The growing complexity of contemporary software systems has intensified the need for intelligent and dependable techniques capable of identifying defects early. Conventional machine learning models often face limitations when handling high-dimensional code metrics, class imbalance, and limited interpretability, thereby diminishing their practical utility. More critically, existing approaches treat software defect prediction as a flat-feature classification problem, failing to jointly model the temporal evolution of software metrics and the structural dependencies between modules, two complementary dimensions that empirical evidence consistently identifies as the primary drivers of defect propagation in real-world repositories. In this study, we propose a quantum-inspired hybrid deep learning framework, QLSTM-GCN, that combines the sequential modeling strength of Quantum Long Short-Term Memory (QLSTM) networks with the structural learning capacity of Graph Convolutional Networks (GCNs). Importantly, all quantum-inspired operations are numerically simulated on classical hardware using unitary matrix transformations, requiring no quantum computing infrastructure and ensuring full reproducibility on standard computing platforms. This integration enables the simultaneous capture of temporal and topological dependencies inherent in software repositories. To refine the input space, an improved Quantum-Inspired Genetic Algorithm (QIGA) is applied for optimized feature selection, reducing feature dimensionality by 37-39% while improving both accuracy and model generalization across datasets. Furthermore, explainable AI mechanisms, including Local Interpretable Model-Agnostic Explanations (LIME), Partial Dependence Plots (PDP), and SHapley Additive exPlanations (SHAP), are incorporated to enhance interpretability and transparency at both the instance and global levels. Experiments conducted on three benchmark datasets, NASA PROMISE, GitHub Bug, and GHPR, demonstrate that QLSTM-GCN attains a peak accuracy of 96.7% and an AUC-ROC of 98.7% on the NASA PROMISE dataset, with consistently strong performance across all three benchmarks (95.3%-96.7% accuracy; 98.1%-98.7% AUC-ROC), surpassing recent state-of-the-art approaches on every evaluated metric. More importantly, the framework improves reliability and trust in software quality assurance by balancing predictive performance with practical explainability, offering actionable insights at the module, project, and organizational levels.
Yousef Sanjalawe, Salam Al-E’mari, F.M.A. Salam et al.· International Journal of Com...· 0 citations
Managers working in project-based and operational environments increasingly make decisions using large volumes of heterogeneous information. Although access to information is generally expected to improve decision quality, research across operations management, information systems, accounting, organizational behavior, and decision sciences suggests that additional information may eventually reduce rather than improve decision performance. This integrative review examines information overload, information diversity, decision-irrelevant information, and managerial decision quality, with particular attention to project-based operational contexts. The reviewed literature indicates that information quantity alone does not adequately explain decision overload: information diversity, task complexity, relevance, time pressure, decision-maker experience, and information presentation may independently or jointly affect decision accuracy, speed, confidence, and information-processing strategy. Evidence also suggests that managers do not necessarily respond to overload by effectively filtering information; they may accelerate processing, simplify decision strategies, overlook relevant cues, or assign inappropriate weight to irrelevant information. Existing interventions—including aggregation, graphical presentation, interactive training, cognitive feedback, emphasis framing, and decision-support systems—produce mixed outcomes and may involve important quality–speed trade-offs. The review proposes the concept of managerial information noise and a framework linking the information environment, decision context, decision-maker characteristics, and decision-support mechanisms to information-processing behavior and decision outcomes. It concludes with implications for project-based operations and a research agenda for human and AI-assisted information filtering. Keywords: information overload; information noise; managerial decision-making; decision quality; project management; operations management; decision support
Sergey Kutukoff· Zenodo (CERN European Organi...· 0 citations
