When does it makes senses to use Dot-Product as similarity measure instead of Cosine?. I have seen there is already question asked about this, However that merely explains the difference between calculation of dot-product & cosine and it does not focus on when should we use one vs another with real world example.
Borrowing from this Quora answer -- which includes some concrete examples of when you might prefer one measure over the other -- it comes down to whether you care about taking the magnitude of the vectors into account. This is highly domain-specific, but the example used there is in information retrieval, where magnitude represents the length of the documents in question. The dot product will take the document length into account, whereas the cosine similarity will not.
In general, you will want to have good reasons specific to your problem space to use one over the other.
When we want cluster
items use distances as similarity measure. For example, we use Euclidean distances(square root of inner product) in k-means clustering as a similarity measure. Squared Euclidean distances are used as a similarity measure in Ward's method of clustering. However, when we want to cluster
variables, we use correlation(cosine) as a measure of similarity.
The relation between dot product and cosine is similar to the relation between covariance and correlation: one is normalized and bounded version of another.
In my experience usual dot product is better when you also care about the number of dimensions two vectors have in common (i.e. non zero values in these dimensions with the same sign). For example, it can be matchings tags or attributes. For usual texts cosine typically works better.