Machine Learning Towards Intelligent Systems: Applications, Challenges, and Opportunities

• Information overload occurs when data input exceeds a person's cognitive capacity and can lead to crucial information being ignored, overlooked, or misinterpreted.Source: §1 · p. 2
• Machine learning enables computers to learn patterns from training data and apply them to similar data without being explicitly programmed.Source: §1 · p. 2
• The review covers challenges in education, healthcare, network security, banking and finance, and social media, alongside prior work and research opportunities for ML.It also discusses potential frameworks for deploying ML to address these challenges.Source: §1 · p. 2

• Organizations use ML to learn from collected data and predict business changes, while development efforts include algorithms and paradigms such as neural networks and support vector machines.Source: §2 · pp. 2, 4
• The global ML market is expected to reach close to $42.5 billion CAD by 2024; cited projections also describe potential EU economic growth and net job creation.The EU projection is up to 20% economic activity growth by 2030, and the World Economic Forum predicted a net of 58 million jobs in the coming years.Source: §2 · pp. 2–3, Figure 1 · p. 3
• Federated learning trains a centralized model from data distributed across locations by aggregating local model updates, addressing privacy, ownership, and locality concerns.The section describes it as bringing code to data rather than data to code.Source: §2 · p. 3
• Frameworks including TensorFlow Federated support training across multiple data sources while keeping each source local.Other examples named include Federated AI Technology Enabler, PySyft, Leaf, PaddleFL, and Clara Training Framework.Source: §2 · p. 3

3.1 Essay Grading

• ML essay scoring could reduce grading workload and subjectivity; future work should compare algorithms and richer NLP features, with potential use in exams and writing tutorials.Mahana et al. (2012) reported an average kappa score of 0.73, while Ramalingam et al. (2018) reported over 80% accuracy.Source: §3.1.2 · p. 5, §3.1.3 · p. 6

3.2 Dropout Prevention

• Dropout prediction is difficult because student and course factors vary, but identifying at-risk students can support interventions; research should test other settings, student attributes, and learning methods.Coussement et al. (2020) reported an AUC above 0.8 for their logit leaf model; Chung and Lee (2019) reported close to 95% accuracy and close to 0.97 AUC for random forests.Source: §3.2.1 · p. 7, §3.2.2 · pp. 7–8, §3.2.3 · p. 8

3.3 Intelligent Tutoring

• Intelligent tutoring must address language errors and other interacting factors; research should test phonemic and history-based features and compare classifiers such as NN and SVM.Barron-Estrada et al. (2017) used Naive Bayes for sentiment analysis of Spanish Twitter messages and reported accuracy above 80%, but did not compare other classifiers.Source: §3.3.1 · p. 8, §3.3.2 · pp. 8–9, §3.3.3 · p. 9

3.4 Course Recommendation

• MOOC recommenders can use student data to match courses to skill goals; research should evaluate skill-related course histories and compare supervised classification algorithms.Aher and Lobo (2013) combined Simple K-means and Apriori, matching real-world course-selection patterns; Mondal et al. (2020) combined K-means with collaborative filtering and reported low errors and high precision and recall.Source: §3.4.1 · p. 9, §3.4.2 · pp. 9–10, §3.4.3 · p. 10

3.5 Personalized Learning

• Personalized learning uses learner profiles and behaviors to tailor scenarios and pathways, relate engagement to academic performance, and predict performance during course delivery.LearnFitII mined learner logs and preferences to update scenarios with a hybrid recommender, while Elfaki et al. (2014) used decision-tree rules for pathway recommendations; other studies linked engagement with performance and used optimized ensembles for performance prediction.Source: §3.5.1 · p. 10, §3.5.2 · pp. 10–11
• Personalized-learning recommendations could incorporate learner motivation, knowledge level, and additional personality traits.Source: §3.5.3 · p. 11
• Comparing classification algorithms to predict student performance during course delivery could help identify students needing assistance and provide personalized improvement plans.Source: §3.5.3 · pp. 11, 13

4 Healthcare

• Healthcare data span heterogeneous formats, and professionals may be overwhelmed; the section surveys ML applications in drug response, diabetes, retinopathy, and cancer detection.ML is presented as a way to identify relevant patterns in complex data.Source: §4 · p. 13

4.1 Drug Response Prediction

• An SVM predicted drug responses for 175 cancer patients with >80% accuracy; proposed extensions include emerging therapies, richer drug features, semi-supervised learning, and adaptation.The opportunities also include adapting models to other drugs, cancer types, and diseases.Source: §4.1.2 · p. 14, §4.1.3 · p. 15

