如何通过LIME包获取变量变动对应的预测概率变化值？

Hey there! I see you've built an MLP model with caret and used lime for interpretability, and now you want to get quantitative values for how variable changes impact prediction probabilities—great goal, since those numbers make your visualizations way more concrete. Let's walk through how to do this, plus share some key literature to support your work.

一、提取量化的变量影响值 #

You can get these quantitative values in two reliable ways: leveraging LIME's built-in explanation results, or manually creating perturbed samples to calculate real model probability changes.

方法1：从LIME解释结果中推导近似影响 #

LIME works by fitting a local linear model to approximate the black-box model's behavior. The feature_weight field in the explanation output is exactly the coefficient of this linear model—it represents the change in prediction probability for the target label when the variable increases by 1 unit. You can organize this info easily:

# 整理错误分类样本的LIME解释结果
explanation_wrong_summary <- explanation_wrong %>%
  group_by(sample_id, label) %>%
  mutate(
    # 变量每变动1单位的预测概率变化（线性近似值）
    prob_change_per_unit = feature_weight,
    # 标注变量当前值与分箱范围（如果是连续变量分箱的话）
    feature_context = ifelse(
      bin_continuous,
      paste0(feature, " (当前值: ", feature_value, ", 分箱范围: ", bin_left, " - ", bin_right, ")"),
      paste0(feature, " (当前值: ", feature_value, ")")
    )
  ) %>%
  select(sample_id, label, feature_context, prob_change_per_unit) %>%
  arrange(sample_id, desc(abs(prob_change_per_unit)))

# 查看整理后的结果
print(explanation_wrong_summary)

For example, if Family has a prob_change_per_unit of 0.05, that means every 1-unit increase in Family boosts the target label's prediction probability by 5%. If your variables are normalized to a 0-1 range, a 0.1 increase would lead to a 0.5% probability lift (0.05 * 0.1)—exactly the kind of specific value you want for your visualization notes.

方法2：手动构造扰动样本计算真实概率变化 #

If you want results that reflect the MLP's actual behavior (not just LIME's linear approximation), you can modify a specific variable in a sample, re-run the prediction, and compare the probability difference:

# 选取错误分类样本中的第一个样本作为演示
target_sample <- test_data_wrong[1, ]
original_prob <- predict(model_mlp, target_sample, type = "prob")

# 模拟Family变量提升0.1的场景
modified_sample <- target_sample
modified_sample$Family <- modified_sample$Family + 0.1
modified_prob <- predict(model_mlp, modified_sample, type = "prob")

# 计算概率变化量
prob_difference <- modified_prob - original_prob
cat("Family变量提升0.1后，各类别预测概率变化：\n")
print(prob_difference)

This method uses the actual MLP model to generate results, so it's more accurate for describing how your specific model responds to variable changes.

二、相关文献参考 #

Here are key references to ground your work in interpretability research:

• Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). "Why Should I Trust You?": Explaining the Predictions of Any Classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. (The foundational paper for LIME, explaining its core logic and design)
• Molnar, C. (2020). Interpretable Machine Learning: A Guide for Making Black Box Models Explainable. (A comprehensive guide covering LIME, SHAP, and other interpretability methods—perfect for deepening your understanding)
• Pedersen, T. L. (2018). lime: Local Interpretable Model-Agnostic Explanations for Machine Learning Models. R package version 0.5.2. (The official documentation/technical note for the R lime package you're using)
• Zhang, H., et al. (2019). Evaluating the Local Explanation Methods for Deep Neural Networks. arXiv preprint arXiv:1902.03814. (Discusses evaluation metrics for local explanation methods like LIME, helping you validate the reliability of your results)

内容的提问来源于stack exchange，提问作者Lotw