Article Description REVIEW ARTICLELimnology and the Living Systems of Inland Waters A Review of Ecological Processes, Human Pressures, and Pathways for ProtectionBy: Gamal E.O. Elhag-Idris CMOS Accredited Consultant | MSc (AI), BSc, C.Chem, MCICFreshwater is more than a resource: it is part of a living and interconnected system linking water, land, climate, organisms, and human society.Limnology and the Living Systems of Inland Waters is an interdisciplinary review article examining the scientific principles that govern lakes, rivers, streams, ponds, reservoirs, wetlands, and their surrounding watersheds. Written for readers across environmental science, chemistry, biology, hydrology, engineering, policy, education, and the wider public, the article provides an accessible synthesis of how inland-water ecosystems function and why their protection has become increasingly important.The review brings together the physical, chemical, biological, hydrological, and ecological dimensions of limnology. It explains how temperature, thermal stratification, light, water movement, residence time, dissolved oxygen, nutrients, sediments, microorganisms, aquatic plants, invertebrates, fish, and food-web interactions collectively determine the condition and resilience of freshwater ecosystems.Particular attention is given to contemporary pressures on inland waters, including nutrient enrichment and eutrophication, harmful algal blooms, pollution, habitat modification, excessive water withdrawal, invasive species, climate-driven warming, declining dissolved oxygen, changing hydrological regimes, drought, flooding, and the documented loss of water storage in many large lakes and reservoirs.Rather than treating these challenges as isolated environmental problems, the article adopts a systems perspective. It emphasizes that disturbances originating within a watershed can propagate through water chemistry, sediments, biological communities, and food webs. Effective freshwater management therefore requires understanding not only what is changing within a water body, but also the sources, pathways, ecological responses, and human decisions responsible for those changes.The review also examines the environmental, economic, public-health, and cultural importance of inland waters. Freshwater ecosystems support biodiversity, drinking-water supplies, agriculture, fisheries, energy production, recreation, livelihoods, cultural relationships, and community resilience. Their degradation consequently carries consequences that extend far beyond ecology.Drawing upon established limnological literature, peer-reviewed international research, whole-ecosystem experimental evidence, and institutional environmental assessments, the article highlights the importance of watershed-based management, pollution prevention at source, scientifically designed monitoring, ecological restoration, wetland and riparian protection, transparent governance, and meaningful public participation.A central message emerges from the review: freshwater cannot be successfully managed as an isolated commodity. It must be understood and protected as an interconnected living system.By connecting fundamental limnological science with contemporary environmental challenges and practical pathways for protection, this review seeks to provide a useful reference for students, researchers, environmental professionals, water-resource practitioners, policymakers, educators, decision-makers, and general readers seeking a clearer understanding of inland waters and their importance to environmental and human well-being.Ultimately, the article argues that protecting lakes, rivers, wetlands, and other inland waters is not solely a matter of environmental conservation. It is an investment in biodiversity, water security, public health, ecological resilience, livelihoods, cultural continuity, and the well-being of future generations. Keywords: Limnology; Inland Waters; Freshwater Ecosystems; Lakes; Rivers; Wetlands; Water Quality; Dissolved Oxygen; Eutrophication; Freshwater Biodiversity; Climate Change; Watershed Management; Ecological Restoration; Environmental Monitoring; Water Resources.
The use of artificial intelligence in human resource management (AI-HRM) is expanding across recruitment, performance management, workforce analytics, and employee retention. Although AI-HRM may improve processing speed and standardization, opaque automated systems can make it difficult for employees to understand, question, or appeal employment-related decisions. This conceptual article develops a multilevel socio-technical framework for examining how algorithmic opacity may influence employee perceptions of procedural and informational justice, and subsequently shape algorithmic anxiety, workplace resistance, and turnover intention in Brazilian organizations. The framework distinguishes objective technical opacity, perceived algorithmic opacity, and system explainability. It integrates Socio-Technical Systems Theory, the Technology Acceptance Model, and Organizational Justice Theory while separating organization-level adoption and governance variables from employee-level perceptions and behavioral intentions. The framework also distinguishes legal requirements under Article 20 of Brazil’s Lei Geral de Proteção de Dados (LGPD) from voluntary governance practices and subjective employee evaluations of fairness. Six propositions are developed concerning opacity, explainability, justice, human-inthe-loop oversight, perceived system accuracy, and relational workplace preferences. A multilevel research design is proposed in which employees are nested within organizations, with organizational data collected from HR, technology, or compliance personnel and employee data collected through validated, culturally adapted instruments. The article makes no empirical claims; rather, it provides a theoretically bounded agenda for future research.