4.2 Diabetes Research

• A single HbA1c marker at the 6.5% cutoff may leave T2D cases undiagnosed, motivating use of additional biomarkers.The section also notes that real-world clinical trials and comparison of classification models remain research opportunities.Source: §4.2.2 · p. 16, §4.2.3 · p. 17
• 86.64% (from 78.71%) with 8-OhdG and 85.63% (from 78.71%) with IL-6 were reported for T2D detection alongside HbA1c.The study used data from 840 patients and reported an optimal HbA1c range of 5.73–6.22%.Source: §4.2.2 · p. 16
• In simulations, CBR(R2R) reduced mean blood glucose, eliminated hypoglycemia, and outperformed standalone R2R in both populations.The simulated scenarios included variation in insulin sensitivity and uncertainty in measurements and carbohydrate intake.Source: §4.2.2 · pp. 16–17

4.3 Retinopathy Detection Through Image Classification

• Ensemble models achieved up to 94% accuracy for diabetic retinopathy detection; proposed opportunities include optimization and CNN–RNN models for multi-label classification.Reported ABCA4-related retinopathy detection accuracies ranged from 86% to 93%.Source: §4.3.2 · p. 18, §4.3.3 · p. 19

4.4 Cancer Detection Through Image Classification

• Breast-cancer detection accuracy ranged from 93%-99%, while prostate-cancer SVM with RBF kernel reached 98%-99%; opportunities include tuning, ensembles, and feature selection.The cancer section also reports 77.8% lung-cancer detection accuracy for the entropy degradation method.Source: §4.4.2 · pp. 19–20, §4.4.3 · p. 21

5.1 Network Intrusion Detection Systems

• Signature-based NIDSs detect known attacks effectively but miss attacks without installed signatures, while anomaly-based NIDSs can detect new attacks but produce high false-positive rates.The section identifies these trade-offs as challenges for efficient and flexible detection of unknown future attacks.Source: §5.1.1 · p. 22
• Javaid et al. (2016) report above 98% classification accuracy; Injadat et al. (2020b) report over 99% detection accuracy on both datasets, with training samples reduced by 74% and features by up to 50%.Injadat et al. (2020b) also report 1-2% improvements in accuracy and false alarm rate against other literature works.Source: §5.1.2 · p. 22
• Further NIDS research could test complex models for real-time detection, optimization techniques, temporal analysis, and reinforcement or transfer learning.The section also calls for extending deep-learning and time-series approaches beyond IDSs to other network-security problems.Source: §5.1.3 · p. 23

5.2 Botnets Detection

• McDermott et al. (2018) report 92%-99% detection accuracy; Pektas¸ and Acarman (2017) report nearly 99.9% accuracy; Chen et al. (2017) report a 13.2% decrease in false positive rate.Chen et al. (2017) also report 93.6% detection accuracy and a 0.3% false positive rate for the RF algorithm.Source: §5.2.2 · p. 24
• Botnet research opportunities include hybrid models for online detection, non-numerical features, and testing optimization techniques.The section frames these as opportunities to explore whether they meet online-framework requirements or improve model performance.Source: §5.2.3 · p. 24

5.3 Intrusion Detection in Vehicles

• In-vehicle ECUs cannot identify the sender of ECU-to-ECU broadcasts on the CAN bus; Kang and Kang (2016) report a real-time DNN response with a 98% average detection ratio.Their DNN was more accurate than a traditional ANN; the authors attribute this to parameter initialization through unsupervised DBN pre-training.Source: §5.3.1 · p. 25, §5.3.2 · p. 25
• Yang et al. (2019a) report 100% accuracy on CAN intrusion and 99.86% on CICIDS2017; Zeng et al. (2019) report 95%-99% precision and 0.92-0.99 F1 using raw traffic.Yang et al. report computational-time reductions of 73.7% to 325.6s and 38.6% to 2774.8s, respectively.Source: §5.3.2 · p. 25
• Vehicular IDS research could test optimization and meta-heuristics, develop more complex or hybrid systems, and detect both known and unknown attacks.Table 3 summarizes network-security challenges, prior works, and research opportunities; Figure 4 maps NIDS, botnet detection, and vehicle IDS to the NIST Cybersecurity Framework’s Detect level.Source: §5.3.3 · pp. 25, 27, Table 3 · p. 26

6.1 Credit Risk Assessment

• Credit assessment can omit dynamic transaction histories and SME-specific factors; proposed research includes hyperparameter tuning, longer-term prediction, and other deep-learning architectures.Zhu et al. (2019)’s hybrid RS-multi-boosting model achieved the highest accuracy, ranging from 67%-84%, on its Chinese supply-chain-finance dataset.Source: §6.1.1 · p. 27, §6.1.3 · p. 28, §6.1.2 · p. 28
• Credit-risk studies report accuracy ranges of 81%-91% for China, 64%-79% for Germany, and 64%-86% for Australia; a SME DBN reached 96%, versus 82% for LR and 87% for SVM.The ranges are Bao et al. (2019)’s detection accuracy for potential models; the DBN comparison is Xu and He (2020)’s result.Source: §6.1.2 · p. 28

