1
$\begingroup$

I have a dataset that contains 2 types of features, one is generated from doc2vec and one is numerical feature. I would like to perform clustering analysis on them. However, due to the size of doc2vec features, if I simply combine them into one array, clustering algorithm would distribute the "weight" on the doc2vec features more, how do I overcome this problem?

For example, for a given label, say I have features from doc2vec that look like this [1,2,3,4,5], and numerical feature [2]. I don't want to simply combine them into [1,2,3,4,5,2] and perform clustering analysis. Ideally, I would like my clustering algorithm to give the numerical feature equal importance as the doc2vec feature.

$\endgroup$
0
$\begingroup$

One way to achieve this is to use a clustering method based on a custom similarity/distance measure. For example you could defined the similarity measure between two instances as:

$$sim(\langle v_1, n_1\rangle,\langle v_2, n_2\rangle)=\frac{1}{2} cosine(v_1,v_2)\ +\ \frac{1}{2} \left(1-\frac{|n_1-n_2|}{max(n_1,n_2)}\right)$$

This measure gives the same weight to the similarity between the vectors ($v_1$ and $v_2$) and the similarity between the numerical values ($n_1$ and $n_2$). Note that since this similarity measure is normalized, you can convert it to a normalized distance measure: $d=1-s$. Of course you should define the exact measure based on what the values represent, this is just an example.

You could use this measure with a hierarchical clustering method or a graph clustering method (with edges based on similarity value).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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