Source-linked AI summary
TNF based Spectral Embedding for Effective Application of Supervised Machine Learning Techniques in Automobile Insurance Fraud Detection
Rohan Yashraj Gupta, Lalith Srikanth Chintalapati, Satya Sai Mudigonda, Pallav Kumar Baruah, Raghunatha Sarma Rachakonda
TL;DR
Insurance fraud carries substantial financial costs, while traditional auditing and expert inspection are costly and time-consuming. The paper evaluates a three-phase automobile-insurance fraud-detection pipeline; MWMOTE, TNFSE2, and Random Forest are reported as the best tested combination, with the strongest reported Accuracy, F1-Score, and AUC-ROC values.
Problem
Insurance fraud causes financial losses, and traditional auditing and expert inspection are costly, inefficient, time-consuming, and require human intervention.
Source: §1 · p. 1Method
The paper evaluates a three-phase automobile-insurance pipeline combining imbalance handling, embedding, and classification, including oversampling, TNF-based embedding, and multiple classifiers.
Source: §1 · p. 2, §2.1 · p. 3, §5 · p. 6Results
M37, combining MWMOTE, TNFSE2, and Random Forest, has the best reported Accuracy (0.9917), F1-Score (0.9918), and AUC-ROC (0.9917) among the models.
Source: §5.2 · p. 9Takeaways & Limitations
For the tested dataset and configurations, MWMOTE with TNFSE2 and Random Forest is the strongest reported combination.
Source: §5.2 · p. 9
Abstract
from arXiv · showhide
Fraud detection is an important area of research in the insurance business due to its financial implications. The primary aim of a fraud detection model is to identify fraud and non-fraud cases with high accuracy along with other important metrics such as Sensitivity, Specificity, Precision, F1-score, False Positive Rate, False Discovery Rate, AUC etc. To achieve this, we need to explore a suitable classification model to identify fraud and non-fraud cases. In this work, we have used auto insurance data set and explored classification models such as Decision Tree (DT), Random Forest (RF), XGBoost, LightGBM and Gradient Boosting Machine (GBM). To overcome the problem of data imbalance, we have employed MWMOTE and TGAN techniques. We have used Topological Node Feature(TNF) based spectral embedding for low dimensional data representation along with some popular embedding methods like MDS, Isomaps and t-SNE. After studying all the 65 possible combinations of these models, we have proposed an innovative method for effective automobile insurance fraud detection. For the given dataset, our results show that using a combination of MWMOTE as a data imbalance handling technique (Phase I), TNFSE2 as data embedding (Phase II) and Random Forest as classification (Phase III) provides the best result in comparison to all other combinations. This work also highlights the efficacy of TNF based spectral embedding in automobile insurance dataset
1. Introduction
Insurance fraud causes substantial financial losses, while traditional auditing and expert inspection are described as costly and inefficient. The paper proposes a three-phase automobile-insurance detection method and introduces TNF-based spectral embedding alongside existing dimensionality-reduction approaches.
1. Introduction
- About 4% of insurance money is lost to insurance fraud, and 21%–36% of auto-insurance claims are described as suspicious, with less than 3% prosecuted.The passage also reports $40 billion in annual non-health insurance fraud losses.Source: §1 · p. 1
- Traditional auditing and expert inspection are costly, inefficient, time-consuming, and require substantial human intervention.The passage notes that organizations are adopting data-mining and machine-learning techniques to analyze large datasets and build fraud-detection models.Source: §1 · p. 1
- The number of dataset features is a training challenge, and prior dimensionality-reduction approaches include MDS, Isomap, t-SNE, and Laplacian Eigenmaps.The passage describes MDS as representing similarity or dissimilarity through geometric distances, Isomap through graph geodesic distances, and t-SNE through probabilities.Source: §1 · p. 2
- The paper proposes TNF-based embedding methods TNFSE1 and TNFSE2 within a three-phase pipeline for handling imbalance, embedding data, and classifying claims.The introduction also situates TNFs in prior spectral-clustering work and describes an affinity measure intended to capture local information.Source: §1 · p. 2
2. Methodology
The methodology organizes automobile-insurance fraud detection into imbalance handling, low-dimensional embedding, and classification. It combines oversampling methods, TNF-based and other embeddings, and five classifiers.
