TL;DR
This paper introduces the Recurrent Sticky Hierarchical Dirichlet Process Hidden Markov Model (RS-HDP-HMM), a novel Bayesian nonparametric model that allows for position-dependent state persistence in time series data. By extending existing HDP-HMM frameworks, RS-HDP-HMM offers more expressive and accurate modeling of complex spatio-temporal phenomena, outperforming previous methods on segmentation tasks.
What Problem Does This Paper Address?
Traditional Hidden Markov Models (HMMs) and their Bayesian nonparametric extensions, like the Hierarchical Dirichlet Process Hidden Markov Model (HDP-HMM), often assume that the probability of staying in a particular state (self-persistence) is stationary. This means the likelihood of remaining in a state is constant over time, regardless of other factors. However, this assumption is unrealistic for many real-world spatio-temporal systems where the behavior of an entity is influenced by its current location or context. For instance, a car is more likely to turn when it's at an intersection, not just because it was previously in a "driving" state. Existing "sticky" HMM variants address self-persistence but still treat it as stationary. This paper tackles the limitation of stationary self-persistence by proposing a model where this probability can dynamically depend on the observed state.
Methods
The core contribution is the Recurrent Sticky HDP-HMM (RS-HDP-HMM). This model builds upon the sticky HDP-HMM and disentangled sticky HDP-HMM by introducing a recurrent component that makes the self-persistence probability dependent on the previous observation. This is achieved by incorporating a logistic regression model to predict the self-persistence parameter based on past observations.
To enable efficient inference in this more complex model, the authors develop a novel Gibbs sampling strategy. This strategy involves:
- Direct Assignment Sampler: This sampler marginalizes out transition distributions and parameters, then sequentially samples latent states and related variables. It's noted for its thoroughness but potentially slower mixing.
- Weak-Limit Sampler: This sampler leverages a finite approximation of the Dirichlet process to improve the mixing rate of the Gibbs sampler, making it more efficient for practical applications.
The paper also utilizes Pólya-Gamma augmentation, a technique that simplifies the inference of binomial probabilities by expressing them as a combination of Gaussian and Pólya-Gamma distributions, facilitating efficient Gibbs sampling.
Datasets
The models were evaluated on the following datasets:
- NASCAR® benchmark: This dataset simulates the trajectory of a car on a track, designed to test motion segmentation. It features four distinct latent states influencing a 2D continuous state.
- Dancing Bee dataset: This dataset captures the trajectories of honey bees performing waggle dances, known for its intricate characteristics and previously used for time series segmentation. It involves three types of motion: waggle, turn right, and turn left.
- Mouse Behavior dataset: This dataset involves monitoring mouse neural activity and behavior during a visual decision-making task. It's used to test the model's ability to handle complex, potentially over-segmented, behavioral data.
Key Insights and Ideas
- Position-Dependent Self-Persistence: The most significant insight is the successful integration of observation-dependent self-persistence into a Bayesian nonparametric HMM framework. This allows models to capture more nuanced temporal dynamics where state transitions are context-aware.
- Improved Segmentation for Spatio-Temporal Data: By relaxing the stationarity assumption, RS-HDP-HMM can better segment data where the underlying states are influenced by spatial location or other observed features.
- Novel Gibbs Sampling Strategies: The development of efficient sampling schemes (direct assignment and weak-limit) is crucial for making this complex model computationally tractable for real-world applications.
- Addressing Over-segmentation: The model shows promise in handling complex behavioral data, potentially mitigating issues like over-segmentation that plague other ARHMM models.
Key Results
- Superior Performance: RS-HDP-HMM consistently outperformed HDP-HMM, sticky HDP-HMM, and disentangled sticky HDP-HMM across all tested benchmarks (NASCAR®, Dancing Bee, Mouse Behavior) in terms of log-likelihood scores and segmentation quality (Weighted F1 and Accuracy).
- NASCAR®: Achieved the highest log-likelihood scores and produced the most coherent results, with segmentation that better differentiated states, especially the "straight-on" states.
- Dancing Bee: Showed comparable accuracy to S-HDP-HMM but produced smoother segmentation, demonstrating better handling of the distinct movement phases.
- Mouse Behavior: RS-HDP-HMM segmentation was significantly smoother and more stable than DS-HDP-HMM, despite using a similar number of states. It also showed greater tolerance for minor noise by considering actual position, not just direction of movement.
- Efficiency: The RS-HDP-HMM exhibited a considerably shorter burn-in time in fitting compared to other models.
Limitations and Future Work
- Computational Complexity: While efficient sampling schemes are developed, the model's complexity can still be a factor for very long time series or a very large number of states.
- Future Work: The authors suggest exploring further extensions and applications, particularly in areas where non-Markovian temporal dependencies are prevalent. They also mention the potential for further refinement of the sampling strategies.
Summary updated