• A
  • A
  • A
  • ABC
  • ABC
  • ABC
  • А
  • А
  • А
  • А
  • А
Regular version of the site

News

Why should I trust You? LIME method for machine learning tasks

On the seminar of our Research Group on November 16, Olga Silyutina told about the ways of applying the LIME method. The classical tasks of machine learning were used as examples: regression and classification, including the demonstration in both R and Python.

Today machine learning is actively used in various fields of industry and science and is also the main direction of our research group. Moreover, most of the existing models are implemented on the basis of the "black box" principle, and the decision to use them is based on the accuracy score and other aggregated metrics, which can often be misleading about the predictive ability of a particular model in real-world situations. LIME is a method that allows you to evaluate the usefulness of different models of machine learning and increase the level of confidence in them by reviewing their work on specific examples from the available sample.

Thus, with the use of LIME it is possible to check the prediction - you can look inside the model and understand which predictors contributed to the final prediction of the model for each case. Olga presented the regression model of the house price, depending on its characteristics. However, more attention was paid to the classifier of texts of medical symptoms, which were checked correlated with diagnoses of qualified specialist doctors.

At the seminar, Olya talked about a very interesting and useful topic which, perhaps, is still too early to tell in the second year of the minor in Data Science, because according to the program it is still far from machine learning. However, it seems to me that the LIME package is very helpful for training. When you do not quite understand how the models work, it is important to look at each case separately to see why the model produced a particular result. We analyzed a couple of real examples of trained models for different tasks: text processing, price forecasting, etc. There were many interesting ideas for experiments. I hope that I will be able to apply this package to my future projects.

Alexander Nikulin
2nd-year student of BA "Sociology and Social Informatics"