Toplot

Reference docs

Visualizations for topic models.

Installation

pip3 install toplot

Getting started

Topic modelling is a Bayesian endevour. After training your topic model with $K$ components, you've inferred the distribution over two latent variables:

• The posterior over the weights (i.e., the topics) of the model $\pmb{W} = [\pmb{w}_1, \dots, \pmb{w}_K]^T$. We assume that the weights have a two-level structure: each weight is composed of categorical variables (or actually, multinomials), each consisting of a set of categories.
• Per training example $i$, the posterior over the hidden units $\pmb{h}^{(i)}$ (topic loadings, also denoted as $\pmb{\theta}_i$ in LDA).

Visualizing weights (the topic/cluster, $\pmb{w}$ or $\pmb{\phi}$)

Toplot expects your topic model's posterior samples to be organized in specific ways. As an example, we draw 1000 samples from "fake" topic weights $\pmb{W}$ containing two categories, body mass index (BMI) and sex, consisting of three and two categories each, respectively.

import pandas as pd
from numpy.random import dirichlet

# Draw 1000 samples from "posterior" distribution.
weight_bmi = dirichlet([16.0, 32.0, 32.0], size=1_000)
weight_sex = dirichlet([8.1, 4.1], size=1_000)
weight = pd.concat(
    {
        "BMI": pd.DataFrame(
            weight_bmi, columns=["Underweight", "Healthy Weight", "Overweight"]
        ),
        "sex": pd.DataFrame(weight_sex, columns=["Male", "Female"]),
    },
    axis="columns",
)

Use bar_plot to visualize the topic weight, including the 95% quantile range:

from toplot import bar_plot

bar_plot(weight)

If you have many multinomials, you can use bar_plot_stacked to reduce the width of the plot. This plot folds the categories (e.g., "Underweight", "Healthy Weight", and "Overweight") belonging to the same multinomial (BMI) into a single bar.

from toplot import bar_plot_stacked

bar_plot_stacked(weight)

To visualize more than one topic at a time, you can make a scattermap with scattermap.

Visualizing hidden units (topic proportions, $\pmb{h}$ or $\pmb{\theta}$)

Next, we plot the hidden units/topic identities $[\pmb{h}^{(1)}, \dots, \pmb{h}^{(m)}]^T$: that is, for each record $i$, the proportion over the components/topics. Let's generate the (average) proportion for $m=30$ records to visualize:

hidden = pd.DataFrame(
    dirichlet([0.6, 0.8, 0.2], size=30),  # 30 records
    columns=["Topic_1", "Topic_2", "Topic_3"],
)

The function plot_cohort computes the distance between all examples (the cohort) and, by default, sorts them accordingly using the travelling salesman problem. Currently, no uncertainty visualization is supported for plot_cohort (like in bar_plot), so you need to pass the posterior average.

from toplot import plot_cohort

plot_cohort(hidden)

You can emphasize the periodicity inherent in the travelling salesman solution by visualizing all the examples using a polar plot:

from toplot import plot_polar_cohort

plot_polar_cohort(hidden)