I wrote about the performance tuning I did in preparation to the Shown HN post: https://linklonk.com/item/277645707356438528
If you have any feedback please add a comment to that post.
395 karma · joined January 9, 2021
Email is "lonk" at the above domain.
I wrote about the performance tuning I did in preparation to the Shown HN post: https://linklonk.com/item/277645707356438528
If you have any feedback please add a comment to that post.
I'm planning to do a "Show HN" post next week for it and would appreciate any feedback that I could address before it. We have about 4 active users and a couple more would be great.
If you know that the content you like will get more visibility then it incentivizes people to upvote what they want other people to see, as opposed to what they themselves want to see.
These systems have far fewer "Likes" than other implicit signals such as which posts the user read (ie, didn't scroll past). So they use all of those signals and likes are unlikely to have a strong impact on the output of the algorithm.
The algorithms are opaque by nature. Even the developers of these deep models won't be able to explain to you why a user's recommendation list is ranked the way it is. Now imagine a user contemplating the consequences of liking a certain item on their future recommendations. It's hopeless.
As a result, the developers learn not to rely on these "like" signals and the users learn to provide likes only to affect the directly observable changes - the like counter.
But what if we designed the “Like” button to be directed not at the others, but at yourself. I mean, when you press the “Like” button it has consequences for you. If you “Like” useless content, then you will get more useless content in the future. Would that change the dynamics of “liking”? Would it make people think more carefully about what they like if their future content recommendations depended on it?
I’m building just such a system at https://linklonk.com
When you upvote an item - you don’t simply increment the counter and make that item rank higher for everyone else. Instead, you connect stronger to other users who upvoted that item before you. The stronger you connect to someone - the more weight their upvote has for you - the higher their other upvoted content ranks in your recommendations. This creates a feedback loop where you use the upvote button no to influence others, but to direct your future recommendations. This is a “filter bubble” - one that you very consciously form.
But as a user I can say that I do behave differently when it comes to upvoting content on LinkLonk. It is much less impulsive because my ratings have consequences for myself. If I find something that I completely agree with but I haven't learned anything new - I would not upvote it on LinkLonk, while I would be tempted to signal boost it elsewhere (like HN). Let's see whether this generalizes and what kind of system behaviour emerges out of it.
To properly test it out we need more users.
You have to be good at prediction what people will like in order to get them connected to you. And this is what a good curator does.
Though there is one important bit of asymmetry: you connect to those who upvoted that item *before* you. That way people who recognize useful information earlier earn more trust. In a sense this is a "proof of work" - to recognize valuable content before it becomes popular.
My hope is that this asymmetrical nature of connections will get less informed people to connect to more informed people. Which is the opposite of the echo-chamber effect - when less informed people are connected to similarly less informed peers.
And yes, I'm slowly preparing to do a "Show NH" for LinkLonk. But I probably need to grow the number of active users a bit before I do that, otherwise the "Shown HN" will bring a lot of clicks that will just bounce.
In a sense this is similar to how users of RSS feed readers control which feeds they subscribe to. They are responsible for the content they consume. What LinkLonk adds is a transparent layer of automation that helps you subscribe/unsubscribe based on your content ratings.
My hope is that LinkLonk will help people get more informed, but I cannot be sure. The project is a live experiment to find out if this is right system of incentives.
LinkLonk also increases your "downvote connection" to other who downvoted that item as you are suggesting. A "downvote connection" is how much weight the other person's downvote has for you. That is, it captures how good their past downvotes were for you and how much their future can be trusted. So there are two kinds of connections:
- Upvote connection - gives others ability to promote/curate good content for you.
- Downvote connection - how much others can bury/moderate bad content.
And as you are also suggesting, the connection is decreased over time without similar ratings. Each time someone you are connected to upvotes something your connection to them becomes slightly weaker. So if you ignore content from a user/RSS feed then it will have lower ranking for you over time. So in practice the downvote button should not be used much at all.
In the paper they had the participants of the experiment to manually pick who they want to follow. But what if the system connected them automatically to high signal-to-noise individuals based on feedback alone?
I've been working on something like this with my hobby project https://linklonk.com - an information network where the connections between you and other users are determined by your ratings of content.
When you upvote an article - you connect to other users who upvoted that article. When you downvote - your connection to those who upvoted it becomes weaker. That way the strength of your connection to other users captures the signal-to-noise ration of those users for you.
The stronger you are connected to someone - the higher their other upvoted items are ranked on the "For you" page.
For example, I upvoted this paper on LinkLonk: https://linklonk.com/item/6534389451373608960 If you also upvote it then you will get connected to me and will see more of my recommendations on the main page. The next user who upvotes it will connect to me and to you, etc.
Since you know that your content ratings have direct effect on what content your future self will see, you are incentivized to think whether each piece of content that you just consumed was truly worth your time. This kind of retrospective thinking is missing when we hit upvote/retweet/like in the existing social systems.
My project is in a very early stage and suggestions/ideas are welcome.
Is supposed to be substituted when it is send as an email?
The problem though is that it is hard to find users you want to subscribe to. You need to inspect their past submissions to see if the signal-to-noise ratio of that user is high enough for you. If you oversubscribe to too many users then your subscriptions will be full of noise.
In my hobby project https://linklonk.com I want to solve this problem of discovering users and feeds that have a high signal-to-noise ratio for you.
It works similar to HN - users submit links and upvote/downvote them. But the way your ratings are used is different.
When you upvote something - you get connected to other users that upvoted the same article/post. The more shared upvotes you have with someone => the stronger you are connected to them => the more weight their other upvotes have for you.
If you downvote something - your connections to those who upvoted that item become weaker, their other upvoted have less weight for you.
Your connections to other users capture how useful their previous recommendations have been for you. And over time you get connected to the best content curators for you. All this done automatically as you rate content. You do not need to manually manage who to subscribe to/unsubscribe from.
The above mechanism works not only for users but also for RSS feeds that posted the content you upvoted. This way LinkLonk works as an RSS reader that prioritizes based on your content ratings.
- Control: what you see is based on your explicit upvotes & downvotes, not what you happened to click on.
- Transparency: you see content from feeds that posted content you upvoted and from other users who upvoted what you upvoted, but did that before you. For example, when you upvote a link it will tell you that "You will get more content from 2 users that also liked it and from 3 feeds that posted it". And when you see a recommendation from other users you can see what likes in common do you have with those users (“We’re showing you this because ...”).
- Fairness: the amount of attention every user and feed gets from you depends on how useful their past recommendations have been. Ie, the higher their signal-to-noise ratio has been for you - the more prominently their other upvoted items will be in your "feed".
Give it a try and let me know what you think. I'd like to do a "Show HN" post for my project soon and your feedback would help me prepare for it.
If you like to read a bit more here are my announcements with discussions:
- https://tildes.net/~tech/u7f/linklonk_a_link_aggregator_with...
- https://www.reddit.com/r/RedditAlternatives/comments/mpqnpl/...