TensorFlow.js Machine Learning

Front Page Forums Webmasters TensorFlow.js Machine Learning

This topic contains 9 replies, has 2 voices, and was last updated by  Marc 1 month, 2 weeks ago.

Viewing 10 posts - 1 through 10 (of 10 total)
  • Author
    Posts
  • #7720 Reply

    James

    Does anyone have any experience with TensorFlow.js?

    We are experimenting with it to do some training and processing of client data. I’m hoping we can leverage some form of neural network to accurately detect bots from humans and eventually use this to entirely automate the fraud controls.

    I think this will take the form of 3 stages.

    1) Collect and submit real-time client-side data via socket.io

    2) Format this data and optimise so that it’s processable effectively in a machine learning model

    3) Run it through TensorFlow to get a % chance of being a bot

    For years I’ve being saying that having blockchain and AI in the same sentence is a joke because blockchains are restricted to small blocks of data and AI requires huge amounts of data. Looks like I might have to eat humble pie and be another one of the projects using it as a buzzword if we can get this working effectively. To be fair I think it is the right tool for the job and could take the human factor out of the fraud systems.

    #7743 Reply

    Marc

    Extremely dangerous to let a computer decide who needs to be kicked out.

    #7758 Reply

    James

    We are running different models and generating a report and then checking it manually as well to start with. It really comes down to do we want humans to make that decision or a machine. There’s the potential for bias and criticism with humans whereas a machine learning algorithm would make a balanced analysis across the network IMO. This isn’t going to affect anyone that isn’t either sending lots of fake referrals or has a publisher mining site that they are sending high volumes of bot/rotator traffic to.

    #7765 Reply

    Josh

    The first 2 days after your updates my network of sites earned 300-500+ more coins then normal. Then on the 22nd my network of sites earned 1/2 what they normally do. Can you elaborate on exactly what factors are being used and the weight on each of them?

    I’m hoping the holidays, a cheater or a bad tweak to your algorithm is causing the extreme drop in day to day earnings after initial increases from your new distribution methods.

    My questions:
    – Are accounts limited in any way? Basically is an account with many sites have the same opportunity per site as an account with only 1 site?
    – What is the algorithm used now to decide who earned a ticket for the lottery system and are the limitations based on sites or accounts?

    #7774 Reply

    Marc

    @ Josh: the same here, earnings at first went up, and dropped to almost nothing starting yesterday (22/11) evening.

    But this was not a real surprise, James wrote clearly in his blogpost that the first days there will be some more changes in the settings from their part.

    Let’s cross fingers and hope the first settings will be the final ones 🙂

    #7824 Reply

    James Bachini
    Keymaster

    My questions:
    – Are accounts limited in any way? Basically is an account with many sites have the same opportunity per site as an account with only 1 site?
    No your sites with multiple domains would be treated exactly the same as if they were coming from a single domain or an account with a single site.

    – What is the algorithm used now to decide who earned a ticket for the lottery system and are the limitations based on sites or accounts?
    It’s based on unique opted in visitors and how long they stay on site. So a unique (24hr) user visits the site who is either already opted in or clicks the opt-in button. The mining starts and some client data is sent to the servers which is recorded for machine learning analysis. Lottery tickets are produced when client “quality” reaches a certain level and then distributed every x minutes. The lottery is run every 30 second block which distributes rewards to pending for 7 days before they are paid in to the clients account.

    We pushed out an update on the first day which meant that anyone with desktop traffic earnt a lot more and anyone with mobile earnt a lot less. This has been rebalanced and we’ve been tweaking the algorithm every day since. There’s going to be another update today and I think we will continue to tweak, analyse data, tweak for a while yet until it’s as optimal as possible.

    #7830 Reply

    Marc

    I sincerely start to think that my website does not meet the high JSE standards… earnings are still very low, so low that I start to wonder if it still makes sense to bother my visitors with ‘consent pop-ups’ and the occasional message from anti virus software.

    Liked it much more before, if it ain’t broke, don’t fix it.

    #7843 Reply

    James

    We have had some issues balancing the publisher earning algorithms. The last update was pushed out yesterday 27th November at around 8am UTC. So hopefully this will make a difference and earnings for the 27th will be higher and then today 28th will be the first full day of stats to compare with.

    #7845 Reply

    Marc

    Hi James,

    Thanks for the update, but I have a bad feeling about it. Website earnings yesterday were even less as the days before (around 35% from what it was on a reasonable day before the big update). Today is still early day, but I need a magnifying glass to find them…

    Maybe you can have a look at my account?

    #7860 Reply

    Marc

    James,

    If I may ask, do you have a cat?

    If so, please keep her away from your computer, it looks like she likes to run over your keyboard and change mining settings constantly.

    Seriously, this is not longer funny. The whole day I mine almost nothing (‘last mining reward received + 1 hour ago…’) and then suddenly BOEM: for one hour or so it’s like I hit the JSE jackpot. Unfortunately, this doesn’t last very long (mostly a bit over one hour) and then mining is over till the next time the cat tries to catch a mouse.

    IMO, if you like to experiment with AI: fine, but better do it with some test sites – never on a live site.

    I wonder if I’m really the only one with this problem? It’s very quite here….

Viewing 10 posts - 1 through 10 (of 10 total)
Reply To: TensorFlow.js Machine Learning
Your information: