Large Language Models for Travel Behavior Prediction

This study demonstrates that large language models, utilized through zero-shot prompting or as embedding generators for supervised learning, offer a flexible and data-efficient alternative to traditional numerical models for predicting travel behavior.

The Gist

Imagine you are trying to predict what a person will do next: Will they drive, take the train, or hop on a bus? Will they be going to work, to a party, or to the grocery store?

For decades, transportation experts have tried to answer this using mathematical calculators. These calculators are like rigid recipe books. If you feed them enough data (thousands of past trips), they can cook up a pretty good prediction. But if you only give them a few ingredients (a small amount of data), the recipe falls apart, and the results taste terrible.

This paper introduces a new chef to the kitchen: Large Language Models (LLMs). Think of these as super-smart, well-read librarians who have read almost every book, article, and blog post on the internet. They don't just crunch numbers; they understand stories, logic, and human common sense.

The researchers tested two ways to use these "librarian chefs" to predict travel behavior:

1. The "Zero-Shot" Oracle (The Instant Expert)

The Analogy: Imagine you walk up to a genius who has never met you before, but you describe your situation: "I'm in a rush, I hate traffic, and I have a monthly train pass." You ask, "How will I get to work?"

The genius doesn't need to study your past data. They just use their vast knowledge of how the world works to say, "You'll probably take the train because it's fast and you already have a pass."

• How it works in the paper: The researchers wrote a detailed prompt (a set of instructions) describing a traveler's situation and the rules of the road (like "people hate traffic"). They asked the AI to guess the travel mode.
• The Result: Even without looking at any specific data about the traveler, the AI guessed almost as well as the traditional "recipe book" models, and sometimes even better when data was scarce.

2. The "Smart Translator" (The Feature Booster)

The Analogy: Imagine you have a very simple calculator that is bad at understanding complex human feelings. But, you have a translator who can turn a messy, emotional story into a clean, organized summary.

You give the messy story to the translator (the LLM). The translator writes a short, high-level summary (an "embedding") that captures the essence of the story. You then feed this clean summary to your simple calculator. Suddenly, your calculator becomes a genius because it's working with the "soul" of the data, not just the raw numbers.

• How it works in the paper: The researchers used the AI to convert travel descriptions into mathematical "vectors" (digital summaries). They then fed these summaries into standard models.
• The Result: When data was very limited, this "Translator + Calculator" combo outperformed the standard models that tried to learn from scratch.

The Good, The Bad, and The "Hallucinations"

The Superpower: Explainability
Old models are like black boxes. They give you an answer (e.g., "Take the train") but can't tell you why.
The LLM is like a detective. It says, "I think you'll take the train because you have a pass, and it saves you 30 minutes compared to driving." This is huge for planners who need to understand why people make choices.

The Weakness: The "Confident Liar"
Sometimes, the AI gets too creative. This is called a hallucination.

• Example: If you ask the AI to compare travel times, it might confidently say, "The train is 50% faster!" when the data actually says it's only 10% faster. It's like a student who knows the concept of math but makes a silly calculation error because they are trying to sound smart.

The Takeaway

This paper is a wake-up call for transportation planners.

• If you have a mountain of data: Stick to your traditional math models; they are still the kings.
• If you have very little data (a small town, a new city, a rare event): Don't panic. Call the "Librarian" (the LLM). It can use its general knowledge of human behavior to make surprisingly accurate guesses without needing to be taught the specific rules of your town first.

In short, we are moving from a world where computers only count to a world where computers can reason about how we move. It's not perfect yet, but it's a powerful new tool in the toolbox.

Technical Summary

Here is a detailed technical summary of the paper "Large Language Models for Travel Behavior Prediction."

1. Problem Statement

Travel behavior prediction (e.g., mode choice, trip purpose) is a fundamental challenge in transportation planning, traditionally addressed using numerical models like Multinomial Logit (MNL), Random Forests (RF), and Neural Networks (NNs). These conventional models rely heavily on large, labeled datasets for calibration. However, they face significant limitations in data-scarce regimes (small sample sizes) and lack interpretability, often functioning as "black boxes."

With the emergence of Large Language Models (LLMs), which possess advanced reasoning and world knowledge capabilities, there is an open question: Can LLMs effectively predict human travel behavior without task-specific training data, and can they serve as a flexible, data-efficient alternative to classical models?

2. Methodology

The authors propose and evaluate two distinct LLM-based frameworks for travel behavior prediction:

A. Zero-Shot Prompting Framework

This approach leverages the pre-trained knowledge of LLMs to make direct predictions without any fine-tuning or supervised training.

