ggplot2无法访问tbl_df数据及LDA模型主题预测准确率对比问题

Hi Chris,

Let's tackle both your LDA model comparison workflow and the tbl_df/ggplot2 issue you're facing:

Here's a structured breakdown of your process with actionable code snippets to implement each step:

• Document-level topic assignment (gamma-based):
  For each document, extract the topic with the highest gamma value (posterior probability of the document belonging to the topic). Using the topicmodels package in R, this can be done with:

  # Assuming lda_model is your trained 15-topic LDA model
  # Quick method to get top 1 topic per document
  doc_gamma_topics <- topics(lda_model, 1)
  
  # If you need manual gamma extraction for verification:
  gamma_df <- as.data.frame(lda_model@gamma)
  doc_gamma_topics <- apply(gamma_df, 1, which.max)
  

• Token-level topic prediction (RTM method):
  After implementing Chang & Blei's (2009) Relational Topic Model to get token-level topic assignments, aggregate to the document level by selecting the most frequent topic per document:

  # Assuming token_topics is a tibble with columns: document_id, token_topic
  doc_rtm_topics <- token_topics %>%
    group_by(document_id) %>%
    count(token_topic) %>%
    slice_max(n, n = 1) %>%
    pull(token_topic)
  

• Merge results for comparison:
  Combine both document-level topic assignments into a single dataset to evaluate prediction accuracy:

  comparison_df <- tibble(
    document_id = seq_along(doc_gamma_topics),
    gamma_assigned_topic = doc_gamma_topics,
    rtm_assigned_topic = doc_rtm_topics
  )
  

Tibbles (tbl_df) play nicely with ggplot2, but common missteps cause extraction issues. Here are quick fixes:

• Use tidyverse syntax directly: You don't need to "extract" columns explicitly—ggplot2 works natively with tibbles. Reference columns by name in the aes() call:

  # Example: Plot distribution of gamma-based topics
  ggplot(comparison_df, aes(x = factor(gamma_assigned_topic))) +
    geom_bar(fill = "#2c3e50") +
    labs(x = "Gamma-Assigned Topic", y = "Document Count")
  

• Convert to base data frame if needed: If you prefer base R column extraction, convert the tibble first:

  comparison_df_base <- as.data.frame(comparison_df)
  # Now you can use syntax like comparison_df_base[, "gamma_assigned_topic"]
  

• Ungroup if your tibble is grouped: If you have leftover grouping from prior dplyr operations, ungroup before plotting:

  comparison_df <- comparison_df %>% ungroup()
  

Let me know if you need adjustments for your specific model setup or dataset structure!

内容的提问来源于stack exchange，提问作者Chris T.