Stochastic Block Models with R

## Building a software package in tandem with machine learning methods research can result in both more rigorous code and more rigorous research.

_Nick Strayer, PhD Candidate, Vanderbilt Biostatistics_

---

# Layout

- What I'm selling
- Getting to the starting line
- Setting up
- The journey
- The destination

---
class: middle

# What am I selling?

---
class: middle

## An idea

![:spacePx 80]
Building an R package alongside a research project serves to strengthen both the package and the research.

]

---
class: middle

## A thing

![:spacePx 80]
`sbmR`: an R package to perform uncertainty-aware unsupervised clustering of graph data. 
]

---

class:middle

# Getting to the starting line

Why a project gets started

---

# But first... jargon break!

---

# The __S__tochastic __B__lock __M__odel?

A network model to cluster nodes into "blocks".

![:centerPic 62](figs/sbm_explained.svg)

---

# An equation for validity

<div style="margin-top:-75px;">
![:centerPic 85](figs/sbm_formula.svg)
</div>

---

# Really why I showed an equation

![:centerPic 32](figs/poisson_tattoo.png)

---
class: middle

## Deciding to start a project

---
# Why SBMs

![:spacePx 50]
- Working with Electronic Health Records (EHR) data

- Questions like "who are these people similar to and why?"

- Needed method that acknowledges the limitations of the data. 
]
.pull-right[
![:spacePx 50]

![:centerPic 100](figs/why_block_models.png)
]

---

# Why write a package?

![:spacePx 70]

No tool existed that:

- Was for R

- Focused on uncertainty

- Allowed us to develop custom methods for bipartite networks etc.

]
.pull-right[
![:spacePx 80]

---
class: middle

# Setting up
## or, laying a solid foundation 🧱

---
class: middle

## Choosing dependencies
.pull-left[

![:spacePx 20]

iGraph is a monolith
- Adding computationally heavy steps in R brought it to a halt 
  
![:spacePx 20]
  
Decided to write algorithms from scratch using `Rcpp`
 
  
]
.pull-right[
![:spacePx 50]
![:centerPic 100](figs/igraph_logo.png)
]

---
class: middle

![:spacePx 90]

## Scafolding with `usethis`

_Use good defaults to setup a solid foundation_

]

---

## `usethis::use_testthat()`

.pull-left[
![:spacePx 180]
No excuses to not have tests...
]
.pull-right[
![:centerPic ](figs/test_that.png)
]

---
## `usethis::use_roxygen_md()`

![:spacePx 160]

Write your documentation with as little friction to publication as possible

]
.pull-right[

<div class = "shadowed" style="margin-top: -3rem;">
![:centerPic 82](figs/roxygen.png)
</div>
]

---

## `usethis::use_data_raw()`

![:spacePx 170]

Setup a consistent workflow for simulations and testing datasets
  
]

---
class: middle

---
class: middle

---

## Unit tests build understanding

![:spacePx 50]

- Toy examples to test basic functions

![:spacePx 30]

- Builds strong foundations for larger models

![:spacePx 30]

- Solidifies intuition about behavior

]
.pull-right[

<div class = "shadowed" style="margin-top: -5rem; margin-right: -1rem;">
![:centerPic 82](figs/unit_test_notes.png)
</div>
]

---

## Simulated data exposes software flaws

![:space 10]

- Stochastic data _will_ find your edge cases

![:space 5]

- Gives an idea of real-world performance

![:space 5]

- Forces you to think about scope early

]
.pull-right[
![:spacePx 90]

![:centerPic 100](figs/clustered_network.png)

]

---
class: middle

---

## Be a translator

![:spacePx 60]

]
.pull-right[
![:spacePx 50]

![:centerPic 90](figs/logic_gates.png)
]

---
class: middle

## Be a copy-editor

![:space 3]

![:space 5]

![centerPic 70](figs/entropy_inefficient.png)

---
class: middle

_Speaking of copy editing..._

![:spacePx 60]
## You still have to write
]
.pull-right[
.iconed[✍️]
]

---

# Doc driven development

![:spacePx 100]

![:spacePx 50]

---

# Vignette or perish

![:spacePx 80]

.bullet_emoji[🥀] If no-one knows __how__ to use your package/method no-one __will__ use it. 
![:spacePx 80]

]

<div style="margin-top: -7rem; text-align: right;">
<img src='figs/vignette_length.png' width = 61%/>
</div>
]

---

---

## A package

![:spacePx 10]

![:spacePx 10]

- R package to fit SBMs

- Investigates uncertainty of found structure by sampling from bayesian posterior

- Provides (a growing list of) visualizations to communicate results

![:spacePx 30]
]

---

## A pkgdown website

![:spacePx 60]

- `usethis::use_pkgdown()`
![:spacePx 20]

- All the docs and vignettes accessible online

![:spacePx 20]

- Don't have to download entire package to see if it does what you want
]

### [tbilab.github.io/sbmR/](https://tbilab.github.io/sbmR/)

---

## Manuscripts

- Glue together your work

![:spacePx 20]

- "Scripts available upon request" becomes `install_github(...)`

![:spacePx 20]

- Not the only avenue for dissemination 
]

---
class: middle

---

## Too much time is spent fiddling

![:spacePx 60]

- `usethis` and strong defaults help a lot

![:spacePx 60]

- Ask "would I end up doing this later anyways?"

![:spacePx 60]

- Set a list of requirements and make you way through it

---

## My project isn't strictly methods development

![:spacePx 100]
- The package infrastructure works well for general data-science projects

![:spacePx 60]
- See [Karthik Ram's work on compendiums](http://inundata.org/talks/rstd19/#/).

---

## I've never made a package before!

![:spacePx 80]

- It's easier than you may think:
  - http://r-pkgs.had.co.nz/ 
  
  
![:spacePx 60]

- If you want people to use your method, an R package is the most effective way to do so

---

## Thanks to
.pull-left[
- My advisor Yaomin Xu
- Lab Biostatistician Siwei Zheng
- Other lab members for thoughtful comments during entire process
- Vanderbilt Biostatistics Development Grant
- My Cat (and wife)

## Links
- Blog: [LiveFreeOrDichotomize.com](https://livefreeordichotomize.com/)
- Tweet: [@nicholasStrayer](https://twitter.com/NicholasStrayer)
- Code: [github.com/nstrayer](https://github.com/nstrayer)
]