I have been modeling data using a hybrid $k$-nearest neighbors (kNN) and multiple linear regression (MLR) and have found the technique to be (at least with my data) much more accurate than either method alone. In order to describe the technique, I will quickly describe kNN and MLR.

In both cases, we are given $n$ observations, say $(x_1, y_1, z_1),...,(x_n, y_n, z_n)$ where $x_i$ and $y_i$ are independent variables (not random) and $z_n$ is a dependent variable subject to some random noise. We want to predict the value of $z$ at some (unobserved) point $(x,y)$.

kNN regression:

Find the $k$ nearest observations to $(x,y)$ and average the corresponding $z$ values. Specifically, if the $k$ closes points to $(x,y)$ are $(x_1, y_1),...,(x_k, y_k)$ then we predict $z = \frac{1}{k}\sum_{i=1}^{k} z_{k}$.


We predict $z$ by building a best fit surface (using least squares) and use the equation of the surface to predict $z$. E.g. we may want to model the surface with a degree 2 polynomial (in two variables), i.e. $$z = \beta_1 + \beta_2 x + \beta_3 y + \beta_4 x^2 + \beta_5 xy + \beta_6 y^2.$$ In that case, we build matrices $X$ where the $i$th row of $X$ is $X_i = \begin{bmatrix} 1 & x_i & y_i & x_i^2 & x_i y_i & y_i^2\end{bmatrix}$ and $Z$ (a column matrix of the $z_i$ values), both matrices have $n$ rows. Then, we determine the coefficients $\beta_i$ by solving the normal equations $\beta = (X^{T}X)^{-1}X^{T}Z$.

Hybrid Method:

One can view kNN method as building a degree 0 polynomial (constant) using only the $k$ nearest data points. The technique that I've used is to build a MLR model from the $k$ nearest neighbors (without the constraint that the polynomial be degree 0). We also don't require that the same model be used for every region--a region being a set of points with the same $k$ nearest neighbors. We use $k$-fold cross validation to determine which polynomial degree is optimal--when $k$ is small, of course the degree of the polynomial is small. This method feels like a natural relaxation/generalization of traditional kNN.

Question: Does this hybrid approach have a name? Is it used frequently? If so, I would love any references that you can provide. I'm in the process of writing a paper that uses this modeling technique and I'd like to call it by the "correct" name (and cite original authors) or be able to honestly claim that it is novel.

Additional Notes:

The surface that is constructed with this method is akin to splines--it is a piecewise polynomial model. However, it is different in that it is not differentiable (and in most cases) probably not even continuous.

  • 2
    $\begingroup$ It sounds like a CART model, a regression tree where there is regression across the leaf-tip instead of a constant value. Oblique tree link, gives oblique splits for categorical data. I do not know how it handles regression. $\endgroup$ Feb 18, 2016 at 16:57
  • 2
    $\begingroup$ This sounds like a local regression method similar to Cleveland's loess and lowess regression en.m.wikipedia.org/wiki/Local_regression It is used frequently for scatterplot smoothing. $\endgroup$
    – Momo
    Feb 18, 2016 at 18:18
  • $\begingroup$ @Momo, this does seem very similar to LOWESS--except that the weighting function of nearest neighbors (1 if among $k$ nearest, 0 otherwise) isn't smooth, and a it can't be computed pairwise (if I know $x$ and $y$ I don't know if there are $k$ values closer to $x$ than $y$ or not). Thanks for pointing out the similarity. $\endgroup$
    – TravisJ
    Feb 18, 2016 at 19:41
  • $\begingroup$ @EngrStudent, this is sort of similar to CART in that there are regions which are modeled with the same polynomial. But, it seems hard to do branching like a tree since definition of the regions involves the data itself in a not obvious way. $\endgroup$
    – TravisJ
    Feb 18, 2016 at 19:42
  • 3
    $\begingroup$ @EngrStudent, all the data (and the scripts I'm using) will be made publicly available late next week (once I've finished writing and have submitted the paper). At this point, I'm not looking for new methods to perform the modeling (or suggestions of better ways to do it); I really want to know is whether or not I "re-invented" the wheel by recreating a known and widely used technique. And, if the technique isn't novel, I want to give credit to its originators. $\endgroup$
    – TravisJ
    Feb 18, 2016 at 20:18

2 Answers 2


There was a paper


Ann. Statist. Volume 22, Number 3 (1994), 1346-1370. "Multivariate Locally Weighted Least Squares Regression" by D. Ruppert and M. P. Wand

discussing a multiple regression where the weights on observations are not 0 or 1 as in your method but vary with the distance from the case you are trying to predict. Multivariate local regression is also discussed in section 2.1.2 of the book "Local Regression and Likelihood" (1999) by Clive Loader. The R package locfit implements univariate and multivariate local regression.


As @Fortranner said, what you describe is locally weighted regression. A particular case, where distances are in a geographical sense, is geographically weighted regression: a reference is Geographically Weighted Regression: The Analysis of Spatially Varying Relationships by A. Stewart Fotheringham, Chris Brunsdon, Martin Charlton, Wiley.

Notice that locally weighted regression affords great flexibility but on the other hand consumes many more degrees of freedom and is prone to overfitting. One is always well advised to test the model(s) on a validation sample.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.