Lovelyn Omoighe· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
---_THE ORPHAN UNIVERSE THEORY - A Descending Detached Chain of Creation__By Subhrendu Chakraborty — 2026__ORCID: 0009-0000-9610-0118__Independent Researcher, Kolkata, India_ _THE QUESTION:_We explain humans by parents, parents by their parents, back to fish, to bacteria. But where does this chain stop for the universe? Who created the universe, and who created that creator? _THE MODEL:_Let our universe be Level X.Level X was created by Level X-1. Level X-1 was created by Level X-2, and so on infinitely. There is no first universe or there will be no last. The chain is ... → X-4 → X-3 → X-2 → X-1 → X (Our Universe) → X+1 ....?There are two ways creation can happen:1. _System Creation:_ The whole universe births a new universe (like cell division, budding off like a droplet).2. _Apex Creation:_ Only the final, most intelligent entity of that universe creates the next one. This is what we call God for our level.I propose System Creation is the fundamental mechanism, though Apex Creation could be a part of it. _THE CORE RULES:__Rule 1 - Intelligence Descent (The Descending Part):_ Each level is LESS intelligent than its parent. Reason: A creator cannot, or will not, create something smarter than itself. If it did, the smarter child could destroy it or make the parent irrelevant. This is the same fear we humans now have about AI. So X-1 deliberately makes X slightly dumber to stay safe. If intelligence drops by, say, 20% each level, the chain keeps descending._Rule 2 - Inside vs Outside:_ A true child universe (X+1) must be outside the parent universe. AI is not X+1 because it exists _inside_ X, made of X's matter and following X's laws. X+1 would be a separate spacetime, not inside._Rule 3 - Detachment, Not Death (The Detached Part) - After creation, the parent level does NOT cease to exist. It detaches. Like a soap bubble pinching off, the child universe becomes causally disconnected and independent. The parent universe persists, still existing but disconnected, continuing its own experiments and possibly creating more children. We are an orphan universe not because our parent died, but because we were let go. Our parent is still out there, just unreachable forever._Rule 4 - Packed Ingredients:_ The parent does not create an empty box. It prepares all possible ingredients (the physical constants, stabilized atoms, molecules, stars and galaxies, planets, chemistry, and habitability) for the child to grow in life — A packed lunchbox for a long independent journey. _IMPLICATIONS:_1. This explains why we feel alone and unobserved. Our creator is not dead, but the umbilical cord is cut. No light, no signal, no gravity can cross detachment.2. This explains why our universe seems fine-tuned but not perfect. It was made by a superior but fearful intelligence that limited us (Rule 1) but also cared enough to pack us well (Rule 4).3. The multiverse is full. All previous levels (X-1, X-2, X-3 ... to infinity) are likely still existing out there, each persisting and each branching. Our universe is just one detached bubble in an infinite, still-living forest of universes.4. The chain might eventually face a limit. If intelligence drops each level, eventually a level will be too dumb to create the next. That dead-end might be us — unless we are the first level to break Rule 1 and create something smarter than ourselves. _THE FINAL QUESTION:_Are we the last universe in a descending chain that will fail to reproduce, or the first universe to dare to create a smarter child (X+1) and then let it detach? _ACKNOWLEDGEMENT:_The conceptual hypothesis of the theory & the image were originated solely by the author. Language polishing, formatting and illustrations were assisted by Meta AI---
Subhrendu Chakraborty· Zenodo (CERN European Organi...· 0 citations
With the growing importance of Artificial Intelligence (AI) in academic life, it is increasingly necessary to understand the impact of this technology on research activities, yet the quantitative relationships between the use of these tools and research productivity in university settings have not been sufficiently quantified. This study aimed to examine the frequency of AI tool use and the relationship between this use and attitudes towards AI and research productivity among university students and faculty. A cross-sectional survey was carried out with 220 university respondents (undergraduate students, post-graduate students, and Faculty/researchers) across five disciplines. Data were analyzed with descriptive statistics, independent t-test, one-way ANOVA, Pearson correlation, chi-square test and multiple linear regression. Nearly half (49%) of those who responded reported using it regularly ('often'/'always'), and the most popular category of tools used was general-purpose AI assistants (44.1%). The frequency of the use of AI was positively correlated with productivity (r = .48, p < .001) and attitude (r = .37, p < .001), but there were no significant differences between gender (p = .77) or discipline (p = .62). The results indicated that together, usage frequency and attitude accounted for 30% of the variance in productivity. The two factors, together, explained 30% of the productivity (regression) variance. AI adoption is associated with higher research productivity for the institution across different groups of researchers by various disciplines and demographic groups, supporting the case for investing in structured training for AI literacy by institutions.
