Utility discounting explains informational website traffic patterns before a hurricane

Ben Priest, Kevin GoldWWW 2013 · DOI

TL;DR

This paper uses psychological models of "utility discounting" to explain why people visit weather websites more frequently in the days leading up to a hurricane. By analyzing web proxy logs, they found that hyperbolic discounting models, which account for time-inconsistent preferences, best fit the observed traffic patterns, suggesting these models could improve predictions for rare, high-impact events.

What Problem Does This Paper Address?

Predicting surges in website traffic, especially for critical events like natural disasters, is a significant challenge. Traditional load prediction models often rely on machine learning, but understanding the underlying human behavior driving these surges could lead to more robust predictions. This paper tackles the question of whether psychological models of decision-making, specifically how people value rewards that are delayed or uncertain, can explain the increased demand for information before a major event like a hurricane. The authors hypothesize that as a hurricane approaches, the perceived utility of weather information increases, but this utility is "discounted" over time due to the uncertainty and the approaching deadline.

Methods

The core of the methodology involves fitting four different utility discounting models to observed website traffic data. The authors collected HTTP request logs from a mid-sized enterprise's web proxy for the days leading up to Hurricane Sandy. They filtered these logs to identify requests to popular weather websites.

The four models they tested are:

  1. Exponential Discounting: A standard model assuming time-consistent preferences, where the value of a reward decreases at a constant rate over time.
  2. Hyperbolic Discounting: A model that captures time-inconsistent preferences, where people tend to discount future rewards more heavily in the short term than in the long term. This is a common observation in behavioral economics.
  3. Green & Myerson Extension: An extension of hyperbolic discounting that introduces a parameter to model sensitivity to the scaling of delay.
  4. Rachlin Extension: Another extension to hyperbolic discounting, this one models sensitivity to the delay itself.

They then analyzed the data by calculating the mean probability that a user would seek weather information within a given hour, interpreting this probability as proportional to the perceived utility of that information. This probability was then used to fit the parameters of each discounting model.

Datasets

The primary dataset used in this paper is a collection of HTTP request lines from the web proxy of a mid-sized enterprise. This data was collected from Friday, October 26, 2012, at 6 pm, until Monday, October 29, 2012, at 9 am, specifically capturing the period leading up to Hurricane Sandy's landfall. The data was filtered to include requests from 1207 distinct users to 19 popular weather-related websites. This is a novel dataset for this type of analysis, as it directly reflects real-world, human-driven behavior in response to an impending crisis.

Key Insights and Ideas

  • Human behavior before crises is predictable: The paper demonstrates that psychological models of decision-making can explain observable patterns in human behavior, such as increased website traffic, in the lead-up to significant events.
  • Time-inconsistent preferences matter: Hyperbolic discounting models, which acknowledge that people's preferences can change over time (they are impatient in the short term but more patient in the long term), provide a better fit to the data than the simpler exponential model. This suggests that human decision-making regarding information seeking isn't always rational and consistent.
  • Utility discounting as a predictive tool: The findings suggest that utility discounting frameworks, which model how people value delayed rewards, can be a valuable tool for predicting traffic patterns for rare events. This could complement existing machine learning approaches.
  • Model complexity vs. fit: While more complex hyperbolic extensions (Green & Myerson, Rachlin) offered slightly better fits, the paper notes that the quantitative difference between them was minimal, highlighting the ongoing challenge in precisely distinguishing between these nuanced models.

Key Results

The paper found that the hyperbolic discounting models (specifically, the Green & Myerson and Rachlin extensions) provided the best fit to the observed weather website traffic data, as indicated by higher R² statistics compared to the exponential model.

  • Model Fit: The R² values for the models were: Exponential (0.6696), Hyperbolic (0.7242), Green & Myerson (0.8191), and Rachlin (0.7984). This shows a clear improvement with the hyperbolic and its extensions.
  • Qualitative Fit: Figure 1 visually confirms that the hyperbolic and its extensions more closely track the observed data, particularly the leveling-off of traffic as the hurricane's arrival became imminent.
  • Practical Implication: The study suggests that these models can explain the surge in informational website traffic before a disaster, offering a psychological basis for load prediction.

Limitations and Future Work

Limitations: * Model Uncertainty: The paper acknowledges that it's difficult to quantitatively distinguish between the two more complex hyperbolic extensions (Green & Myerson and Rachlin), suggesting that their better fit might be a matter of taste rather than a definitive superiority. * Representativeness: While the authors assume the filtered corpus of weather website requests is representative of the enterprise userbase's behavior, this is an assumption. The data is also limited to one specific event (Hurricane Sandy) and one enterprise network.

Future Work: * Broader Application: The authors suggest extending these findings to predict traffic for other types of deadlines or critical events, such as academic deadlines or other types of emergencies. * Refining Models: Further research could focus on better distinguishing between the various hyperbolic discounting models or developing new models that capture the nuances of human decision-making in crisis situations. * Integration with Load Prediction: Exploring how these psychological models can be integrated into existing hierarchical load prediction systems is a promising avenue.

Summary updated

BibTeX
@inproceedings{priest2013utility,
  author = {Ben Priest and Kevin Gold},
  title = {{Utility discounting explains informational website traffic patterns before a hurricane}},
  booktitle = {The Web Conference},
  year = {2013},
  doi = {10.1145/2487788.2487808},
}
Figure