How to improve data quality to better meet the needs of AI projects

​Translator | Cui Hao

Reviewer | Sun Shujuan

Opening Chapter

In today's society, the development of artificial intelligence has become the focus of global enterprises and governments. However, another problem closely related to artificial intelligence has been ignored: poor data quality.

Artificial intelligence algorithms rely on reliable data to produce optimal results—if the data is biased, incomplete, inadequate, or even inaccurate, the consequences can be devastating.

Artificial intelligence systems that identify patients’ diseases are a good example of poor data quality leading to adverse consequences. When there is insufficient data, these systems can produce erroneous diagnoses and inaccurate predictions, leading to misdiagnosis and delayed treatment. For example, a University of Cambridge study of more than 400 tools used to diagnose Covid-19 found that AI-generated reports were completely unusable due to the use of flawed datasets.

In other words, if the data isn’t good enough, AI initiatives will have devastating real-world consequences.

What does “good enough” data mean?

There has always been a huge debate about what is “good enough” data. Some say good enough data does not exist. Others say that “too good” data can lead to analysis paralysis (Translator: should refer to overfitting) – while HBR points out outright that bad information will cause machine learning tools to fail to work.

At WinPure, good enough data is defined as “complete, accurate, valid, and can be used with confidence in risky business processes, the level of data depends on personal goals and business circumstances.”

Most companies struggle with data quality and governance, even though they won’t admit it. This torture continues to increase the tension of the project and overwhelm them. You can imagine that they are under tremendous pressure to deploy artificial intelligence plans to maintain a competitive advantage. Sadly, issues like dirty data are unlikely to be discussed in the boardroom until it causes the project to fail.

How does bad data affect artificial intelligence systems?

Data quality issues arise at the beginning of the process when algorithms learn based on training data. For example, if an AI algorithm is fed unfiltered social media data, it will extract abuse, racist comments, and misogynistic remarks, as demonstrated by Microsoft's AI bot. More recently, the inability of AI to detect dark-skinned people has also been blamed on problems with training data.

How does this relate to data quality?

Lack of data governance, low awareness of data quality, and siled views of data are the main culprits of poor data quality.

what to do?

When companies realize there is a problem with data quality, they panic about recruiting. By blindly hiring consultants, engineers, and analysts to diagnose and clean data, hoping to solve the problem as quickly as possible. Unfortunately, as the months passed, the problem didn't seem to go away, despite millions of dollars spent. Taking a knee-jerk approach to data quality issues rarely helps.

Real change starts at the grassroots.

If you want your AI/ML project to move in the right direction, take these three key steps.

Recognize and acknowledge data quality issues

First, assess data quality by building a culture of data literacy. Bill Schmarzo is a powerful voice on this, recommending using design thinking to create a culture where everyone understands and contributes to the organization's data goals and challenges.

In today’s business environment, data and data quality are no longer the sole responsibility of IT or data teams. Business users must be aware of issues such as dirty data issues and inconsistent and duplicate data.

So, start by making data quality training a valued organizational effort and empowering teams to identify poor data attributes.

With the checklist below, you can use it to track data quality.

Data Health Checklist

• How to capture, store and manage data?
• How many data sources are connected to your central database, and how well is the data disseminated?
• How well are you managing your data? Have you implemented data governance standards? How much of the data is structured, semi-structured or unstructured?
• How much do you spend manually fixing your data compared to automated data management? How do your teams coordinate with each other when accessing and processing data? Are there frequent internal conflicts between IT and business users?
• What is your data quality status? Is your data timely, complete, accurate, unique and follows standardized rules?

Develop a plan to meet quality metrics

Businesses often make mistakes when it comes to data quality. For example, data analysts are hired to complete routine data cleaning tasks instead of focusing on planning and strategic work. Some businesses use data management tools to cleanse, deduplicate, consolidate and purge data without a plan. Unfortunately, tools and talent cannot solve problems in isolation. Strategies that meet the data quality dimensions are the fundamental solution to the problem.

The strategy must address data collection, labeling, processing, and matching data to AI/ML projects. For example, if an AI recruiting program selects only male candidates for technical positions, then the program’s training data is clearly biased, incomplete (not enough data on female candidates is collected), and inaccurate. Therefore, this data does not serve the true purpose of the AI ​​project.

Requirements for data quality extend beyond the daily tasks of cleaning and repairing data. So, data integrity and governance standards need to be set before starting a project. It saves projects from falling into failure!

Ask the right questions and set accountability

There is no universal standard for “good enough data or level of data quality.” Instead, it all depends on the enterprise’s information management systems, data governance guidelines, knowledge of team and business goals, and many other factors.

But before starting the project, there are a few questions to ask the team:

• What are the sources of our information and what are the methods of data collection?
• What issues can impact the data collection process and threaten positive outcomes?
• What information does the data convey? Does it meet data quality standards (i.e. the information is accurate, completely reliable and constant)?
• Is the designated person aware of the importance of data quality and low quality?
• Are roles and responsibilities defined? For example, who needs to maintain a regular data cleaning schedule? Who is responsible for creating master records?
• Is the data fit for purpose?

Ask the right questions, assign the right roles, implement data quality standards and help your team address challenges before they arise!

Summary

Data quality is more than just fixing typos or errors. It ensures that AI systems are not discriminatory, misleading or inaccurate. Before launching an AI project, it is necessary to address data quality challenges by addressing flaws in the data. Additionally, launch an organization-wide data literacy program to connect each team to the overall goal.

Translator Introduction

Cui Hao, 51CTO community editor and senior architect, has 18 years of software development and architecture experience and 10 years of distributed architecture experience.

Original title:​​Is Your Data Good Enough for Your Machine Learning/AI Plans?​​, Author: Farah Kim​

The above is the detailed content of How to improve data quality to better meet the needs of AI projects. For more information, please follow other related articles on the PHP Chinese website!