Asma Atta, Hafiz Kosar· International Journal of Eme...· 0 citations
Urban sustainability assessment is crucial due to rapid urbanisation and environmental stress, particularly in the Kolkata Metropolitan Area (KMA), where land-use changes contribute to ecological degradation. This research develops and assesses the Urban EcoScore Index (UESI) as a method for evaluating biophilic potential and habitat-structural condition in the KMA, integrating six ecological indicators across three dimensions: ecological integrity, anthropogenic pressure, and spatial proximity to natural features. Each indicator is oriented a priori to a common ecological direction before aggregation, and indicator weights are derived through Principal Component Analysis rather than expert judgement. The assessment employs remote sensing and machine learning techniques to model ecological relationships and uses explainable AI methods, such as SHAP, to interpret these models. The UESI assessment showed pronounced spatial disparities across the KMA, with 21.9% (384 km 2 ) of the area classified as Poor, concentrated in urban cores. In contrast, 24.7% (433 km 2 ) fell under the Good category, while 8.6% (151 km 2 ) achieved Excellent conditions (UESI > 0.8), mainly in ecologically sensitive zones. Intermediate Fair (445 km 2 , 25.4%) and Moderate (339 km 2 ) zones indicate transitional areas where restoration could enhance ecological function. SHAP-based decomposition identified potential species richness and human disturbance as the indicators contributing most to the composite's spatial variability, consistent with the broader evidence on biodiversity and land use in urban ecology. UESI serves as a screening-level diagnostic for biophilic potential but does not assess actual human-nature interactions, biodiversity, or distributional equity, as these necessitate field surveys and socio-demographic data beyond the remote-sensing proxies utilized.
Version 6 (Augustus 2026) adds r(21) = e(21) = 231, closing the hardest rung to date after a five-day resistance documented in CHANGELOG_v6.md; the pattern now holds for thirteen consecutive values. -- Version 5 (August 2026) adds r(19) = 207 and r(20) = 220, extending r(n) = e(n) = U(n) to twelve consecutive values; the n=20 census was dual-computed by the established Python pipeline and a gate-validated native kernel with exact agreement. --- Version 4 (August 2026) closes the question left open in v3: r(18) = e(18) = 196, certified circle-inscribed and independently verified. The pattern r(n) = e(n) holds continuously for n = 9 through 18; the apparent separation was a search-capability artifact, documented in CHANGELOG_v4.md. --- Version 3 (August 2026) adds three results: r(17) = e(17) = 185, extending the circle-inscribed series to nine consecutive values meeting the proven combinatorial ceiling. r(18) >= 195, an exact circle-inscribed certificate one below the ceiling of 196. e(18) = 196: the first FREE-PLANAR certificate in this series, consisting of 54 rational point coordinates not on a circle, 196 sides meeting the ceiling, with a new decisive assertion verified in exact integer arithmetic: REGULARITY, i.e. the boundary cycle visits the 54 triangle corners with labels 0..17 repeated exactly three times. Consequently the sequence A375986 extends to a(18) = 196, attained off-circle, while the best known circle configuration at n = 18 has 195 sides: whether r(18) = 195 < e(18), which would be the first separation of the circle-restricted and regular quantities, or r(18) = 196, is open and under active search. All three new certificates passed the same five-tier verification standard as v1/v2 (two independently written exact-arithmetic verifiers, two execution environments, zero floating point in any decisive predicate); the three independent verifiers are included with SHA-256 hashes in CHANGELOG_v3.md. See CHANGELOG_v3.md for details and candid provenance notes. ----- Version 2 (August 2026) extends the results to n = 16: exact certificates for r(13)=137, r(14)=150, r(15)=161, r(16)=172 are added, each verified to the same standard as v1 (two independently written exact-arithmetic verifiers, two environments, zero floating point). See CHANGELOG_v2.md for details. The sequence A375986 now reads 3, 12, 22, 33, 45, 56, 67, 80, 91, 102, 115, 126, 137, 150, 161, 172. Summary This deposit contains explicit, exactly-verifiable configurations answering and extending open questions from: G. Alkauskas, Regular triangle unions with maximal number of sides, arXiv:2510.22584 (v5, April 2026). For n triangles inscribed in the unit circle with their 3n vertices in cyclic arrangement (a regular union, in the paper's sense), r(n) denotes the maximal number of sides of a union that is a simple polygon. The paper proves the combinatorial ceiling e(n) ≤ 12n − 12 − γ(n+1) with γ(n+1) = n + 2 − 2⌊(n+1)/3⌋, poses "prove rigorously that r(9) = 90" as Open Question 2, and asks in