Mitigate Bias - Data Intervention

1. Mitigate Bias - Data Intervention

Now we will look at another set of tools and techniques to actually mitigate bias. Let's follow the same format: data and model. You can mitigate bias issues by applying several techniques, although the actual approach depends on the use case, usually, it involves multiple methods. First, let's talk about methods to mitigate biases by underpinning data. Teams should continuously improve and diversify the data sources and in notations included in their training data sets as data gaps can be harmful. This is due to over-representation and/or insufficient training data for specific subgroups that might lead to drastic performance differences across various subgroups. To improve the training data for machine learning models, you should aim for these goals: Ensure all groups have sufficient data representation, ensure data does not contain harmful labels or media that are outside your intended use case, such as pejorative labels or poor quality data, ensure your data does not skew towards harmful biases like stereotypes or bias correlations as it may lead to unintended correlations in your model, and ensure your data does not have unintended biases in the authorship or who collected the data, the way it was described, or who annotated or labeled it, or the content of the data itself, or what was captured. To achieve these goals, here are some techniques that help: Refine your data collection pipeline when you identify bias, especially when associated with data filters in the pipeline, resample the dataset to balance data, augment with more data, like other existing datasets, synthetic data, or new data collection, relabel your data by removing harmful labels, updating labels to current standards, and adding labels that were missed during the initial notation effort. Let's dive into each of the techniques mentioned. First, review and refine the pipeline when you identify bias in your data. When there are filters to include or exclude data in your data collection pipeline, ensure that you think through the implications of what is being filtered in or out, both intended and unintended. Every time you add a filter to your data, you heavily bias it in some way. Refining the pipeline can help achieve higher quality and more diverse data. There are also limitations on refining the pipeline. Best practices about fair data collection depend on features. For example, if you want to collect a fair human face image dataset, you can use the Monk Skin Tone or MST scale, developed by Dr. Ellis Monk in partnership with Google. The MST scale provides a 10-shape scale of skin tones that can be used to evaluate datasets and ML models for better representation. However, this kind of best practice might not be available for some sensitive features, and it can be challenging to define it from the beginning. Next, balance the data by resampling the dataset. For example, you can upsample or downsample the minority or majority group examples with existing data. Resampling can occur at various stages of the ML workflow and often depends on the use of the dataset. The benefit of this technique is that it's relatively cheap, no new data is required if you are upsampling underrepresented groups. However, be aware of a limitation potential to overfit highly underrepresented groups and pay attention to amplification, leading to stereotyping bias. Here is an example of balancing data. Imagine you are training a machine learning model using a dataset with flight attendant data. In this dataset, only a handful of male flight attendants exist in the data. For this case, it would be beneficial to consider downsampling the female attendance group or upsampling the other group. A limitation of this technique is that each individual example could be over-represented. To offset this, it's recommended to explore and collect more data for the minority group and augment with data. Now, augmenting with other existing datasets, synthetic data, or new data collection is an important technique for mitigating bias in data. However, note that augmenting with data enlarges the data pool, whereas balancing data by resampling operates on the existing data, and by collecting data from multiple datasets, you could increase representation of minority subgroups or failure cases. To add new examples for underrepresented minority subgroups, you can generate synthetic data. If budget and time are not pressing concerns, you can also add new examples by collecting new data. There are a few benefits to augmenting with more data. Additional data with improved fairness characteristics might be already available for widespread use. By increasing minority subgroup representation, you can balance the data without downsampling majority subgroups. Augmenting with synthetic data can be potentially less expensive than curating a new dataset to fill in these gaps. However, each of the techniques has its own limitations. For [inaudible] product uses with rare data attributes, usage of additional out-of-domain data may introduce technical complexity on the relevant MF tasks to overcome domain gaps. Training on synthetic data can be problematic, especially for simulator-generated data. For example, it's important to pay attention to whether a trained model is treating synthetic images differently from natural images in order to ensure the model doesn't cheat by simply changing scores for synthetic images, and oversampling might cause over-representation for the minority group, like upsampling. Now, of all the listed techniques, new data collection is the most expensive and time-consuming. Coming back to our previous example, we could also try a different approach to improve representation and reduce bias in our training model by augmenting our data with a new dataset with more male flight attendants. Lastly, relabel your data by removing harmful labels, updating labels to current standards, and adding labels that were missed during the initial annotation effort. This method can improve overall data quality. However, this may not change underlying sample bias. A real-life example of how relabeling data can be used to reduce bias in machine learning models is through sentiment analysis. Imagine you're training a machine learning model to classify movie reviews as positive or negative. Your training data might contain reviews with biased language, such as using stereotypical phrases or making generalizations about certain groups of people. To reduce bias, you could relabel some of the reviews by having human experts carefully review them, then assign more neutral labels. This results in helping the model learn to associate sentiment with the actual content of the reviews instead of biased language.

2. Let's practice!

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.