doubts on how to embedd transaction data!

Follow the full discussion on Reddit.
Hi, I'm working on anomaly detection, and I have transactions data with various fields, it's unstructured, it has categorical and numerical features, my goal is to see which transactions are similar and which is anomaly and I couldn't think of any statistical machine learning methods, as the data is very complex, so thought of using a autoencoders, maybe based on the reconstruction we can say which is similar or not, problem is i don't know how to embedd this data, where the representation captures all the complexities of the data, so how should I approach this problem? Ps: kinda of a noob, so please feel free to correct if I was wrong somewhere and if I'm not clear in some points! Thanks