6.2 Bankruptcy Prediction

• Bankruptcy predictions can provide financial institutions early warning of potential client failures, but prior models’ performance depends heavily on collected features and data.Suggested work includes testing additional models and kernels, optimization methods against computational complexity, and deep-learning architectures beyond CNNs.Source: §6.2.1 · pp. 28–29, §6.2.3 · p. 30
• Barboza et al. (2017)’s bagging, boosting, and RF models achieved 86%-87% testing accuracy versus 72%-76% for LR and ANN; Kim et al. (2019)’s GOSVM outperformed non-SVM and other SVM-based models during hold-out prediction.Lin et al. (2019) found bagging ensembles had lower false-positive rates than single classifiers, and genetic-algorithm feature selection enabled better classifier performance than information gain.Source: §6.2.2 · p. 29

6.3 Currency Crises Prediction

• For currency crises, RF-DWT achieved 89%-90% accuracy versus 84%-85% for LR, while Kinkyo (2020)’s model reached close to 73%, at least 5% above other models.McMahon (2019) reported that an SVM predicted the majority of crises in 17 emerging markets; Alaminos et al. (2019) reported 97%-99% testing accuracy across regions with DNDT.Source: §6.3.2 · p. 31
• Currency-crisis research opportunities include testing models in specific fields such as housing, comparing classifiers and their complexity, and exploring other deep-learning models.The review describes currency crises as economically damaging and identifies early warning as a potential means to prevent or manage them.Source: §6.3.1 · pp. 30–31, §6.3.3 · pp. 31–32

7.1 Pharmacovigilance

• Voluntary ADR reports may be delayed, incomplete, duplicated, or under- or over-reported, motivating social-media data as a supplement; lexicon methods also struggle with informal language and spelling variation.Supervised approaches require substantial manually annotated data, often from domain experts.Source: §7.1.1 · pp. 33–34
• For ADR identification, Cocos et al. (2017)’s RNN achieved an F1-measure of 0.755, versus 0.63 for lexicon matching and 0.65 for CRF.Other studies reported precision of 91%-94% for ADRs and 79%-87% for medical events, and precision of 81%-84% on CADEC versus 64%-73% on Twitter.Source: §7.1.2 · p. 34
• Pharmacovigilance research opportunities include modeling tweet context and semantics, enriching ADR lexicons, temporal analysis of drug–ADR patterns, and testing HMMs and transfer learning.Patki et al. (2014) found their approach only marginally successful at classifying drugs as normal or blackbox, citing limited annotated data and noisy text.Source: §7.1.3 · pp. 34–35, §7.1.2 · p. 34

7.2 Social Media and Vaccines

• In vaccine studies, users more often exposed to negative HPV-vaccine opinions were more likely to subsequently post negative tweets; a classifier’s accuracy ranged from 87.6% to 94.0%.Dunn et al. (2015) analyzed temporal tweet sequences and user connections; the reported relationship is an association, not a causal finding.Source: §7.2.2 · pp. 35–36
• Vaccine-related studies reported precision of 71%-78% for HPV-vaccine sentiment, around 80% for negative safety sentiment, and correlation r = 0.90 with CDC flu-vaccination survey data.The survey-data comparison concerns Huang et al. (2017)’s Twitter-based tracking results.Source: §7.2.2 · p. 36
• Vaccine sentiment research could address changing social connections with adaptive models, update datasets for evolving slang and abbreviations, and test CNNs and RNNs.The review notes that connections captured in a fixed-period dataset may not reflect later changes.Source: §7.2.3 · p. 36

7.3 Social Media and Politics

• Political sentiment studies reported 97%-99% accuracy and F1-scores of 0.94-0.97 for the 2016 US election, while another study’s RNTN estimate of a 9.2% Democratic advantage was close to the actual 8.6%.The 2012 US election study also found that its Twitter-based lead matched the election outcome.Source: §7.3.2 · pp. 37–38
• Political sentiment research opportunities include testing SVM and ANN, adding linguistic and contextual features, and comparing CNNs and RNNs with RNTN.The review describes social media monitoring as a way to analyze politically relevant information and public opinion.Source: §7.3.1 · pp. 36–37, §7.3.3 · p. 38

8 Conclusion

• The review surveys challenges and prior work across education, healthcare, network security, banking and finance, and social media, and identifies further ML research opportunities.Figure 7 summarizes challenges and previous or potential ML techniques addressing them.Source: §8 · p. 40, Figure 7 · p. 40