hey can you give me some idea for my own algorithm implementation in machine learning? as i m working with fraud detection problem, i have tried, stacking, cascading, meta, hybrid but not getting improved result. even i checked and tried feature engineering again checked used feature selection method which now contains appropriate features. i want that my precision,recall and f1-score should be improved in comparison of rf and it should consider my own work from research perspective. what should i do,can any one guide me?
I would suggest using XGBoost and data augmentation. Perhaps two projects we carried out could serve as a guide:
Article Improvement of Bank Fraud Detection Through Synthetic Data G...
Article Improvement of Distributed Denial of Service Attack Detectio...
It sounds like you’ve already explored most ensemble and feature selection options, so further gains will likely come from rethinking the learning strategy itself rather than just swapping classifiers.
One approach is combining Reinforcement Learning with Nature-Inspired Optimisation, as shown in my Deep Guard work, where Deep Q-Learning dynamically adjusted detection strategies and Artificial Gorilla Troop Optimization fine-tuned features. While we applied it to deepfake detection, the same adaptive policy learning and search space optimisation can target fraud patterns, potentially pushing precision/recall beyond Random Forest baselines.
A second lever is data governance–driven feature refinement, as seen in our banking fraud detection study, where rigorous standardisation and lineage tracking improved PCA–Decision Tree models, cutting false positives by 35% and increasing model stability.
By integrating these — adaptive learning for decision logic and governance-optimised feature inputs — you can create a genuinely novel fraud detection framework that advances both research and practical performance.
Please check these papers for more details
Data Governance :- Conference Paper Efficacy of Data Governance a Cutting Edge Approach to Ensur...
Deep Guard work :-
Article Deep Guard: Fortifying Digital Authenticity with Deep Q-Lear...
D. Shah ,
Fraud detection is a field that it is adversarial and as such the learning target suffers from dataset shifting. The type of shift depends on the type of fraud being analyzed.
More details on the particular type of fraud that you are trying to discover and the data source are required to offer helpful advise.
Regards
D. Shah ,
Which one of the Kaggle datasets (If you want to send a private message with the information I have no problem)?
To give you a brief background on why the specific dataset is important, do consider the following:
1) Fraud by its nature is a hidden activity. If it were overt it would not be fraud. For the person to carry out a successful fraud scheme, the scheme must be hidden. This means that separating fraud data from real data is a difficult task.
2) Fraud by its nature is adversarial. That means that as soon as it is detected, the adversary modifies it behavior to hide the fraud again. This usually applies to fraud at scale and with adversaries that take advantage of jurisdictions.
3) The fraud can vary considerably whether it is internal or external threat.
4) If the fraud is internal, the fraudster has knowledge of the workings of the system and can hide the fraud more efficiently (assuming it is an intelligent actor). If the fraud is external it is usually easier to detect because of lack of internal system workings.
5) depending which data of the system and accounting mechanism is being analyzed, the attack may be easier to detect.
If you want a work that is valuable to both fraud analysts and not just for ML benchmarking, do get a dataset that is well designed and has a rationale in design.
Regards
Do you mind letting me know about the machine learning/deep learning model used in performing the feature method and finding the detection of fraudlent transaction/activities? The machine learning model used to identify the fraud impacts the outcome significantly while calculating accracy and f-1 score for the inference. Generally, In detection problems consists of categorical/numerical values to apply feature engineering principles on the raw datasets. It could be a significant factor to choose the right engineering principles for the model depends on the type of data. These principles could avoid misimputance, redundancy, regulation errors, distribution skewness of the datasamples etc... Once, the dataset is normalized using standard or adhoc based normalization, PCA based processes for higher dimensionality features applying the advacned techniques based on the linear/nonlinear nature of the datasample result in better confusion matrix parameters including accuracy, f-1 score (in this case). One of the ML model useful in this scenario would be Naive Bayes( Bayesian approach with Probabilisitc measures). There are other techniques including SVM, Clustering, CNN with Recurance/Generative could aid in measurement of these parameters with positive result.
- Badges
- Science topic