Hi Stephan,
Let's go at it step by step, so we can determine what the problem is.
I am using EA 3.10 Build 504.
I create a deployment diagram called, say, "PC Deployment". In this diagram I have created a node called "Community_proj_PC": this is not an instance (object), but a representation of the classifier itself. I have stereotyped it as "pc" (it could be something else, though).
In the same project, I have created four components: Linux, PostGreSQL, StarOffice and community_system, and only the last component is stereotyped (as executable).
I have dragged the components inside the node (again: as classifiers, not as instances), and I have "Use Stereotype Icons" checked in my Diagram Properties.
In this diagram, both nodes and components are shown with their respective icons.
As Mikkel has said, I drag the components out of the node, and an aggregation relationship appears. I drag the components inside the node again, because that is the place I want them to be.
I open my relationship matrix, choose the root of the project tree as both source and target; choose Node and Component as "Type", Aggregation as "Link Type", Bi-Directional (or Target -> Source) as "Direction". The matrix correctly shows that "Community_proj_PC" has the components. I save the matrix as a new profile (Profiles -> Save as New Profile).
Are you using instances of nodes or instances of components? In that case, you are right, and EA is showing instances with object (and not with node or component) icons. The "Community_proj_PC" has the PC icon in the upper right corner, but otherwise it is an object icon (even though I have selected "Instance has classifier style" in the Diagram Behaviour dialog).
In the case of instances there is problem, and I want to call Sparx attention to it (although, in all modesty, there is probably some good reason for this).
However, as to your original question, I do get a correct matrix if I compare object vs object, choose aggregation and the appropriate source to target (or biderectional) as Direction. I wouldn't like to use the object vs object matrix, because it does get confusing: in this matrix I have all sorts of objects, and not only instances of nodes and components.
I hope this helps.
Jaime