CDR analysis

[dimespi]

Hello,

I'm doing some analysis on Call Detail Records (CDR). My dataset is similiar to this: https://neo4j.com/blog/neo4j-call-detail-records-analytics/

Here are the fields from my dataset :

• source (operator)

• called_number

• calling_number

• calling_date

• country_code_from

• country_code_to

• usage

• service_name (SMS, DATA, VOICE)

  • SMS-OUTGOING
  • SMS-OUTGOING-ROAMING
  • SMS-INCOMING
  • DATA-OUTGOING
  • DATA-OUTGOING-ROAMING
  • VOICE-OUTGOING
  • VOICE-OUTGOING-ROAMING
  • VOICE-INCOMING
  • VOICE-INCOMING-ROAMING

If the service_name is SMS, the usage value will be set to 1.
If the service_name is DATA, the called_number and country_code_to will be empty.

I'd like to apply some machine learning algorithms and predictions for fraud/anomaly detection. I'm wondering wich one would be best for my use case? Kmeans, RandomForest, NaiveBayes, TimeSeries?

I found this:

• https://github.com/git4impatient/telco/blob/master/kmeansCDRwithNormalizationOfFeatures2019.pdf
• https://rpubs.com/Manura/milano
• https://github.com/awslabs/machine-learning-for-telecommunications/tree/master/source/industry/telec...

I'm using py2neo and MLlib.

[jennifer_reif]

What kinds of fraud or anomalies are you looking for in this data set? I think understanding a bit more about your use case would help me narrow down the better options.

Cheers,
Jennifer

[rsagar4]

Hi dimespi, were u able to find any example code on CDR analytics.. Great if you share link for example... also do have sample dataset for this.. Thanks in advance..