mixed-methods approach was used for quantitative scoring of the answers and qualitative thematic analysis to describe the development of individual students’ conceptual understanding. The effect of prior knowledge was studied by dividing the students into two groups based on the baseline performance. The topic of evolution showed the most significant differences in prior knowledge and in knowledge gains. Students with lower prior knowledge experienced desirable shifts in their explanatory models about evolutionary processes, but an equal amount exhibited fragmented learning trajectories. Students with a more robust prior understanding of such an emergent phenomenon can more likely integrate scientific concepts in their knowledge framework. Implications for teaching in higher education are discussed.