Understanding and Mitigating Biases in Big Data
Stéphanie LOPEZ, Ph.D.
Data Scientist (PhD), Principal Investigator, wants to meet the daunting challenges of Lung Cancer Screening
In the era of digital transformation, big data plays a pivotal role in shaping industries, driving innovation, and informing decision-making. As massive amounts of data are generated and analyzed daily, big data applications have the potential to revolutionize our lives—from improving healthcare outcomes to optimizing business operations.
However, with this immense potential comes the responsibility to understand and address the biases that can emerge in big data. These biases, if unaddressed, can lead to discriminatory practices and faulty conclusions. As our roles evolve with these technologies, it is crucial for us, as both users and data providers, to recognize and mitigate these challenges.
This article delves into the dimensions of big data, explores the sources of biases, and emphasizes the importance of tackling these issues to ensure fair and ethical use of technology in our increasingly data-driven world. Understanding these stakes is essential to harness the true power of digital transformation responsibly and equitably.
Quick Reminder About Big Data Applications
Big data applications are characterized by the 5Vs: Volume, Velocity, Variety, Veracity, and Value. These dimensions define the scope and challenges of big data, and understanding them is crucial in addressing biases.
Volume
Volume is perhaps the most defining characteristic of big data, referring to the sheer quantity of data or the total bytes it consumes. This poses long-term challenges regarding storage capacity and scalability. For instance, Between 2020 and 2025, the total amount of global healthcare data is projected to increase from 2,300 to 10,800 exabytes. Read
for further information.
Velocity
Velocity pertains to the speed at which data is generated, collected, and processed. For AI applications to provide relevant real-time answers, they must handle data with high velocity efficiently.
Variety
Variety impacts biases directly. Diverse datasets are essential to avoid creating discriminatory AI applications. For example, AI used in hiring processes may unintentionally exacerbate gender or racial biases if the training data isn't sufficiently varied.
Some examples of biased algorithms due to a lack of representativity in data
Veracity
Veracity involves ensuring the accuracy and reliability of data. In medical fields, it’s crucial that data, like thoracic scans, are correctly associated with medical records. Qualitative data, such as patient feedback, should have an uncertainty measure to mitigate impact on algorithm training.
See this illustration of false positive and false negative cases.
Value
The value dimension raises critical questions about the quality versus quantity of data. Is it better to use a large dataset with general veracity or a smaller, thoroughly vetted one? Additionally, the carbon footprint of AI should be considered. A more extensive dataset demands more energy, so we must weigh the performance gains against environmental costs.
Biases in Big Data
Biases in big data can arise from multiple sources, and understanding these is essential for creating fair and effective AI systems. For those who speak French, I recommend reading Agnés Chambion's insightful post on cognitive biases here. For others, a tool like DeepL can help translate it.
Data Bias
Data bias occurs when the dataset used to train the AI is not representative. This can result in AI models that perpetuate existing prejudices or create new forms of discrimination.
Model Bias
Model bias happens when the algorithms themselves are biased. This could be due to the way they are programmed or the assumptions built into their design.
领英推荐
Interpretation Bias
Interpretation bias arises when the outputs of AI models are misinterpreted, leading to incorrect conclusions or decisions.
What about generative AI?
The emergence of generative AI algorithms such as GAN lead to powerful tools but also challenges in ethics.
Generative Adversarial Networks (GANs) are powerful tools capable of creating realistic synthetic data, used for improving image resolution, image denoising and image reconstruction in medical image analysis for example. In addition, due to regulatory constraints (GDPR), few data are accessible and one can generate synthetic data that can be used to augment training datasets for machine learning models, thereby improving their performance and robustness.
But they present significant ethical challenges. These include potential misuse for spreading misinformation, privacy concerns due to generating data resembling real information, risks of embedding biases from training data, and implications for security and creative industries. Addressing these issues requires careful regulation, transparency in AI development, ethical guidelines for data use, and ongoing research into fairness and accountability in AI applications. Balancing the benefits of GANs with ethical considerations is crucial to harness their potential responsibly.
Advice for Mitigating Bias
Integrate Final Users
Final users possess domain knowledge and can help verify datasets and evaluate AI results. This is crucial, especially with open-source datasets from AI challenges. For instance, the BraTS dataset used for glioma detection included incorrect grade classifications that Dequidt et al. meticulously corrected.
Learn how to use an empathy map.
Human-AI Collaboration
Remember, the goal is human-AI collaboration, not competition. While AI can outperform humans in some tasks, humans ensure reliability and accountability. Studies indicate that AI can sometimes nudge humans toward lower sensitivity, emphasizing the need for careful integration .
Empower Users
Users should understand AI mechanisms and potential biases. Transparency about AI capabilities and limitations is essential. AI is not magic; educating citizens helps foster a realistic understanding of its role and impact.
Validate with Real-Life Data
Always test AI systems with real-life data to ensure their validity and reliability.
For example, the MICCAI Brain Tumor Segmentation (BraTS) dataset is well-known dataset for machine-based deetection. It also provides grade prediction. But in this article, the research team higlights a key limitation: it does not differentiate between WHO-defined LGG and HGG. Instead, it groups grades I, II, and III as "lower grade gliomas," while its HGG category only encompasses grade IV glioblastoma multiforme. Consequently, the researchers propose a new classification scheme aimed at enabling AI research teams to train classification algorithms using their annotation set, to align with clinical reality.
Cost-Benefit Analysis
Consider the cost-benefit ratio of your AI applications. Is the incremental performance gain worth the extensive data collection effort? Sometimes, AI might be better suited to identifying biases rather than making high-level decisions.
In this article, we can read the results of a study conducted by researchers at the University of Massachusetts "to determine how much energy is used to train certain popular large AI models. According to the results, training can produce about 626,000 pounds of carbon dioxide, or the equivalent of around 300 round-trip flights between New York and San Francisco –?nearly 5 times the lifetime emissions of the average car."
Conclusion
In conclusion, addressing biases in big data requires a comprehensive approach involving diverse datasets, accurate data validation, user empowerment, and a careful balance between data quantity and quality. By integrating human expertise and maintaining transparency, we can harness the full potential of AI while minimizing the risks of bias.
While technological advancements offer promising avenues to mitigate biases in big data, true progress necessitates a dual approach: addressing biases both technologically and socially. By cultivating diverse teams encompassing gender, cultural, and interdisciplinary perspectives, we can identify biases, develop inclusive AI tools, and ensure that technological innovations benefit society as a whole. Embracing diversity not only enhances the accuracy and fairness of AI systems but also fosters a more equitable and empathetic future where technology serves everyone equitably.
For additional insights, you can read this article on globalcio.com, where the author asks chatGPT to propose 5 new V to big data applications:
Bravo chère Stephanie et merci pour ce partage ????