liveBook · Manning

Bandwidth parameter

As explained in Build AI Drug Discovery Pipelines

Bandwidth is the scale parameter controlling the spatial range and smoothing of a kernel density estimate.

The bandwidth is the scale parameter that controls how far each observation influences a kernel density estimate. It determines the estimate’s smoothing: nearby points receive influence through the kernel, while distant points contribute less. section 7.3.3

Adjust bandwidth to balance local detail against noise:

  • Use a smaller bandwidth for a localized estimate; it emphasizes points very close to the position being estimated but can overfit noise. section 7.3.3
  • Use a larger bandwidth for a smoother estimate; it incorporates more distant points but can oversmooth and hide distribution structure. section 7.3.3

Selecting bandwidth for pharmacophore distributions

A scikit-learn KDE can search bandwidths from 10^-3 to 10^3 with GridSearchCV, selecting a value by cross-validation before estimating pharmacophore pair distributions.

How Build AI Drug Discovery Pipelines covers it

How deep it goes
Defines bandwidth mathematically, explains its smoothing tradeoffs, and demonstrates cross-validated selection without deriving the estimator.
Examples
Uses Gaussian KDE, a bimodal distribution, and pharmacophore pair distributions to illustrate bandwidth selection and consequences.
What you can do after
Apply it via scikit-learn's KernelDensity and GridSearchCV to select smoothing for one-dimensional distribution data.
How it connects
Connects bandwidth with kernel density estimation, kernel functions, smoothing, overfitting, underfitting, and cross-validation.

Also explained in

Compare how 2 books explain bandwidth parameter →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build AI Drug Discovery Pipelines. The section links point to the parts of the book it draws on.