Learning Continuous Time Bayesian Networks in Non-stationary Domains

Simone Villa, Fabio StellaIJCAI 2018 · DOI

TL;DR

This paper introduces "non-stationary continuous time Bayesian networks" (nsCTBNs), a novel probabilistic graphical model that allows the relationships between variables to change over time. They developed a method to learn these networks from data, demonstrating their effectiveness on macroeconomic time-series data, which often exhibits evolving dependencies.

What Problem Does This Paper Address?

Many real-world systems, from biological networks to financial markets, are not static. Their underlying relationships and dependencies evolve over time. Traditional Bayesian networks (BNs) and even their dynamic extensions (DBNs and CTBNs) often assume these relationships are constant (stationary). This assumption breaks down when dealing with phenomena like business cycles, gene regulatory network changes, or evolving financial market structures. The paper addresses the challenge of modeling and learning these dynamic, non-stationary systems where the very structure of dependencies can change.

Methods

The core contribution is the introduction of the non-stationary continuous time Bayesian network (nsCTBN). This model extends continuous time Bayesian networks (CTBNs) by allowing the parent set of each node (i.e., the variables that directly influence it) to change over specific time intervals, called epochs.

Key methodological aspects include:

  • Structurally Non-Stationary Model: Unlike standard CTBNs, nsCTBNs are defined by a sequence of graphs, each representing the dependency structure for a specific epoch. Transition times between these epochs mark structural changes.
  • Extended Bayesian Score: To learn these nsCTBNs, the authors extend the Bayesian score commonly used for learning CTBNs. This allows them to infer both the structure and the transition times.
  • Learning Settings: They propose three settings for structural learning:
  • Known Transition Times (KTT): The exact times of structural change are provided.
  • Known Number of Epochs (KNE): The number of distinct structural periods is known, but their exact timing is not.
  • Unknown Number of Epochs (UNE): Both the number and timing of structural changes are unknown. This is the most challenging setting.
  • Discretization for Real-World Data: For continuous real-world time-series data, they employ a binary discretization strategy using a look-back window to convert continuous variables into discrete states.

Datasets

The paper evaluates their nsCTBN model on:

  • Synthetic Datasets: Used to compare nsCTBNs against nsDBNs under different knowledge settings (KTT, KNE, UNE) in terms of accuracy, precision, recall, and F1-measure.
  • Real-World Datasets:
    • Drosophila, Saccharomyces cerevisiae, and songbird datasets were used to compare nsCTBNs against state-of-the-art algorithms like TSNI, nsDBN, and TVDBN under the UNE setting.
  • Macroeconomic Dataset: This is the primary real-world dataset discussed in detail. It comprises 17 financial/economic time-series from the United States, spanning from January 1986 to March 2015, with varying granularities (daily, monthly, quarterly). This dataset was used to discover business cycles and evolving relationships among economic variables.

Key Insights and Ideas

  • Modeling Evolving Relationships is Crucial: The paper strongly emphasizes that many real-world phenomena are inherently non-stationary, and models that assume static relationships will fail to capture important dynamics.
  • Business Cycles as a Prime Example: The application to macroeconomic data highlights how nsCTBNs can effectively identify periods of structural change that correspond to economic business cycles, revealing how financial and economic variables interact differently during expansions and recessions.
  • The Power of Unknown Epochs: The UNE setting is particularly significant, as it allows the model to discover the number and timing of structural shifts directly from the data, making it a more general and powerful tool.
  • Discretization Strategy Matters: The choice of look-back period for discretizing continuous time-series data is presented as a critical parameter for successful learning.

Key Results

  • Competitive Performance: nsCTBNs are shown to be highly competitive with state-of-the-art algorithms on both synthetic and real-world datasets.
  • Macroeconomic Insights: On the macroeconomic dataset, the learned nsCTBN identified three epochs with transition times aligning with significant US economic turning points (around July 2000 and November 2007). The model revealed known relationships (e.g., personal income and unemployment) and potentially less obvious ones (e.g., non-farm payrolls and S&P500 index).
  • Figure 2 Analysis: The distribution of transition times learned by the nsCTBN on the macroeconomic data showed peaks around the identified economic turning points, and the posterior probability distribution favored a model with three epochs.
  • Figure 3 Visualization: The learned structures for each epoch in the macroeconomic dataset visually demonstrate how the dependencies between economic variables changed over time.

Limitations and Future Work

  • Discrete Variables Assumption: The current nsCTBN model assumes that variables are discrete. This requires a discretization step for continuous data, which can lead to information loss.
  • Parameter Setting Difficulty: Learning the optimal parameter settings for structural learning, especially in the UNE setting, is noted as being extremely difficult.
  • Future Directions:
    • Asynchronous Parent Set Changes: Developing a modeling approach where each node can change its parents set asynchronously, rather than synchronously at epoch transitions.
  • Streaming Domain Classification: Applying nsCTBNs to object classification tasks in streaming data environments.

Summary updated

BibTeX
@inproceedings{villa2018learning,
  author = {Simone Villa and Fabio Stella},
  title = {{Learning Continuous Time Bayesian Networks in Non-stationary Domains}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/804},
}
Figure