Page 432 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Model Training Several AI and ML techniques are employed to identify key factors contributing
and Fine-Tun- to delays in the RSD process. These include:
ing • Linear Regression: A statistical ML technique used to quantify the rela-
tionship between processing time and various predictors such as case
characteristics, resource allocation, or workflow stages. It helps in identi-
fying which variables are significantly associated with longer delays.
• Regression Trees and Decision Trees: These tree-based models provide
a visual and interpretable way to understand how different combinations
of variables (e.g., country of origin, type of procedure, staff availability)
contribute to variations in processing time.
• Random Forests: An ensemble learning method that builds multiple deci-
sion trees and aggregates their outputs for improved prediction accuracy.
Random forests help in ranking the most important predictors of delays
and capturing complex patterns that may not be evident in simpler models.
Model Fine-Tuning plays a critical role in ensuring that these ML models are
both accurate and generalisable. Techniques such as cross-validation, hyper-
parameter optimisation, and regularisation are used to adjust the models for
better performance. This process helps avoid overfitting, improves predictive
accuracy, and ensures that the insights derived from the models are robust
and actionable.
In addition to these ML methods, AI-driven process mining is used to analyse
event logs and uncover the actual workflows within the RSD process. This
technique provides insights into bottlenecks, inefficiencies, and deviations
from standard operating procedures, offering a dynamic, data-driven view of
how cases move through the system.
To avoid the risk of exposing personally identifiable information through
reverse engineering of our ML model, particularly through the data used to
train it, data will be encrypted and some variables eliminated from the model.
This will provide a minimum dataset (data minimisation) with enhanced privacy
and data protection.
Testbeds or Pilot deployments are expected to take place in the next stage of the proj-
Pilot Deploy- ect, targeting specific locations in which RSD is conducted across a range of
ments geographical areas.

Code reposito- Not available
ries

2 Use Case Description

2�1 Description

RSD is the process through which governments or UNHCR decide if a person seeking
international protection qualifies as a refugee under international, regional, or national law.
This process is crucial for ensuring that refugees can exercise their rights under international
law. Each year, UNHCR carries out RSD procedures under its mandate in around 50 countries
where fair and efficient asylum systems are not in place.

396