• Prompt Design: The authors engineered a comprehensive prompt structure containing four key components:
  1. Task Description: Defining the prediction goal (e.g., mode choice) and available options.
  2. Structural Data: Organizing travel attributes (time, cost) into a dictionary format for clarity.
  3. Descriptive Data: Converting individual attributes (e.g., "regular train user") into natural language sentences rather than binary variables.
  4. Domain Knowledge & Reasoning Guides: Explicitly instructing the LLM on how to weigh trade-offs (e.g., "People prefer lower cost/time") and providing comparative calculations (e.g., "Swissmetro saves 39% time") to mitigate arithmetic errors and hallucinations.
• Output: The LLM generates both the prediction and a natural language explanation (Chain-of-Thought reasoning).

B. Embedding-as-Features Framework

This approach uses LLMs as a feature extractor for supervised learning, specifically designed for small-sample settings.

• Process: Instead of generating a direct answer, the LLM converts the input travel scenario (prompt) into a high-dimensional text embedding vector.
• Integration: These embeddings serve as input features for traditional supervised learning models (MNL, RF, NN).
• Goal: To leverage the LLM's ability to capture high-level semantic relationships and behavioral nuances that are difficult to encode via manual feature engineering, thereby boosting performance when labeled data is limited.

3. Experimental Setup

• Datasets:
  1. Swissmetro Stated Preference Survey: 1,004 individuals, 9,036 responses. Task: Predict mode choice (Train, Car, Swissmetro).
  2. US National Household Travel Survey (NHTS) 2017: Task: Predict trip purpose (Working, Social, Recreation, Others).
• Baselines: Multinomial Logit (MNL), Random Forest (RF), and Neural Networks (NN).
• LLMs Tested: GPT-3.5, GPT-4, Llama 3.1 (8B and 70B).
• Scenarios:
  • Large Data: 1,000 training samples.
  • Small Data: 10 training samples (to test data efficiency).
  • Zero-Shot: No training data for the LLM.

4. Key Results

Performance Comparison

• Large Data Regime: Traditional supervised models (MNL, RF, NN) trained on large datasets achieved the highest accuracy and F1-scores.
• Small Data Regime:
  • Traditional models suffered significant performance degradation with only 10 training samples.
  • Zero-Shot LLMs outperformed all supervised models trained on small datasets. For example, GPT-3.5 zero-shot achieved ~58.6% accuracy in mode choice, surpassing MNL (46.9%) and RF (45.5%) trained on small data.
  • Embedding Approach: Using LLM-generated embeddings as features for supervised models consistently improved performance over manually engineered features in small-sample settings.

Interpretability

• LLMs provided natural language explanations for their predictions.
• Case Studies:
  • Success: LLMs correctly identified trade-offs (e.g., choosing a faster mode despite higher cost) and explained them logically.
  • Failure Modes: The study observed hallucinations (inventing facts not in the input) and reasoning errors (misinterpreting the impact of an annual pass). However, even when predictions were wrong, the reasoning was often plausible, offering diagnostic value.

Ablation Studies

• Prompt Sensitivity:
  • GPT-3.5: Highly sensitive to prompt design. Removing "reasoning guides" (comparative calculations) caused a 5.8% drop in accuracy. Structured data formats were also critical.
  • GPT-4: Demonstrated high robustness, with performance drops of less than 1% across all ablation scenarios, indicating superior internal reasoning capabilities.

5. Key Contributions

1. Novel Frameworks: Introduced two complementary LLM frameworks (Zero-Shot Prompting and Embedding-as-Features) specifically tailored for travel behavior prediction.
2. Data Efficiency: Demonstrated that LLMs can achieve competitive performance in low-data regimes where traditional models fail, offering a solution for scenarios with scarce historical data.
3. Interpretability: Showed that LLMs can generate human-readable justifications for decisions, bridging the gap between predictive accuracy and explainability in transportation modeling.
4. Prompt Engineering Insights: Provided empirical evidence that explicit reasoning guides and structured data inputs are crucial for smaller models (like GPT-3.5) but less critical for advanced models (like GPT-4).

6. Significance and Future Work

This study establishes LLMs as a viable, flexible alternative to classical econometric and machine learning models in transportation. It suggests a paradigm shift where natural language reasoning can complement or replace numerical calibration, particularly in data-constrained environments.

Limitations & Future Directions:

• Hallucinations: The risk of LLMs generating false information requires mitigation for safety-critical applications.
• Generalization: Future work needs to test on more diverse datasets and additional tasks (e.g., route choice, departure time).
• Optimization: Systematic research is needed to standardize prompt engineering and explore few-shot learning strategies to further reduce data requirements.

In conclusion, the paper argues that LLMs offer a data-efficient and interpretable path forward for travel behavior prediction, capable of outperforming traditional methods when data is limited while providing transparent reasoning for decision-making.