Training your Machine Learning algo to be fair, unbiased –

Data topped with advanced algorithms like machine learning and data mining are proving to add new revenue streams in the business world. Despite the accomplishments, data, artificial intelligence or machine learning has achieved, it continues to express some amount of unfairness and bias, which affects not just business but humanity at large.If the machine algorithm is queued that a certain type of person is more likely to commit fraud then it may end up depriving that person the resources, products or services. This is because the machine may believe that a person with a certain background and ethnicity is likely to commit fraud. Speaking on how ethics in AI have become more life-impacting, Dale Vaz, Head of Engineering & AI, Swiggy said that the choice of model matters. Many new models like deep learning being highly sophisticated aren’t explainable. Meaning it is unexplainable why the model made a certain choice because of the complexity of how it works.“So sometimes you might have to go back to more explainable models which are simpler in nature and allow you to justify and substantiate on why a certain recommendation was made,” he said. He insists on the importance of feedback on the algorithms once they are deployed to make sure it is being tested. For this, Swiggy gathers sources of inputs from its end-user. They study what their customers are telling the service agents and analyze if they are happy with it.Adding to how Lowe’s India is making its algorithms fair. Vijay Nair, Senior Director- Data Analytics, Lowe’s India, said, “We use test and control approach no matter how we built the algorithm and how confident we are on the same. We have a model taking a decision and a person taking the same decision to compare and see if there is any bias. If there is any bias, we then go back to the source of data to investigate further in terms of what’s going wrong. There could be things like sampling bias that could occur so we have teams that can address the same.”According to Nair, it is extremely important to have explainable models and this is where techniques like neural networks help. Neural networks are hard to implement but can be still successful from a biased perspective. “So if you are taking a model and asking the business to implement the model, half the time the battle is won when you can explain why the black box or the ML algorithm is coming up with what it is coming up with. It can be challenging because it’s hard to explain what input leads to a certain output,” said Vijay.Subramanian M S, Head of Analytics, Bigbasket said, “We avoid black-box algorithms so we have a hierarchical set of algorithms. Most are heuristic and rule-based where the algorithm is very clear which allows to break this down and make sure that is there is no bias creeping the algorithm. We keep improving the complexities of algorithms by using a machine learning model and try avoiding any deep learning models that clearly can not be taken down into the component piece to understand how the algorithm operates,” Wherever Big Basket uses personalizations such as recommendations, they try not to use aggregate data from other sources that could let bias to creep in. They use customers’ own buying behavior to generate their own recommendations. The aggregate data only gets used in forecasting, routing that could help deliver efficiency and does not affect the customers. “Talk to internal customers and customers, get constant feedback to see if there is any bias. If there is, re-work the algorithms to avoid them. In critical situations, manual interventions are important to have therefore ensured the right decision is made,” he concluded.