求推荐支持多主体拟合等特性的R/Python隐马尔可夫模型（HMM）包

Great question! Let's break down the best tools in R and Python that check all your boxes for a multi-subject (panel/longitudinal data) HMM with discrete multivariate emissions, covariate-driven transition probabilities, and parameter constraints:

R Packages #

depmixS4 #

This is the go-to package for flexible HMMs in R, and it directly supports every feature you need:

• Multi-subject support: Use the group parameter to specify your panel data's subject identifiers, fitting either shared-parameter or mixed-effect HMMs across subjects.
• Discrete multivariate emissions: Supports multinomial and categorical families for multi-output discrete data; you can define separate emission models for each output variable.
• Covariate-dependent transitions: Define transition probabilities using formula syntax (e.g., transition = ~ age + treatment) — the package uses logit links to map linear covariate combinations to valid probabilities.
• Parameter constraints: Use the constraints argument to fix parameters, enforce equality across states/subjects, or define custom linear constraints. You can also manually adjust parameters with setpars() for more complex rules.

Quick code snippet to illustrate the setup:

library(depmixS4)

# Assume your data has: id (subject ID), y1/y2 (discrete outputs), cov1/cov2 (covariates)
hmm_model <- depmix(
  response = list(y1 ~ 1, y2 ~ 1),
  data = your_panel_data,
  nstates = 2,
  family = list(multinomial(), multinomial()), # Multivariate discrete emissions
  transition = ~ cov1 + cov2,                  # Covariates drive transitions
  group = id                                   # Multi-subject panel structure
)

# Add constraint: state 1→2 and state 2→1 share the same cov1 coefficient
constraint_mat <- matrix(c(NA, 1, 1, NA), nrow = 2)
hmm_model <- setpars(hmm_model, constraints = list(transition = constraint_mat))

# Fit the model
fit_model <- fit(hmm_model)
summary(fit_model)

mHMMbayes #

If you prefer a Bayesian framework, this package is built explicitly for multi-subject HMMs:

• Native support for longitudinal/panel data, with options to model subject-specific random effects alongside shared group-level parameters.
• Handles discrete multivariate emissions out of the box.
• Covariates can be included in transition probability models using logit-transformed linear predictors.
• Parameter constraints are enforced via custom prior distributions (e.g., fixing a parameter to a value by setting a narrow prior, or linking parameters across states/subjects through hierarchical priors).

Python Packages #

hmmlearn (Customized) #

While hmmlearn doesn't natively support covariate-driven transitions or multi-subject models, it's highly extensible to meet your needs:

• Multi-subject handling: Group your panel data by subject, and either fit separate models with shared parameters or build a joint model that pools information across subjects.
• Discrete multivariate emissions: Combine your multi-output discrete data into a single categorical variable (mapping each combination to a unique level) or define a custom emission probability function.
• Covariate-dependent transitions: Override the _compute_transition_matrix method to calculate transition probabilities using a logistic regression (or other link function) that incorporates your covariates.
• Parameter constraints: Use scipy's optimization constraints in the fit() method to fix parameters or enforce equality/inequality rules.

PyMC3/PyMC4 (Probabilistic Programming) #

For maximum flexibility, use a probabilistic programming framework to build a fully custom HMM:

• Multi-subject support: Implement hierarchical models to share group-level parameters while allowing subject-specific variation (e.g., random effects for transition coefficients).
• Discrete multivariate emissions: Define joint categorical distributions for your multi-output data, or model each output separately with correlated emission probabilities.
• Covariate-driven transitions: Directly model transition probabilities as functions of covariates (e.g., using tt.nnet.sigmoid to convert linear covariate combinations to valid probabilities).
• Parameter constraints: Explicitly define equality/inequality constraints between parameters, or fix parameters to specific values during model setup.

Here's a simplified skeleton of a PyMC3 model:

import pymc3 as pm
import theano.tensor as tt
import numpy as np

# Assume covariates_data is a (n_samples, n_covariates) array, observed_data is multi-output discrete
n_states = 2
n_outputs = 2

with pm.Model() as hierarchical_hmm:
    # Covariate coefficients for transition probabilities (state x state x covariate)
    beta = pm.Normal("beta", mu=0, sd=1, shape=(n_states, n_states, covariates_data.shape[1]))
    
    # Calculate transition probabilities using logit link
    trans_probs = tt.nnet.sigmoid(tt.dot(covariates_data, beta))
    
    # Emission probabilities for each state and output
    emission_probs = pm.Dirichlet("emission_probs", a=np.ones((n_states, n_outputs)), shape=(n_states, n_outputs))
    
    # Initial state distribution
    init_dist = pm.Dirichlet("init_dist", a=np.ones(n_states))
    
    # Hidden Markov Model
    states = pm.HiddenMarkovModel(
        "states",
        trans_probs=trans_probs,
        emission_probs=emission_probs,
        init_dist=init_dist,
        observed=observed_data
    )
    
    # Sample from posterior (Bayesian inference)
    trace = pm.sample(2000, tune=1000)

内容的提问来源于stack exchange，提问作者Joep van der Plas