Abstract
The development of the Internet makes it easier to get data. However, it is more urgent to obtain effective information on the large data which is redundant. Based on the risk prediction system of power intelligent automation, this paper tries to solve the problems of enterprise risk management, and tries to make human beings free from the heavy human and mental labor. In order to solve the problem of large data analysis, this paper constructs a close triangle case based reasoning model based on the data mining technology of NT-SMOTE and the case based reasoning technology. Compare with NT-MDA, NT-Logit, NT-Probit, NT-DT, NT-SVM and other methods, the experimental results show that NT-CBR is more effective than other methods of intelligent risk prediction.
Introduction
In the era of big data, the primary challenge for enterprise risk management is not the acquisition of information, but the filtration of noise to identify critical signals. Within "power intelligent automation" systems, data is often highly imbalanced. Typically, datasets contain thousands of healthy operational samples (negative class) against a minute fraction of "at-risk" events (positive class). Standard statistical models and machine learning algorithms, such as Support Vector Machines (SVM) or Decision Trees (DT), often struggle with this imbalance, frequently ignoring the minority class to maximize overall accuracy. This paper introduces a hybrid risk-prediction model, NT-CBR, designed to accurately identify rare but catastrophic risks, thereby reducing the manual analytical burden on human managers.
Methodology
The proposed system utilizes a two-stage approach to handle imbalanced data and perform robust prediction.
- Data Balancing with NT-SMOTE: To address the "needle in a haystack" problem, the researchers employ the Neighborhood Triangular Synthetic Minority Over-sampling Technique (NT-SMOTE). Unlike standard over-sampling that simply duplicates rare cases, NT-SMOTE leverages the geometry of "adjacent triangle spaces" to generate realistic synthetic data points for the minority class. This creates a balanced training set without biasing the model toward the majority class.
- Prediction via Case-Based Reasoning (CBR): CBR mimics human cognitive problem-solving by finding historical cases most similar to a new situation and adapting past solutions. This analogical approach is often more robust than pure statistical boundary-finding for financial crises, which tend to follow recognizable patterns.
The NT-CBR model was benchmarked against five other methods: Multiple Discriminant Analysis (MDA), Logit, Probit, Decision Trees (DT), and Support Vector Machines (SVM), all preprocessed with NT-SMOTE.
Experimental Setup
The study validated the model using data from the Chinese tourism and hotel industry, serving as a proxy for financial risk in complex systems.
- Features: Each sample included 41 financial indicators.
- Timeframes: Analysis was conducted at three intervals: one year ($t-1$), two years ($t-2$), and three years ($t-3$) prior to a financial distress event.
- Data Imbalance: The raw dataset was heavily skewed, with approximately 200 "healthy" samples for every 11 "troubled" samples.
Results and Discussion
Experimental results across 30 repeated trials demonstrate that NT-CBR significantly outperforms traditional optimization methods.
Key findings include:
- Superior Sensitivity: At $t-2$ (two years before a crisis), NT-CBR achieved a 100% sensitivity rate, successfully identifying every at-risk case. In comparison, the Decision Tree (NT-DT) only achieved 59.33% sensitivity.
- Balanced Performance: NT-CBR maintained high scores across Sensitivity, Specificity, and the G-mean (a metric for balanced performance) across all timeframes.
- Reliability: While methods like NT-MDA performed reasonably well, NT-Logit and NT-SVM exhibited high volatility and poor sensitivity, making them less reliable for real-world risk management where missing a single positive case is costly.
The core insight is that risk prediction is better handled by analogy (CBR) than by fitting complex multi-dimensional curves. By using spatial geometry for over-sampling, the model fills data gaps effectively without the overfitting common in random over-sampling.
Conclusion
The NT-CBR model provides a robust framework for intelligent risk prediction in big data environments. While validated on financial data, the model offers a theoretical foundation for building intelligent systems in power automation and other enterprise safety domains. Future work should focus on industry-specific feature selection and dynamic maintenance of the case database to adapt to evolving economic conditions.
Summary updated