Mebone Fraud

Fraud detection is a difficult task. The proportion of fraudulent electronic transactions is very small compared with the legal ones. With the growing number of web sites, merchants, mobile applications and fraudsters gaining in expertise the total amount of the fraudulent operations is increasing. The problem is how to find this little proportion of fraud operations between millions of legal ones. One of the biggest concerns of banks, transaction processors, insurance companies, merchants and online stores is the loss of confidence of the customer. Fraud has a deep impact in recurring purchases or potential clients. Mebone Fraud uses two approaches to achieve fraud detection:

* Machine learning techniques to find fraudulent operations.
* The ability of the fraud analyst to construct rules for fraud detection based on his expertise.

The first approach finds fraud faster than the authorization system and gives the analyst clues to construct new rules. The techniques used fall in the field of artificial intelligence (AI). The latter lets the analyst find fraud by mastering the rules and adapt them to new fraud patterns. This is known as intelligence amplification (IA). These two approaches reinforce themselves achieving more accuracy in fraud detection over time.

Mebone Fraud enables the possibility
to catch fraud in electronic transactions
even before the fraud is commited.

Fraud detection techniques range from simple rule engines to Neural Nets. Mebone Fraud uses Bayesian Networks. Each technique has owns its pros and cons:

Requirement	ANN	BBN	Rules
Analyst knowhow	-	YES	YES
Analysis Performance	YES	YES	N/A
Learning Rate	YES	YES	N/A
Feedback	-	YES	YES
Reaction to change	-	YES	N/A
Fraud detection	YES	YES	YES

Analyst Know How. Usually, Fraud analysts retain lots of knowledge and intuitions from their everyday work. It is important to have the possibility to take advantage of this knowledge using the ability of the system to understand it. Mebone Fraud allows this knowledge transfer using rules expressed in the analyst's natural language. The system takes care of building syntactically correct rules avoiding misleading errors. The Bayesian net is constructed using the analyst knowledge also, but the net proposes the analyst, ways to make a better net design. This is not possible using neural nets.
Analysis Performance. At peak moments, an authorization system can manage hundreds of operations per second. Any operation is a potential fraud, so the system must cope with a high analysis performance. Both techniques (BBN and ANN) accomplish successfully this task. In the case of rules, the analysis time grows because of the amount of information being analyzed (real time and historical data).
Learning Rate. Automatic systems learn from past experience. Both systems (ANN&BBN) need to scan past data to tune their fraud sensibility. While Neural nets need lots of time (even retraining) Bayesian Networks are much faster.
Feedback. It is important for an automatic learning mechanism to explain its decisions. In the case of rules, the explanation is straightforward: one is good if it finds out fraud. BBN gives clues about the fraud detection indicating which of the variables used is responsible of the changing fraud pattern. Looking at that variable it can be found which of its possible values is increasing fraud. For example if one of the variables is "Country", the BBN indicates not only that the fraud is increasing abroad, it also indicates which country is responsible. ANN keeps this information hidden in its data structure.
Reaction to change. It is crucial to find out a fraud pattern as soon as possible. Every technique has its own inertia. This is inevitable, so the system must reduce this inertia as much as possible. The inertia in rules resides in the analyst mind. The inertia in ANN and BBN is inside its data structures. BBN are much quicker than ANN in detecting a new pattern. In some cases the change in pattern is so deep that the ANN must be retrained, consuming lots of human and computer resources.

The combination of state of the art
techniques, makes Mebone Fraud an
unique tool for fraud detection.

Graphical Interface is simple, based on the "Overview and zoom" paradigm. The alarms window presents quickly all the information that the analyst need in a simple way. Alarms are prioritized within each window. Lower-priority alarms do not appear in the visible portion of the alarm window.
Mebone Fraud users can drill down to the alarm details with a single click. One mouse action brings up an elaborate table with data about the object being analyzed. The analyst can build different rule sets one for each detection strategy and all of them can be activated at the same time. Analysts using this tool have a cadre of tireless automated assistants searching for fraud 24 hours a day: Intelligence Amplification.
Mebone Fraud assistance for rule developers consists on a text-based grammar-driven real time parser/editor called Rule Builder. Sophisticated real time parsing is used to present all grammatically valid alternatives as the rule is written and all valid alternatives when the rule is being edited. Syntax error cannot be made using this approach. It provides a rich extensible language that can express a wide range of fraud preventing ideas in a natural way.
All the complex information is presented in graphics to have a quick visual understanding of what is happening, allowing to make fast, real time decisions.

Graphical Interface is simple, based on

the "Overview and zoom" paradigm.

Mebone offers the information that the

analyst needs in a simple way.

Fraud analytics often involve many measurements made over many days. An example is the average activity of a merchant over the last three months. Without access to historical data, filling up this analytic would take many hours, which is clearly unacceptable. Many attempts to use general purpose AI tools in fraud detection assume away the nasty details that must be accommodated in a real time system. Mebone Fraud includes specialized database engines to accomplish complex historical calculations in seconds over the last three months. Information can be retrieved from five years in the past.

On the side of automatic detection, the system must run faster than the authorization mechanism. While the authorization makes its own decisions based on card limits, for example, can also launch asynchronous requests to the anti fraud system to get its fraud evaluation. Mebone Fraud can also be integrated with SMS brokers to communicate with the cardholder at the moment of purchase.

