I have noticed that was wasting too much time on reading tons of messages in telegram chats worrying to miss something useful. E.g. I have a chat with other people living in my house about some problems, solutions and joint ventures, producing over 300 messages a day.
So I decided to apply a little machine learning to this problem in order to save time and still be aware of problems being discussed in useful chats.
- Parse telegram
- Text analytics
Fails:
At first I used options provided by Telegram desktop to download data, but it takes a day to be approved and is downloaded in not very useful formatI tried LDA for text modeling, but it didn't fit well enough with the taskTSNE to look at 300-d vectors in 2-d was a bad idea - just a messKNN approach was to slow, not approved. HDBScan showed to be not relevant for the task. KMeans - too simplified