Theory gaps in Computer Science

Representative open questions

Showing 30 of 132 — one per source paper, highest-quality first.

• AN ITERATIVE GLMM–XGBOOST ALGORITHM WITH GROUP-AWARE CONDITIONAL PERMUTATION IMPORTANCE FOR EXPLAINING MULTILEVEL ITEM RESPONSE DATA (2026) · doi

  The paper introduces group-aware conditional permutation importance for explaining GLMM–XGBoost predictions but does not provide guidance on interpreting interaction effects identified through this approach when predictors operate at different hierarchical levels (person-, cluster-, item-level), particularly in cross-level interactions.

• Securing IoT Devices with PUFs: Mitigating Aging and Tampering through Cryptography and Machine Learning (2026) · doi

  The serial test frequency statistics νi1…im are noted as non-independent, yet the paper does not provide guidance on correcting bias or variance inflation that may occur when applying this test to PUF output sequences in aging or tampered IoT device scenarios.

• Blockchain-integrated machine learning framework for transparent smart contract vulnerability detection (2026) · doi

  Random Forest demonstrated inherent robustness with nearly identical performance before and after SHAP application, but the specific ensemble-based averaging mechanisms that provide this interpretability-independent robustness on smart contract vulnerability detection have not been formally analyzed. This requires theoretical investigation of why tree ensemble averaging outperforms gradient boosting for unbalanced vulnerability classification.

• Inferring High-Dimensional Dynamic Networks Changing with Multiple Covariates (2026) · doi

  The Hamming distance analysis reveals that radiation doses of 0 Gy and 0.05 Gy produce nearly identical network structures with only marginal differences attributed to random variation, but no threshold or statistical test is established to distinguish true dose-dependent network changes from stochastic fluctuations in CVN estimation.

• Large language model based machine translation for universal multilingual understanding and translation quality enhancement (2026) · doi

  The comparative analysis shows GPT-4 outperforms other models consistently across language pairs (En-De, En-Cs, En-Zh, En-Ru, De-En, Cs-En, Zh-En, Ru-En), but does not investigate the specific linguistic phenomena or grammatical structures where GPT-4 advantages emerge, particularly for morphologically complex languages like Russian and Czech.

• Quantum Information Framework for Neural Network Generalization: A Comprehensive Experimental Analysis (2026) · doi

  The framework evaluates generalization on test sets but does not investigate the relationship between quantum information metrics (von Neumann entropy, effective rank) and overfitting detection, or whether these metrics can predict generalization gaps before evaluation on held-out test data.

• Comparative analysis of deep learning algorithms for rolling element bearing fault classification under variable loads and speeds (2026) · doi

  A discrepancy was observed between high AUC-ROC values (0.98-1.00 for all fault classes) and lower overall classification accuracy (0.928), indicating a gap between class separability in feature space and final label assignment. Future work should investigate whether this gap results from suboptimal decision thresholds, class boundary definition, or feature representation issues specific to rolling element bearing fault detection networks.

• LLM-Powered Silent Bug Fuzzing in Deep Learning Libraries via Versatile and Controlled Bug Transfer (2026) · doi

  The failure case analysis reveals that TransFuzz cannot distinguish compiler-level optimization artifacts from genuine functional bugs in gradient computation, as demonstrated by the torch.compile flex_attention case where numerical differences from compiler optimizations were misclassified. Developing domain-specific oracles that incorporate expert-level reasoning about compiler internals and floating-point precision behavior is needed for silent bug fuzzing in compiled deep learning operations.

• Quantum-SpinalNet: a hybrid deep learning approach for mammographic breast cancer detection (2026) · doi

  The paper demonstrates that integrating biologically-inspired SpinalNet with quantum-inspired DQNN achieved synergistic performance improvements (93.8% accuracy, Dice = 0.89), but does not experimentally isolate whether these gains derive from reduced overfitting via layer-wise modular processing or from improved feature abstraction through quantum probabilistic reasoning.

• Ocean: Object-aware Anchor-free Tracking with Matching-relation Learning (2026) · doi

  The paper demonstrates that online update benefits are differentially masked by restart mechanisms on VOT datasets versus non-restart datasets (GOT-10K, LaSOT), but does not investigate whether alternative performance estimation methods could decouple these effects or provide more reliable generalization metrics for anchor-free trackers.

• Automated design of heuristics for resource-constrained project scheduling problem via regression algorithms (2026) · doi

  The paper reports that AllPR (all priority rules) demonstrated clear advantages on small project instances but lost this advantage on large projects, yet the underlying causes for this performance degradation with problem scale are not analyzed. Research should investigate how priority rule combination effectiveness diminishes as project size increases and solution space expands in the resource-constrained project scheduling problem.

• From unstructured text to structured reasoning: a hybrid knowledge graph for Indonesian sentencing analysis (2026) · doi

  While the paper demonstrates that objective entities (F1 > 90%) correlate with standardized formats and interpretive entities (F1 < 80%) with legal reasoning variation, it does not propose or validate methods to explicitly model this epistemological distinction in the hybrid knowledge graph structure for improved entity disambiguation.

• On the interface between linguistics, computer science and psychiatry: analyzing textual key-factors affecting BERT-based classification of schizophrenia in social media texts (2026) · doi

  Topicality effects on semantic coherence salience in schizophrenia detection remain theoretical; the paper cannot yet determine whether BERT captures deeper grammatical or coherence disruption patterns related to topic versus simply capturing differences in linguistic information density and discourse register across genres. Cross-topic coherence analysis with controlled semantic and syntactic complexity is required.

• Predicting Employee Attrition: A Machine Learning Approach in Human Resource Analytics (2026) · doi

  The feature importance analysis reveals divergent rankings between Gradient Boosting and Random Forest models for secondary attrition predictors, with Gradient Boosting prioritizing workload variables (Overtime, Stock Option Level) while Random Forest emphasizes demographic factors (Distance from Home, Years at Company). The paper does not investigate whether these differences stem from model architecture bias or represent genuine contextual variations in attrition mechanisms across employee subgroups.

• Hepatitis C Diagnosis using Supervised Machine Learning Algorithms and Ensemble Learning Techniques (2026) · doi

  The study demonstrates that SVM achieves perfect recall (100%) but 0% specificity and ROC-AUC of 0.50 in hepatitis C diagnosis, indicating poor differentiation of healthy cases. The underlying mechanisms causing this severe specificity collapse in SVM compared to ensemble methods like AdaBoost and Random Forest require investigation through kernel selection analysis and class boundary examination.

• Understanding the Dynamics of Trust and Engagement in E-Commerce Recommender Systems: Trends and Influences (2026) · doi

  Current theoretical frameworks for trust and engagement in e-commerce recommender systems are predominantly based on research from China and Western markets, with African informal and community-driven recommendation networks critically underrepresented. Future work must empirically investigate how trust conceptualizations differ across African, South Asian, and Latin American e-commerce contexts to validate whether existing engagement models capture culturally diverse user interactions with recommender systems.

• Deep Learning Based Fish Species and Freshness Detection Using Convolutional Neural Networks (2026) · doi

  The three freshness categories (Fresh, Medium, Spoiled) used for CNN classification lack correlation with objective freshness metrics such as bacterial load, pH levels, or volatile organic compound concentrations; validation against biochemical freshness indicators is absent.

• Unified URL and QR Based Phishing Detection Framework (2026) · doi

  While ethical considerations address data protection and transparency, the paper does not evaluate adversarial robustness: how the model performs against evasion attacks (e.g., homograph attacks, obfuscated URLs, dynamically generated QR codes with steganographic encoding).

• A Comprehensive Review of Machine Learning Techniques for Student Academic Performance Prediction (2026) · doi

  Causal inference methods for determining which interventions are most effective for specific student subgroups in academic performance prediction have not been systematically applied or compared. Research is needed to establish causal models that move beyond correlation-based predictive approaches.

• ON DOMINATION IN COARSE DEG-CENTRIC GRAPHS (2026) · doi

  For coarse deg-centric graphs where the coarse deg-centric transformation yields an empty graph K_{2n-1} (such as in flower and blossom graphs), the relationship between the original graph structure and why certain graph operations consistently produce empty graphs in the coarse deg-centric representation is not theoretically explained.