Learning Evolution

Highly optimized algorithms makes
Mebone Fraud capable to analyze card
or electronic transactions in real time.

Fraud detection is a difficult task. The proportion of fraudulent electronic transactions is very small compared with the legal ones. With the growing number of web sites, merchants, mobile applications and fraudsters gaining in expertise the total amount of the fraudulent operations is increasing. The problem is how to find this little proportion of fraud operations between millions of legal ones. One of the biggest concerns of issuer banks, transaction processors, merchants and web sites is the loss of confidence of the customer. Fraud has a deep impact in recurring puerchases or potential clients. Mebone Fraud uses two approaches to achieve fraud detection:

* Machine learning techniques to find fraudulent operations.
* The ability of the fraud analyst to construct rules for fraud detection based on his expertise.

The first approach finds fraud faster than the authorization system and gives the analyst clues to construct new rules. The techniques used fall in the field of artificial intelligence (AI). The latter lets the analyst find fraud by mastering the rules and adapt them to new fraud patterns. This is known as intelligence amplification (IA). These two approaches reinforce themselves achieving more accuracy in fraud detection over time.

Fraud detection techniques range from simple rule engines to Neural Nets. Mebone Fraud uses Bayesian Networks. Each technique has owns its pros and cons:

Requirement	ANN	BBN	Rules
Analyst knowhow	-	YES	YES
Analysis Performance	YES	YES	N/A
Learning Rate	YES	YES	N/A
Feedback	-	YES	YES
Reaction to change	-	YES	N/A
Fraud detection	YES	YES	YES

Analyst Know How. Usually, Fraud analysts retain lots of knowledge and intuitions from their everyday work. It is important to have the possibility to take advantage of this knowlegde using the ability of the system to understand it. Mebone Fraud allows this knowledge transfer using rules expressed in the analyst natural language. The system takes care of building syntactically correct rules avoiding misleading errors. The bayesian net is constructed using the analyst knowledge also, but the net proposes the analyst, ways to make a better net design. This is not possible using neural nets.
Analysis Performance. At peak moments, an authorization system can manage hundreds of operations per second. Any operation is a potencial fraud, so the system must cope with a high analysis performance. Both techniques (BBN and ANN) accomplish successfully this task. In the case of rules, the analysis time grows because of the amount of information being analyzed (real time and historical data).
Learning Rate. Automatic systems learn from past experience. Both systems (ANN&BBN) need to scan past data to tune their fraud sensibility. While Neural nets need lots of time (even retraining) Bayesian Networks are much faster.
Feedback. It is important for an automatic learning mechanism to explain its decissions. In the case of rules, the explanation is straightforward: one is good if it finds out fraud. BBN gives clues about the fraud detection indicating which of the variables used is responsible of the changing fraud pattern. Looking at that variable it can be found which of its possible values is increasing fraud. For example if one of the variables is "Country", the BBN indicates not only that the fraud is increasing abroad, it also indicates which country is responsible. ANN keeps this information hidden in its data structure.
Reaction to change. It is crucial to find out a fraud pattern as soon as possible. Every thechnique has its own inertia. This is inevitable so the system must reduce this inertia as much as possible. The inertia in rules resides in the analyst mind. The inertia in ANN and BBN is inside its data structures. BBN are much quicker than ANN in detecting a new pattern. In some cases the change in pattern is so deep that the ANN must be retrained, consuming lots of human and computer resources.

Graphical Interface is simple, based on the "Overview and zoom" paradigm. The alarm window presents quickly all the information that the analyst need in a simple way. Alarms are prioritized within each window. Lower-priority alarms do not appear in the visible portion of the alarm window.
Mebone Fraud users can drill down to the alarm details with a single click. One mouse action brings up an elaborate table with data about the object being analyzed. The analyst can build different rule sets one for each detection strategy and all of them can be activated at the same time. Analysts using this tool have a cadre of tireless automated assistants searching for fraud 24 hours a day: Intelligence Amplification.
Mebone Fraud assistance for rule developers consists on a text-based grammar-driven real time parser/editor called Rule Builder. Sophisticated real time parsing is used to present all grammatically valid alternatives as the rule is written, and all valid alternatives when the rule is being edited. Syntax error cannot be made using this approach. It provides a rich extensible language that can express a wide range of fraud preventing ideas in a natural way.
All th complex information is presented in graphics to have a quick visual understanding of what is happening, allowing to make fast, real time decisions.

Fraud analytics often involve many measurements made over many days. An example is the average activity of a merchant over the last three months.Without access to historical data, filling up this analiytic would take many hours, which is clearly unacceptable. Many attempts to use general purpose AI tools in fraud detection assume away the nasty details that must be accomodated in a real time system. Mebone Fraud includes specialised database engines to accomplish complex historial calculations in seconds over the last three months. Information can be retrieved from five years in the past.

On the side of automatic detection, the system must run faster than the authorization mechanism. While the authorization makes its own decisions based on card limits, for example, can also launch asynchronous requests to the anti fraud system to get its fraud evaluation. Mebone Fraud can also be integrated with SMS brokers to communicate with the cardholder at the moment of purchase.

Learning Evolution

Evendor

Mebone Fraud enables the possibility to catch fraud in electronic transactions even before the fraud is commited.

The combination of state of the art techniques, makes Mebone Fraud an unique tool for fraud detection.