🏗️ Overall Corruption Analysis Pipeline (Fig. 1)

Figure 1 (page 2). The paper’s pipeline starts from survey/perception data and data processing to build a dataset of organizations. In parallel, corruption frequency statistics are used to build a weight vector. The core step is weighted-attribute K-means, producing meaningful clusters, followed by cluster analysis & labeling.

📊 Data Collection & Organization Attributes (Section III)

Each public-sector organization is represented using 7 attributes: Integrity, Accountability, Independence, Resource, Transparency, Co-operation, and Awareness. A multidimensional survey is used to score each attribute using question-level ratings of 1 (minimum), 3 (mid-point), or 5 (maximum). For each organization and attribute, the final value is the average across surveyed users.

🧾 Sample Survey Assessment (Table I)

Table I (page 2). Example questionnaire block used to compute an attribute score (Transparency) for an organization.

Dataset size (as reported): data collected from 70 survey participants; dataset includes 20 public-sector organizations; each described by 7 attributes with values in [1, 5].

📌 Why Attribute Weighting Matters (Fig. 2, Table II)

Standard K-means uses Euclidean distance and effectively treats all attributes as equally important. The paper argues this is misleading for corruption analysis because some corruption types occur much more frequently (for example, bribery and negligence). Therefore, the model builds a weight vector using corruption-type frequency statistics and applies weighted distance in clustering.

Figure 2 (page 2). The paper reports household victim percentages by corruption type, including Bribe (49.8) and Negligence (39.9) as the most frequent categories, followed by smaller-frequency types. These frequencies motivate higher weights for attributes linked to high-frequency corruption types.

⚖️ Weight Vector of Attributes (Table II)

Table II (page 2). Normalized attribute weights used by the proposed model.

🧠 Core Model: Attribute-Weighted K-means (Section IV)

The clustering uses a weighted distance (instead of standard Euclidean distance). Each attribute belongs to a weight partition, and the distance between organizations is computed by multiplying each squared attribute difference by its attribute weight.

The algorithm initializes K = 3 cluster centers using uniform attribute values from the set {1, 3, 5}, assigns organizations to the nearest weighted-distance centroid, and iteratively updates cluster centers until the stress function falls below a threshold.

📈 Results: Cluster Centers & Separation (Table III, Fig. 3)

📐 Resultant Cluster Centers (Table III)

Table III (page 3). Cluster centers found by the proposed weighted K-means.

Figure 3 (page 3). Scatter plot of organizations using the two most heavily weighted attributes: Integrity and Accountability, showing cluster separation.

🏷️ Cluster Labels (Table IV)

Table IV (page 3). Cluster labeling derived from weighted sums of cluster-center attributes.

✅ Evaluation vs Reference Corruption Report (Tables V–VI)

To validate reliability, the paper compares the model’s cluster labels against a reference corruption report (TIB National Household Survey 2017). Performance is summarized using precision, recall, and F1-score per cluster, along with macro and weighted averages.

📊 Performance Analysis (Table V)

Table V (page 4). Precision/recall/F1 by cluster, plus macro and weighted averages.

🧾 Sector-by-Sector Comparison (Table VI)

The paper lists sector corruption percentages from the reference report and compares them against the model’s predicted class. (Below is the paper’s table rendered into HTML.)

Table VI (page 4). Side-by-side comparison between reference sector classification and the model’s predicted class.

☁️ Real-Time Cloud Web Application Architecture (Fig. 4)

Figure 4 (page 4). The paper’s deployment design uses a Flask-based web UI where a user selects an organization, answers questionnaires, and submits scores. The backend processes scores, stores the dataset in a database, runs the proposed clustering on a cloud server, and returns results with visualizations (charts/tables/plots).

🚀 Practical Impact (as described)

• Early segmentation: Uses perception data to group sectors as least/moderate/high corruption, supporting prioritization of audits and interventions.
• Adaptable weighting: Attribute weights reflect corruption-type frequency, so the model emphasizes dominant corruption channels.
• Scalable collection: Web-based design supports larger-scale data collection and repeated re-analysis over time.

⚠️ Limitations & Future Directions (as stated)

• Collect more data across countries, more organizations, and more diverse populations to improve generalization.
• Expand survey coverage to better represent the full spectrum of public-sector experiences.

📝 Citation

@inproceedings{pramanik2020public,
  title={Public sector corruption analysis with modified K-means algorithm using perception data},
  author={Pramanik, Anik and Sarker, Amlan and Islam, Zabirul and Hashem, MMA},
  booktitle={2020 11th International Conference on Electrical and Computer Engineering (ICECE)},
  pages={198--201},
  year={2020},
  organization={IEEE}
}