2. Methodology
- The proposed workflow handles data imbalance in phase I, embeds data in phase II, and classifies claims in phase III.The schematic is identified as the proposed method.Source: §2 · pp. 2–3, Figure 1 · p. 3
2.1 Phase I
- MWMOTE and TGANs oversample fraudulent claims, using minority-cluster weighting in MWMOTE and a learned minority-sample distribution in TGANs.The TGAN discriminator is trained until it can no longer distinguish real from synthetic data, after which the model generates minority samples.Source: §2.1 · p. 3
2.2 Phase II
- The TNF representation uses graph degree, clustering coefficient, and a recursively aggregated summation index to construct claim affinities before spectral embedding.The affinity matrix expresses how closely claims are related; the embedding procedure constructs a normalized representation from Laplacian eigenvectors.Source: §2.2 · pp. 3–5
- TNFSE1 uses degree and clustering coefficient, while TNFSE2 adds the summation index to form the TNF-based affinity for spectral embedding.The paper compares these variants with MDS, Isomap, and t-SNE.Source: §2.2 · pp. 3, 5
- The comparison embeddings are MDS, Isomap, and t-SNE, which represent data through Euclidean distances, graph geodesic distances, and affinity probabilities, respectively.The passage describes t-SNE as useful for revealing structure and multiple clusters while reducing crowding in the embedded space.Source: §2.2 · p. 5
2.3. Phase III
- Phase III classifies claims with Decision Tree, Random Forest, XGBoost, LightGBM, and Gradient Boosting Machine models.The paper describes Random Forest as an ensemble of decision trees and XGBoost as differing from GBM in its use of the loss function.Source: §2.3 · p. 5
3. Data Description & preparation
The study uses a preprocessed automobile-claims dataset with a 70:30 train-test split.
3. Data Description & preparation
- The carclaims.txt dataset contains 15,420 claims, including 6% fraudulent claims, and 32 features.The passage reports 6 ordinal and 25 categorical features.Source: §3 · p. 5
- Preprocessing engineered date-related features, discarded one record with missing values, and one-hot encoded categorical features.The engineered features include accident date, reporting date, and reporting delay.Source: §3 · p. 5
- The dataset was randomly split 70:30 into training and test sets while keeping the fraud and non-fraud ratios aligned across the split.Source: §3 · p. 5
4. Performance Metrics
The study evaluates model performance with eight metrics covering fraud detection, non-fraud classification, and overall predictive performance.
4. Performance Metrics
- The evaluation uses Sensitivity, Specificity, Precision, False Positive Rate, False Discovery Rate, Accuracy, F1 Score, and AUC-ROC.Source: §4 · p. 6
- Precision is the number of fraud claims identified as fraud divided by all claims the model identifies as fraud.Source: §4 · p. 6
- False Discovery Rate is defined using non-fraud claims identified as fraud over all claims identified as fraud by the model.Source: §4 · p. 6
5. Results and discussion
The results progress from raw-data classifiers to imbalance handling and then embedding across three studies. MWMOTE, TNFSE2, and Random Forest form the best reported combination, while TNFSE2 constructs affinities from multi-level claim features.
5.1 Study 1 results
- On raw data, classifiers had sensitivity below 0.2 and specificity above 0.9, indicating poor fraud identification despite strong non-fraud classification.The authors attribute the skew toward non-fraud predictions to training on imbalanced data; Table 1 presents baseline classifier results.Source: §5.1 · pp. 6–7, Table 1 · p. 7
5.2 Study 2 results
- After MWMOTE/TGAN balancing, M7 led AUC-ROC (0.9900), accuracy (0.9901), and F1-score (0.9902), while M15 was best on false-positive and false-discovery rates.The authors report higher AUC-ROC across models M6–M15 than in the unbalanced study.Source: §5.2 · p. 7, Table 2 · p. 7
- M37 (MWMOTE, TNFSE2, and Random Forest) led Accuracy (0.9917), F1-Score (0.9918), and AUC-ROC (0.9917); AUC and F1 showed no significant improvement beyond 16 dimensions.The dimension comparison spans 1–20 eigenvectors, and the study used 16 dimensions for TNFSE2.Source: Table 3 · p. 8, §5.2 · p. 9, Figure 2 · p. 9
5.4 Business interpretation explaining the effectiveness of TNFSE2 embedding
- TNFSE2 builds affinity from degree, clustering coefficient, and summation index across claim-neighborhood levels; higher affinity indicates more related claims, which the authors link to better fraud/non-fraud classification.The vicinity graph links claims whose Euclidean distance is below a fixed ε, representing similarity despite minor attribute differences.Source: §5.4 · p. 10
6. Conclusion and Future Work
The paper proposes a three-phase automobile insurance fraud detection method and reports that MWMOTE, TNFSE2, and Random Forest performed best in the tested combinations. Future work would examine other insurance lines.
6. Conclusion and Future Work
- MWMOTE, TNFSE2, and Random Forest in phases I, II, and III, respectively, gave the best performance in AUC, Accuracy, and F1-Score.The comparison covered various combinations tested on an automobile insurance dataset.Source: §6 · p. 10
- The proposed automobile insurance fraud detection method combines three phases, with multiple model combinations tested on the dataset.Source: §6 · p. 10
- Future work proposes applying the methods to other lines of business, such as health insurance, and identifying an effective phase-by-phase model combination for each.Source: §6 · p. 10