Machine learning will only become more commonplace at enterprise level, but knowing the difference between black-box and white-box models is crucial to making the right decision for your organisation.
Machine learning (ML) is fast becoming a major area of interest for organisations of all types and it is helping to power everything from cybersecurity defence to recruitment and chatbots. While discussion about the benefits has piqued the interest of many businesses, deciding how to implement ML can be overwhelming.
Siliconrepublic.com spoke to Dr Ryohei Fujimaki, founder and CEO of data automation platform DotData, about the key differences between black-box and white-box ML models and the growing calls for transparency in data science.
What are the key differences between black-box and white-box models?
White-box models are the type of models which one can clearly explain how they behave, how they produce predictions and what the influencing variables are.
There are two key elements that make a model white-box: features have to be understandable, and the ML process has to be transparent. These models include linear and decision/regression tree models.
On the other hand, black-box models, such as deep-learning (deep neural network), boosting and random forest models, are highly non-linear by nature and are harder to explain in general.
With black-box models, users can only observe the input-output relationship. For example, input the customer profile then output customer churn propensity score. But the underlying reasons or processes to produce the output are not available. Black-box models often result in 1pc to 3pc better accuracy than white-box models, but you sacrifice transparency and accountability.
Why is transparency becoming increasingly important in data science?
More and more ML/AI models are being implemented in businesses to accelerate automated decision-making. As more organisations are adopting data science into their business process, there are increasing social concerns about decisions made based on personal/discriminatory information.
For example, in the use case of a loan application, race and gender should not be used to determine people’s eligibility for a loan product. Black-box models exacerbate this issue, where less is known about what influencing variables are actually driving the final decision. White-box models help organisations stay accountable for their data-driven decisions, and compliant with law and legal audits.
Traditionally, such transparency was required in only certain industries such as financial services. However, people are recently paying more and more attention to how their personal data is used, and new regulations require accountability of ML/AI models.
Are organisations moving away from black-box ML models in general? What will it take for white-box models to be implemented in a broader sense?
It varies from case to case, as both white-box and black-box models have pros and cons. In general, white-box models are preferred in core business functions and also when using personal information, where stronger accountability is required. It is important to understand their characteristics of both and to choose the right approach for your needs.
As a data scientist, it is important to balance the trade-off between interpretability and accuracy resulting from the difference between white-box models and black-box models.
The industrial best practice is always to start with white-box models wherever possible, and only to try black-box models if the 1pc to 3pc accuracy improvement is really needed at the cost of model interpretability.
How can white-box models explain their behaviour as opposed to the relative mystery of black-box decision-making?
There are two important aspects: one is model expression and the other is feature expression. For the model expression, for example, neural networks apply high-dimensional and non-linear transformations to the input, which in principle humans cannot understand.
In contrast, a linear model, as its name indicates, calculates a score based on ‘weight (coefficient) times feature value’ – the logic is simple enough for humans to understand, and the models are fairly transparent in that one can easily explain how these models generate predictions.
Additionally, white-box models produce prediction results alongside influencing variables, making prediction fully explainable. This is especially critical in situations where a model is used to support a very big business decision or to replace an existing model, and model developers have to defend their models and justify model-based decisions to other business stakeholders.
Features also have to be understandable. Data scientists are often math- and statistics-oriented and create complex features (AKA exploratory variables). Also, deep learning (neural networks) computationally generates features. However, it is not possible to understand such deep non-linear transformations. Incorporating this type of feature will make the model a black box.
How will automation tools alter data science?
While substantial investments are being made into data science across many industries, the scarcity of data science skills and resources limits the advancement of AI and ML projects within organisations.
In addition, one data science team is only able to execute several projects a year given the iterative nature of the process and the manual work that goes into data preparation and feature engineering.
In the next two to three years, data science automation platforms will be a growing trend and will impact data science in multiple ways.
Automation will help democratise the data science process, empowering the broader data and analytics team to execute on all projects. It reduces the skill barrier to data science to enable broader participation in the data science process, and frees up data scientists for higher-value tasks.