Researchers develop tools to help data scientists make features used in machine learning models easier to understand for end users.
The explanatory methods that help users understand and trust machine learning models often describe the extent to which certain features used in the model contribute to its predictions. For example, if a model predicts a patient’s risk develop heart diseaseA physician may want to know how strongly a patient’s heart rate data affects that prediction.
But if those features are so complex or complex that the user can’t understand them, is the explanation method any good?
MIT researchers are trying to improve the interpretability of features so that decision makers will feel more comfortable using the outputs of machine learning models. Based on years of field research, they developed a classification method to help developers create features that will be easier for their target audience to understand.
“We found that in the real world, even though we are using modern ways to interpret machine learning models, there is still a lot of confusion stemming from features, not from model itself,” said Alexandra Zytek, a PhD student in electrical engineering and computer science and lead author of an introductory paper to taxonomy.
To build the classification method, the researchers identified the attributes that make the features interpretable for five types of users, from artificial intelligence experts to those affected by the project. prediction of the machine learning model. They also provide guidance on how modelers can convert features into formats that are easier for ordinary people to understand.
They hope their work will inspire modelers to consider using interpretable features early in the development process, rather than trying to work backwards and focus on post-factual interpretation.
MIT co-authors include Dongyu Liu, a postdoc; visit Professor Laure Berti-Équille, director of research at IRD; and senior author Kalyan Veeramachaneni, principal research scientist at the Laboratory for Decision and Information Systems (LIDS) and leader of the Data to AI team. They feature Ignacio Arnaldo, a lead data scientist at Corelight. Research published in the June issue of the Association for Computing Computing Special Interests Group on Knowledge Discovery and Data Mining peer-reviewed Discovery Newsletter.
Lessons in the real world
Features are input variables fed to machine learning models; they are usually plotted from the columns in the data set. Data scientists typically select and process features for models, and they are primarily focused on ensuring features are developed to improve model accuracy, not on whether humans decision makers can understand them or not, explains Veeramachaneni.
For several years, he and his team worked with decision makers to identify usability challenges of machine learning. These domain experts, most of which lack machine learning knowledge, often distrust models because they don’t understand the features that affect predictions.
For one project, they collaborated with clinicians in an ICU hospital, who used machine learning to predict the risk a patient would face complications after heart surgery. . Some features are presented as aggregated values, like trends in a patient’s heart rate over time. Although features coded in this way are “ready models” (models that can process data), clinicians do not understand how they are computed. They wanted to see how these composite characteristics related to baseline values, so they could identify abnormalities in a patient’s heart rhythm, Liu said.
In contrast, a group of academic scientists preferred synthetic features. Instead of having a feature like “number of posts students have made on discussion forums”, they want to have related features grouped together and labeled with terms they understand. , such as “join”.
“With interpretability, one size does not fit all. As you go from one area to another, there are different needs. And interpretability itself has many levels,” said Veeramachaneni.
The idea that one size doesn’t fit all is key to researchers’ classification. They identify attributes that might make features more or less understandable to different decision makers, and outline which attributes are likely to be most important to specific users.
For example, machine learning developers might focus on having model- and predictive-compatible features, which means they are expected to improve the model’s performance.
On the other hand, decision makers without machine learning experience may be better served by human-powered features, which means they are described in a way that is natural to the user and easy to understand, which means they refer to real-world metrics the user can reason about.
Taxonomy says if you are creating interpretable features, to what extent are they interpretable? You may not need all levels, depending on the type of domain specialist you are working with,” says Zytek.
Put interpretability first
The researchers also outline feature engineering techniques that a developer can use to make features easier to understand for a particular audience.
Feature engineering is a process in which data scientists transform data into a format that a machine learning model can process, using techniques such as data aggregation or normalization of values. . Most models also cannot handle categorical data unless they are converted to codes. These transformations are often nearly impossible for ordinary people to decompress.
Zytek says that making the features conceivable may involve undoing some of that encryption. For example, a common feature design technique organizes data intervals so that they all contain the same number of years. To make these features easier to understand, one can group the ages using human terms, like infant, toddler, child, and teen. Or instead of using a transformed feature like average pulse rate, an interpretable feature could simply be actual pulse rate data, adds Liu.
“In many areas, the trade-off between interpretable features and model accuracy is really small. For example, when we worked with child welfare screeners, we retrained the model to only use features that met our definitions of interpretability and performance loss. yield is almost negligible,” said Zytek.
Building on this work, the researchers are developing a system that allows model developers to deal with complex feature transformations in a more efficient way, to generate human-centered explanations. for machine learning models. The new system will also convert algorithms designed to interpret model-ready data sets into formats that decision makers can understand.
Written by
Source: Massachusetts Institute of Technology
