How to represent relation between users as a feature?

Question

I'm developing a model for unsupervised anomaly detection. I have a dataset representing communications between users (each example represents a communication): there are many features (time, duration, ...) and the ids of sender and receiver. My question is: how to represent the link between those two users? I have several ideas, but each of them seems to have serious drawbacks:

1. Use id as is. Drawback: even if ids are integers, they have no numerical sense (id 15 is not 3 times id 5) and I think this may mislead the system
2. Use sort of vectors: for example, with 3 users: user1 = (0 0 1), user2 = (0 1 0), user3 = (1 0 0). Drawback : the number of users may vary over time, thus the number of features would vary as well and I would have to re-train my model.
3. Graph theory: I've heard of that way of representing data, which could fit perfectly my data model. Drawback: I've absolutely no knowledge in graph analysis
4. Assign each user a id which is a prime number. That way a communication could be represented in an unique way as the product of the 2 ids. Drawback: as for point 1, ids do not have a "numerical sense"

What do you think may be the better way to represent these relations?

Answer 1

There's a couple of approaches you can take depending on the nature of your data. It sounds like you're trying to detect social anomalies in your data so you need to model the communication boundaries between them which leads to some sort of graph representation.

If you don't have too many users in the system (say $n$) then you can create an $n\times n$ matrix, $M$ over a time period that represents the communications between users. The component $M_{ij}$ could be either $1$ or $0$ if user $i$ and $j$ communicated or the number of times that they communicated.

If you have more data then you would want to represent the data in terms of nodes and edges. Nodes would be the users and edges would be the presence of a communication. This can be done manually or by using a library such as NetworkX.

Here's a Python tutorial on getting started with graph network analysis.

If you're doing this at a large scale then you might want to use a graph database such as Neo4J.