development of early warning systems to detect, predict

Ensuring the Integrity of the European food chain

Development of early warning systems to detect, predict and assess

food fraud

Hans Marvin, Bram Steen & Yamine BouzembrakRIKILT Wageningen UR, Wageningen, the Netherlands

Rabin NesloUniversity Medical Center Utrecht , the Netherlands

Partner(s) logo(s)

Outline

Data sources of food fraud: development of European Media Monitor (EMM)

Prediction of food fraud: Bayesian Network (BN) modelling approach

3rd FoodIntegrity Conference, Prague, 6-7 April 2016

WP8 objectiveDevelop a structured approach for collecting and analysing information regarding potential drivers of the EU food chain fraud events and frequency of fraud incidents for commodities

Existing: RASFF, UPS, EMA

Develop new: EMM

Data sources

Application of Bayesians Network modelling

Prediction models

ENDUSER (industry, authorities)


https://www.ukoug.org/membership-new/case-studies-testimonials/case-studies/

https://www.ukoug.org/membership-new/case-studies-testimonials/case-studies/

The Europe Media Monitor (EMM) provides advanced analysis

systems for monitoring of both traditional and social media.

EMM applies text mining techniques to screen different types of

media on the world wide web: websites, databases, blogs, ..etc.

EMM is updated every 10 minutes, 24 hours per day.

EMM gathers reports from news portals world-wide in 60

languages.

EMM contain 3 portals: NewsBrief, NewsExplorer and MedISys

(http://emm.newsbrief.eu/overview.html)

EMM; characteristics


MedISys: public health related topics

http://medusa.jrc.it/medisys/homeedition/en/home

No collection of publications on food fraud


EMM Food Fraud Filter Design steps

Step 4: Evaluation and improvement of the filter.

Analyse the articles Relevance evaluation Key words improvement

Step 3: The design of the food fraud filter in EMM

EMM system

Step 2: Validation of the keys words by Food fraud experts

Prof. Saskia Van Ruth (RIKILT) Dr. Hans Marvin (RIKILT) Dr. Karen Everstine (USA)

Step 1: Definition of food fraud key words

Scientific articles Food Fraudarticles USP Database EMA Database RASFF


News papers Blogs Databases Websites

Text mining tool

- Updated every 10 minutes.- 24/7- 60 languages

- Automatic retrieval of reports- Automatic data storage- Automatic data processing

- Data visualisation using

EMM Food Fraud Filter Design steps

3 4

1

2

- 6000 websites- ...etc

- 600 keywords.- 8 languages.- ...etc


Food fraud reports (overall) in MedISys (period September 2014 to December 2015; N = 1114)

Data visualisation using ArcGIS 4

Food fraud reports (milk) in MedISys (period September 2014 to December 2015)

Data visualisation using ArcGIS 4

EMM results

Developed in September 2014 Tested in the period September 2014 to

December 2015 Number of articles collected 1144 Number of relevant articles ca. 75% Will be public available by end of April 2016


Food fraud alerts/ reports: a comparison

NEW

Outline

Data sources of food fraud: development of European Media Monitor (EMM)

Prediction of food fraud: Bayesian Network (BN) modelling approach



Can we help the customs controller to decide what type of

fraud should be checked?

Food fraud is reported in Rapid Alert System for Food and Feed (RASFF)


Prediction of food fraud type using BN

• All notifications reported in the RASFF database under the hazard category “adulteration/fraud” from the period 01/01/2000 to 31/12/2013 (N = 749).

• Machine learning technique: expectation-maximization-algorithm.

1. Variable identification

3. BN model validation

Variable name Node name StatesFraud type Fraud HC, Illegal-importation, Tampering, CED, Expiration Date,

LabellingNotification type Notification information, border-rejection, alertProduct category Product alcoholic, molluscs,..., wildNotification year Year 2000, 2001,..., 2013Notifying country Notified United kingdom, Portugal,...,Austria (i.e. member countries to

RASFF)Country of origin Origin United States, Japan,...,Brazil (i.e. countries from which the

product was imported)

2. Learning the BN model

• RASFF food fraud notifications reported in 2014 were used to validate the BN model (N=88).

Predicted 80% of cases correctly


Modelling of Fraud in RASFF=> statistic relationships between all parameters


Applications of such BN models

Provides understanding of relationships between all parameters

Supports the evaluation of the effects of mitigations measures (scenario’s)

Allows forecasting/ prediction


Extended BN model that connects many drivers and parameters relevant to food fraud

Some examples

Economic drivers:• Prices of the fraudulent product at the time of detection• Price spike around the period of detection• Trade volumes of the product between the country of

detection and country of origin• Complexity of the food chain

Parameters of the country of origin & detection• Indices: corruption index, food safety index, governance

index, legal system index, press index, human development index and technology index.


BNs in holistic approach of food safety linking 36 data sources (18 databases and 8 expert judgements)


Extended BN model to assess food fraud


The type of food fraud may depend on country of origin


Conclusions

An EMM filter has been created that collects media reports on food fraud worldwide and will be online end of April 2016

Automatic retrieval of these reports from the EMM filter has been realised

Models based on Bayesians network can be used to predict the type of food fraud as reported in RASFF

BN models can be used to evaluate relationships between food fraud parameters


The project has received funding from the European Union’s Seventh Framework Programme for research, technological development and demonstration under grant agreement No. 613688.

www.foodintegrity.eu


development of early warning systems to detect, predict

Documents