Should I choose cluster based / density based approach?
Yes, but depends on how good the algorithm is, and how effectively is the algorithm identifying important clusters. There is a lot of literature where people have taken the clustering approach.
If I were to go with the clustering approach, I would go with the hierarchical clustering technique, cause it would help me re-check my clusters and build hierarchies in them so that the seriousness of the fraud can also be explored.
How do I find rare & suspicious events that parents should be notified
for with the help of machine learning ?
I would point out an answer which I have written for a very similar question, which also explores the aspect of fraud detection and anomaly analysis. And as I have said in the answer, the genetic algorithm still goes strong when it comes to fraud detection, owing to it's neat method of fitness calculation of the data points.
Any suggestions on how to present the results visually ?
Generally such charts are presented through pie charts and network graphs. you can also use infographics if you have a nice idea about design and aesthetics.
Is a week of data enough for training ?
Generally, data science teams at Paypal, Palantir, etc churn years of data for sensible insights. As you say it's a class project, data of several months should do good. You wouldn't really get anything suspicious or sensible from a week of data.
Are there any good open sources in Java to detect anomaly ?
Yes, there are some:
- Genetic algorithms library
- A Distributed Online Machine Learning Framework in java