CS229 Final Project: An AI approach to fighting financial crime using the UK companies register

1. Corporate informations are downloaded from Company `house, the UK companies registrar.

2. Data from companies is of poor quality and required corrections: (i) names of countries are not standardized and sometimes name of cities or postcodes are entered instead. Use NLP to recognize countries names and clean-up the database. (ii) identical names of individuals and corporate entities are treated wrongly as separate entities in the Company House Database, requiring multiple searches to take place to correct this issue.

3. We convert the raw inputs of the model (nationality and status corporate/individual of the officers and the beneficial owners) for each corporate using a bag of words approach into a vector of size 492, representing (245 countries+ one unknown)* 2. We multiply by 2 to represent an individual officer and a corporate officer.

Given the number of countries, we also created two other vectors: (i) a vector with regions and also included the European Union list of tax havens (American Samoa, Belize, Dominica, Fiji, Guam, Marshall Islands, Oman, Samoa, Trinidad and Tobago, United Arab Emirates, Vanuatu, US Virgin Islands the "Blacklist") as at 17 May 2019: https://ec.europa.eu/taxation_customs/sites/taxation/files/eu_list_update_17_05_2019_en.pdf (ii) a vector with a mix of regions and individual countries (aggregating Africa, Antarctica, North America, South America, Asia, other Oceania and Middle-East), and providing detailed countries for Europe, Eastern Europe and Central Asia, and Caribbean.

4. Using Snorkel, we created four labeling functions and generated the final label for each data point.

5. The dataset has been split between train 90%, dev 5%, and test 5%.

6. The following models have been applied on the train and dev datasets. -SVM: sigmoid, gaussian, 5th and 10th degree polynomial kernels -Logistic regression -NN: containing just fully connected layers with 2, 5 and 10 hidden layers -CNN: LeNet-5 variation for 1D data

7. The most promising models were CNN, 2 hidden layers NN and SVM/Logistic regression

8. Discussion and conclusion We could show that a model can detect patterns in the data indicating companies andlimited partnerships more likely to be involved in money-laundering using the residence and legalpersonality (individual or corporate) of its officers and beneficial owners.