• An Approximate Decomposition of a Multivariate Polynomial and Its Application (2026) · doi

  Algorithm 2 computes approximate decompositions for indecomposable multivariate polynomials using Hamming distance, but the authors explicitly state that 'when it is indecomposable, we cannot determine the extent of the difference between the true solution and the output.' This requires developing theoretical bounds or optimality analysis to quantify how far the approximate decomposition is from an optimal solution.

• AI-Powered Resume Screening and Ranking System (2026) · doi

  The paper does not address how the AI-Powered Resume Screening System handles implicit or non-obvious qualifications, transferable skills, or alternative certifications that may not match exact keywords in the job description. The keyword matching technique may systematically disadvantage candidates with unconventional career trajectories or those who use domain-specific terminology different from standard job requirement vocabulary.

• A Systematic Review of Machine Learning Approaches For AI-Driven Fraud Detection in Loyalty Programs (2026) · doi

  No quantitative framework exists for choosing between interpretable tabular models and graph-based approaches based on program size and fraud taxonomy breadth, preventing loyalty operations with intermediate-scale datasets from objectively deciding when graph learning complexity is justified versus when simpler tabular models suffice.

• A deep learning approach to multi-marginal optimal transport via Hilbert space embeddings of probability measures (2026) · doi

  Condition (A5) relies on universal approximation theorems (Hornik 1991 for bounded σ, Kidger-Lyons 2020 for ReLU), but the paper notes the ReLU result applies to L²(Rd; Rd) rather than the required L²(μ1) space; rigorous verification that this approximation property holds for all measure distributions μ1 in the problem is missing.

• LISArD: learning image similarity to defend against gray-box adversarial attacks (2026) · doi

  While LISArD demonstrates resistance to Sparse-RS and shows that adversarial training (AT) does not significantly improve defense against this attack, the underlying mechanism explaining why the image similarity learning component succeeds where AT fails for this specific perturbation type is not theoretically analyzed.

• Utilizing Cascade Deep Metric Learning for the Kellgren-Lawrence Grading of Knee Osteoarthritis Classification from X-Ray Images (2026) · doi

  The paper demonstrates that CDML with VGG19 outperforms deeper architectures (ResNet50, EfficientNetB0), but does not investigate whether this advantage generalizes to other shallow/medium-depth backbones or explain the mechanistic reasons why the cascade ordinal-aware learning design interacts differently with network depth in knee OA classification.

• An Acoustic Feature-Based Ensemble Learning Approach for Chicken Health Detection (2026) · doi

  The paper demonstrates that RF with MFCC features minimizes false negatives while maintaining low false positives, but the mechanism by which different acoustic feature representations (MFCC vs. alternative spectral or temporal features) interact with specific classifier architectures (SVM, LR, RF) to produce these distinct error patterns is not theoretically analyzed or explained.

• From Machine Learning to Heterogeneous Models: A Comprehensive Study on Fine-Grained Emotion Detection in Arabic Text (2026) · doi

  The sequential hybrid architecture achieved the best performance (F1=0.802), but no analysis was conducted on why intermediate predictions from MARBERTv2 followed by BiGRU and LR improve learning on hard instances compared to alternative sequences or different numbers of base models in the sequential pipeline.

• IoT-Driven Transformer Health Monitoring and Fault Detection: Advancing Reliability in Emerging Power Systems (2026) · doi

  Humidity monitoring (≤60% RH threshold) is included as a parameter affecting insulation health, but the paper provides no data on the correlation between measured humidity levels and actual insulation degradation rates, dielectric strength loss timelines, or the relationship between ambient humidity and internal winding moisture accumulation in different transformer designs. Quantitative validation of humidity thresholds against insulation failure mechanisms is absent.

• Nowcasting of Airport Low Visibility Based on Machine Learning (2026) · doi

  The referenced studies on hierarchical classification methods for visibility prediction and relevance analysis-based data balancing techniques do not specify the optimal threshold values for visibility category boundaries (e.g., 50m, 200m, 1000m) or explain how these thresholds should be adjusted across different airports with varying operational constraints.