Model biases are the systematic errors or distortions that affect a machine learning model's predictions or decisions.
Below are some prevalent types of biases in the realm of Machine Learning:
1. Data Bias: Data bias occurs when the training data is biased or unrepresentative of the population.
Types of data bias:
- 1. Sampling Bias*: Biased sampling methods can lead to unrepresentative data.
- 2. Selection Bias*: Selecting a subset of data that is not representative of the population.
- 3. Confirmation Bias*: Selecting data that confirms preconceived notions.
2. Algorithmic Bias: Algorithmic bias occurs when the machine learning algorithm itself introduces bias.
Types of algorithmic bias:
- Optimization Bias: Biased optimization methods can lead to biased models.
- Regularization Bias: Biased regularization techniques can lead to biased models. Read more
- Model Complexity Bias: Complex models can be more prone to overfitting
3. Cognitive Bias: Cognitive bias occurs when human developers introduce bias into the model.
Types of cognitive bias:
- Anchoring Bias*: Relying too heavily on initial information.
- Confirmation Bias*: Seeking confirming evidence.
- Availability Heuristic*: Overestimating the importance of vivid events.
4. Model Evaluation Bias: Model evaluation bias occurs when the evaluation metrics are biased. Types of model evaluation bias:
- Evaluation Metric Bias*: Using biased evaluation metrics.
- Data Split Bias*: Biased data splitting methods can lead to biased evaluation.
5. Deployment Bias: Deployment bias occurs when the model is deployed in a biased manner. Types of deployment bias:
- Model Updating Bias*: Biased model updating methods can lead to biased models.
- Model Monitoring Bias*: Biased model monitoring methods can lead to biased models.
To mitigate model biases, it's essential to:
- Use diverse and representative data.
- Select unbiased algorithms and techniques
- Evaluate models using unbiased metrics
- Monitor and update models regularly
- Use techniques like regularization and data augmentation
Here's an example of confirmation bias:
Example:
Suppose John is a big fan of Apple products and believes that they are superior to other brands. One day, he reads an article online that compares the battery life of Apple iPhones with that of Samsung Galaxy phones. The article concludes that iPhones have longer battery life.
John is thrilled to see this result and immediately shares the article on social media, saying "See, I told you Apple is the best!" However, he fails to notice that the article only compared the battery life of one specific iPhone model with one specific Samsung model, and that the results may not be generalizable to all iPhones and Samsung phones.
Moreover, John ignores other articles and reviews that suggest Samsung phones may have better battery life in certain situations. He only seeks out and gives credence to information that confirms his pre-existing belief that Apple is superior.
What's happening here?
John is exhibiting confirmation bias by:
- Seeking out confirming evidence*: He's actively looking for information that supports his belief.
- Ignoring disconfirming evidence*: He's ignoring or downplaying information that contradicts his belief.
- Interpreting ambiguous evidence in a biased way*: He's interpreting the article's results in a way that supports his belief, without considering alternative explanations.
Consequences: John's confirmation bias leads him to:
- Reinforce his existing beliefs*: He becomes more convinced that Apple is superior, without considering alternative perspectives.
- Miss out on valuable information*: He ignores or dismisses information that could help him make more informed decisions.
- Make suboptimal decisions*: He may make decisions based on incomplete or biased information, which could lead to suboptimal outcomes.