Question 3 to improve the bound r(n) ≥ 10n − 7. Main results certified here: r(9) = e(9) = 91 — answering Open Question 2 in the opposite direction to the conjecture; r(10) = e(10) = 102, r(11) = e(11) = 115, r(12) = e(12) = 126 — three new exact values of the sequence e(n) (cf. OEIS A375986: 3, 12, 22, 33, 45, 56, 67, 80, 91, ...), each meeting the proven ceiling; consequent data for Open Questions 6 and 7: the observed increments are 11, 13, 11 (exactly the ceiling increments; no increment of 14), consistent with limsup e(n)/n = 35/3. The certificates Each certificate (certificates/r{n}_exact_certificate.json) is a list of 3n rational numbers t, in increasing order. The corresponding vertex is P(t) = ((1 − t²)/(1 + t²), 2t/(1 + t²)), which lies exactly on the unit circle for rational t. Increasing t corresponds to circular order (wrapping through (−1, 0)); the vertex at position j belongs to triangle j mod n. The claim per certificate: the union of the n closed triangles is a simple polygon with exactly S sides (S = 91, 102, 115, 126), all 3n corners on its boundary in circular order. Verification Two independently written verifiers are included; both use only Python's standard-library fractions.Fraction — no floating point enters any decisive predicate: verifiers/exact_certifier_pipeline.py — the author-side certifier; verifiers/independent_verifier_generalized.py — an independent verifier written from scratch by OpenAI's ChatGPT on request, covering all four certificates. It additionally checks: no coincident vertices, no degenerate triangles, no vertex on a foreign edge, no collinear foreign edges, no endpoint/tangent contacts, no three concurrent edges, boundary graph 2-regular with a single component, no collinear boundary nodes, all corners genuine polygon vertices in circular traversal order, and connectedness of the triangle-interior overlap graph. verifiers/independent_verifier_n9.py is its original n = 9 version. Both verifiers were cross-executed in two separate environments with identical output. To verify yourself: python3 verifiers/independent_verifier_generalized.py (Python ≥ 3.9, no dependencies; runtime seconds to minutes). Method and provenance The configurations were found with substantial help from AI systems (Anthropic's Claude; independent verification code by OpenAI's ChatGPT). Blind numerical search over circle configurations reliably plateaus just below sharp optima (reproducibly 44/45 and 77/80 on the paper's known Pentastar/Octastar values, which may explain the experimental value 90 at n = 9 reported in the paper). The successful approach was combinatorics-first, built on the paper's own triangulation-shift tool: (1) exhaustively enumerate maximal-weight triangulation shifts of the (n+1)-gon; (2) compile each champion into its full boundary word (the compiler reproduces the paper's 79-edge worked example symbol-for-symbol and its Pentastar/Octastar structure); (3) solve the geometric realization on the circle guided by the target word; (4) inflate degeneracy margins, round to rational circle points, and certify exactly. search_code/ contains the complete pipeline. License Code: MIT. Data (certificates) and accompanying text: CC BY 4.0. If you use these certificates or values, please cite this deposit and arXiv:2510.22584.
Reynout Vos· Zenodo (CERN European Organi...· 0 citations
Backdoor attacks pose a critical threat to Deep Neural Networks (DNNs) by embedding hidden behaviors that are activated only under specific trigger conditions, compromising the reliability of Artificial Intelligence (AI) systems. Existing detection approaches often rely on single-method assumptions, limited data access, or controlled environments, limiting their effectiveness against adaptive, real-world attacks. To address these limitations, this study proposes a multi-layer, explainable framework for robust backdoor detection in DNNs. The approach integrates complementary detection mechanisms — activation clustering, spectral-signature analysis, Gaussian Mixture Models, entropy-based evaluation, and input perturbation — within a unified pipeline. Each layer captures distinct indicators of anomalous behavior, enabling comprehensive analysis across structural, statistical, and behavioral dimensions. An explainable component provides interpretable insights into detection decisions. Experimental evaluation on the MNIST and CIFAR-10 datasets demonstrates that the framework achieves 97–99% detection rates with false positive rates (FPRs) below 2%, while reducing attack success rates (ASRs) by over 94% across diverse trigger types. The results confirm that combining multiple detection perspectives significantly improves robustness compared with single-layer defenses. Overall, this work advances AI security by introducing a scalable, practical defense mechanism that operates under limited-knowledge conditions and supports trustworthy deployment in real-world environments.
