Accountanalysis

Comment Deleted

@kanad_bahalkar It makes me happy that accountanalysis is useful to you. At the moment the tool is used by very different groups (researchers/journalists and marketers) which would mean that such suggestions need to be customized for each group. I can imagine it as an option in the account settings. "I am a marketeer" / " I am a researcher". And depending on what someone choose, we show additional info. "Most Tweets come from automation tools.", "Account seems has no awake/sleep pattern", "Most active in the morning". We have to figure out, if we should re-calculate those things depending on how users filter Tweets or not. Quite a big feature. Until I am able to determine it's impact, it has a low priority. - Tooltip for "selected/retrieved Tweets" is already in work. "Retrieved" tells you how many Tweets there are in total in the current analysis. "Selected" refers to the number of Tweets that are currently filtered. You can click on any feature (hour of weekday, time frame, Tweet type, replied to account) to filter for those Tweets. And combine filters from all charts. "Selected" will always tell you how many Tweets pass those filters. - Daily rhythm. Google maps only shows 6am to 2am of a single day at a time. accountanalysis shows all 24 hours of 7 days at a time, because we are more interested in general patterns than in current volume. The daily rhythm of @producthunt becomes easier to read, once you exclude buffer, because it posts once per hour 24/7. - The search box is important, but isn't used as much as the crossfiltering. It's not like an search engine where you run several searches after each other. It's more like opening a file and then working with that file for longer. - Most people don't need the explanations, but it makes it easier for new users to get started.

@__tosh Thanks for hunting accountanalysis! Identification of groups of bots starts with single accounts in most cases. Sometimes when looking at a trending topic there a weird Tweets or an account is suddenly followed by several similar looking accounts. Once you have the first suspicion, you investigate. Just visiting the profile of the account can be enough. If your suspicion grows, you can run the account through accountanalysis. The vast majority of bots isn't very sophisticated and the daily activity pattern already shows that the account is active 24/7. But even if that's not the case, you can look at the other features of the account. Bots often tweet in many languages and about things that have nothing to do with each other. In accountanalysis you would look at the language and URL graph. Once you are certain that it is a bot, you can try to find accounts that are operated by the same entity. Again, you look for patterns that are specific for that group. Maybe the interface (app aka source) they use or they all share a specific link or use a weird typo. With that information you go back to the Twitter search to find more accounts like them. Once again, running them through accountanalysis allows you to see similarities. Once you found a pattern that has few false positives, you would need a tool that allows you to gather many accounts at once and test them for patterns. I wrote a Python script for that (twecoll3) and use Gephi to create a network visualizations of the whole network of bots. Hopefully it will be possible to do that directly in accountanalysis one day. If anyone is interested in such stuff, I post bot hunting threads on Twitter from time to time. - Network of accounts, operated by humans, that engage in inauthentic behaviour that can be identified through special hashtags: https://twitter.com/luca/status/... - Bots that were active before the national election in Germany: https://twitter.com/luca/status/... - Network of Reply spam bots, identifiable by their typos: https://twitter.com/luca/status/...

@bavaturka Thanks! We decided to not use AI to make the analysis more transparent and comprehensible. There are other tools like botometer, botrnot and similar ones which focus on bot detection through machine learning. While they are nice to calculate scores on big numbers of accounts, they are inaccurate when looking at single accounts. Some researchers use such fully automated tools to get a list of suspicious accounts and then inspect those accounts with accountanalysis.