Skip to content

Supervised Learning Example

This example demonstrates supervised topic modeling and temporal analysis.

Example 1: Supervised Topic Classification

Dataset Description

  • Domain: Customer reviews
  • Size: 3000 documents
  • Language: English
  • Labels: 5 product categories
  • Goal: Discover category-aligned topics

Step 1: Prepare Labeled Data

CSV format with labels: 

text,label
"Great laptop with fast processor and long battery life",Electronics
"Comfortable running shoes with good arch support",Sports
"Delicious coffee beans with rich aroma",Food

Step 2: Preprocess in Supervised Mode

python prepare_data.py \
    --dataset reviews \
    --model theta \
    --model_size 0.6B \
    --mode supervised \
    --vocab_size 5000 \
    --batch_size 32 \
    --gpu 0

Step 3: Train with Supervision

python run_pipeline.py \
    --dataset reviews \
    --models theta \
    --model_size 0.6B \
    --mode supervised \
    --num_topics 15 \
    --epochs 100 \
    --batch_size 64 \
    --gpu 0 \
    --language en

Step 4: Compare Modes

Train both supervised and zero-shot for comparison:

# Zero-shot (ignores labels)
python run_pipeline.py \
    --dataset reviews \
    --models theta \
    --model_size 0.6B \
    --mode zero_shot \
    --num_topics 15 \
    --gpu 0

# Supervised (uses labels)
python run_pipeline.py \
    --dataset reviews \
    --models theta \
    --model_size 0.6B \
    --mode supervised \
    --num_topics 15 \
    --gpu 0

Results comparison:

Mode TD NPMI Label Alignment
Zero-shot 0.85 0.41 0.62
Supervised 0.83 0.38 0.89

Supervised mode achieves better label alignment with slight reduction in diversity.

Step 5: Topic-Label Analysis

import json
import numpy as np

# Load results
with open('result/0.6B/reviews/supervised/metrics/evaluation_results.json') as f:
    results = json.load(f)

# Analyze topic-label correspondence
# Topics 0-2: Electronics
# Topics 3-5: Sports
# Topics 6-8: Food
# Topics 9-11: Books
# Topics 12-14: Clothing

Example 2: Temporal Topic Evolution

Dataset Description

  • Domain: Academic papers
  • Size: 10000 documents
  • Language: English
  • Time range: 2015-2023
  • Field: Machine learning

Step 1: Prepare Temporal Data

CSV with year column: 

text,year
"Deep learning approaches for image recognition...",2015
"Transformer architectures for natural language...",2019
"Large language models and emergent capabilities...",2023

Step 2: Preprocess with Time Information

python prepare_data.py \
    --dataset ml_papers \
    --model dtm \
    --vocab_size 8000 \
    --time_column year

Step 3: Train DTM Model

python run_pipeline.py \
    --dataset ml_papers \
    --models dtm \
    --num_topics 30 \
    --epochs 150 \
    --batch_size 64 \
    --hidden_dim 512 \
    --learning_rate 0.002 \
    --gpu 0 \
    --language en

Step 4: Analyze Topic Evolution

DTM tracks topic changes over time:

Topic 5: Deep Learning (2015-2018) - 2015: convolutional, neural, network, classification - 2016: deep, learning, layers, training - 2017: residual, connections, skip, depth - 2018: architectures, design, efficient, mobile

Topic 12: Attention Mechanisms (2017-2020) - 2017: attention, mechanism, sequence, encoder - 2018: self-attention, multi-head, transformer - 2019: bert, pre-training, fine-tuning, downstream - 2020: scaling, models, parameters, performance

Topic 18: Large Language Models (2020-2023) - 2020: gpt, generation, language, model - 2021: prompting, few-shot, in-context, learning - 2022: instruction, tuning, alignment, human - 2023: emergent, capabilities, scaling, laws

python -m visualization.run_visualization \
    --baseline \
    --result_dir /root/autodl-tmp/result/baseline \
    --dataset ml_papers \
    --model dtm \
    --num_topics 30 \
    --language en \
    --dpi 300

Visualizations show: - Topic birth and death - Word probability changes over time - Topic intensity trends