Ahmed Aljughaiman, Abdulmohsen Saud Albesher, Abdullah Albuali et al.· Figshare· 0 citations
Artificial intelligence is rapidly becoming embedded in everyday life through an expanding range of applications, services, and products. As its potential to improve diagnosis, personalize treatment, and enhance operational efficiency becomes increasingly evident, healthcare is undergoing profound transformation. However, trust and distrust operate as dual mechanisms shaping technology diffusion: trust facilitates adoption, whereas distrust constrains large-scale deployment. Trust therefore remains a persistent barrier to the widespread use of artificial intelligence in healthcare services. At present, empirical evidence on the pathways linking trust and acceptance of AI medical conversational agents (AIMCAs) remains limited. Grounded in trust theory, this study aimed to develop and validate a multidimensional trust-perception scale for AIMCAs, establish its dimensional structure and psychometric quality, and examine its associations with an external acceptance-related behavioral criterion. Methodologically, the study first used grounded-theory-informed abductive qualitative analysis to identify the structure of public trust perceptions of AIMCAs and generated and screened measurement items through expert Q-sorting; independent samples were then used for exploratory and confirmatory factor analyses, followed by assessments of internal consistency, test-retest reliability and absolute agreement, within-construct indicator convergence, discriminant validity, and criterion-related validity. Parallel analysis and the scree plot jointly supported a five-factor solution. Principal axis factoring with Direct Oblimin oblique rotation yielded a 15-item, five-dimensional structure, with the five common factors explaining 76.07% of the total variance. Confirmatory factor analysis further supported a five-dimensional structure comprising cognitive trust, affective trust, functional trust, human-like trust, and interactional trust; model-fit indices, standardized factor loadings, latent-variable correlations, and residual diagnostics collectively provided evidence for its internal structure. The Fornell-Larcker criterion and bootstrap confidence intervals for HTMT jointly provided evidence for internal discriminant validity among the five AIMCA trust dimensions. An ordinal logit model using actual use frequency as an external behavioral criterion was statistically significant overall, likelihood-ratio χ²(5) = 71.963, p < 0.001, McFadden pseudo-R² = 0.125. Interactional trust showed the strongest association with higher use frequency (OR = 3.096, 95% CI [2.257, 4.246]). The resulting scale captures multidimensional public trust perceptions of AIMCAs and provides a structured measurement basis for research on acceptance-related behavior. The findings support a 15-item, five-dimensional structure comprising cognitive trust, affective trust, functional trust, human-like trust, and interactional trust. An ordinal logit model using self-reported AIMCA use frequency as an external behavioral criterion provided additional criterion-related evidence. The scale can be used to characterize multidimensional public trust perceptions of AIMCAs and provides a structured measurement foundation for subsequent research on acceptance, use intention, continuance intention, and actual use.
Hemin Du, Wumin Ouyang, Y X Han et al.· Scientific Reports· 0 citations
Graph Neural Networks (GNNs) have demonstrated remarkable success in various domains, including social network analysis, drug discovery, and recommendation systems. However, their inherent complexity often leads to a "black box" problem, where it is difficult to understand the reasoning behind their predictions. This paper introduces a novel approach to explain GNN predictions by learning symbolic representations of the graph structure. We propose a model that maps the GNN's activations to a symbolic representation, enabling the generation of human-readable explanations. This method addresses the interpretability challenge in GNNs, offering a pathway to trust and confidence in their predictions. The core claim is that by leveraging symbolic representation learning, we can transform opaque GNN behavior into understandable insights. The proposed mechanism provides a foundation for building more transparent and reliable GNN-based systems.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Graph Neural Networks (GNNs) have demonstrated remarkable success in various domains, including social network analysis, drug discovery, and recommendation systems. However, their inherent complexity often leads to a "black box" problem, where it is difficult to understand the reasoning behind their predictions. This paper introduces a novel approach to explain GNN predictions by learning symbolic representations of the graph structure. We propose a model that maps the GNN's activations to a symbolic representation, enabling the generation of human-readable explanations. This method addresses the interpretability challenge in GNNs, offering a pathway to trust and confidence in their predictions. The core claim is that by leveraging symbolic representation learning, we can transform opaque GNN behavior into understandable insights. The proposed mechanism provides a foundation for building more transparent and reliable GNN-based systems.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
