3 ms·

Interesting. I'd thought about some sort of meta-search service running over Google, Bing, etc., and aggregating results, that was tied in with some way of rati

Interesting. I'd thought about some sort of meta-search service running over Google, Bing, etc., and aggregating results, that was tied in with some way of rating recommending URLs. I've never got around to doing it and one of the reasons (excuses) is that I haven't yet figured out a way to stop bad actors exploiting it by, e.g., mass downvoting URLs of competitors. I did think about some sort of moving average, or bias toward more recent votes, which would at least somewhat mitigate against transient actions of this kind, but it's harder to deal with a sustained attack like this.

jacobobryant 6y ago
I've thought about that a fair amount also. I think the key is personalization. If bad actors downvoted a bunch of items, it would only affect people who already shared a lot of similar votes with those actors. That is still exploitable: if I know you like A, B, and C, (perhaps those are popular items), I could also upvote those items and then downvote another item D. I have an idea for countering this: keep track of history. So instead of just measuring ratings overlap ("of items you've both rated, 75% of items were rated the same"), you keep track of the performance of recommendations made from that user ("of Alice's items that we recommended to Bob, Bob upvoted them 25% of the time"). I haven't thought about this super deeply yet, but I think to beat that, you'd have to actually provide valuable data to the system in the first place.
- jacobobryant 6y ago
  To generalize that a bit, I guess I'm talking about a reputation network. You could visualize it with a directed graph that has edges weighted from -1 to 1 (0 default). Recommendation + rating history is used in some way to assign those weights, and then future recommendations are taken primarily from other users who have high weights from your perspective.

jacobobryant 6y ago

I've thought about that a fair amount also. I think the key is personalization. If bad actors downvoted a bunch of items, it would only affect people who already shared a lot of similar votes with those actors. That is still exploitable: if I know you like A, B, and C, (perhaps those are popular items), I could also upvote those items and then downvote another item D. I have an idea for countering this: keep track of history. So instead of just measuring ratings overlap ("of items you've both rated, 75% of items were rated the same"), you keep track of the performance of recommendations made from that user ("of Alice's items that we recommended to Bob, Bob upvoted them 25% of the time"). I haven't thought about this super deeply yet, but I think to beat that, you'd have to actually provide valuable data to the system in the first place.

jacobobryant 6y ago
To generalize that a bit, I guess I'm talking about a reputation network. You could visualize it with a directed graph that has edges weighted from -1 to 1 (0 default). Recommendation + rating history is used in some way to assign those weights, and then future recommendations are taken primarily from other users who have high weights from your perspective.