4 ms·
It has been for a long time. PageRank was never designed for adversarial scenarios. It reminds me of KPIs like lines of code - it's only useful if it cannot b
by yarg 4y ago
It has been for a long time.
PageRank was never designed for adversarial scenarios.
It reminds me of KPIs like lines of code - it's only useful if it cannot be manipulated.
"When a measure becomes a target, it ceases to be a good measure."
https://en.wikipedia.org/wiki/Goodhart%27s_law https://en.wikipedia.org/wiki/Goodhart%27s_law
- PaulHoule 4y agoPageRank is gameable but harder to game than some of the competing algorithms (such as a straight link count.) The real consequence PageRank had was that it got web pages to stop linking to each other…. Google gaslighted people into removing the competition they could have had navigating from one page to another through links. (e.g. web directories)
- wolpoli 4y agoIndeed, after Google, links slowly disappeared and no one could really "surf" the web anymore.
- DaiPlusPlus 4y agoGoogle didn’t gaslight the web: Web directories died out because they go stale, fast, and require far more human curation than is economically viable. I remember Yahoo effectively deprecating their directory pages long before I heard of Google, back when I was using Dogpile and CompuServe.
- PaulHoule 4y agoI never saw a web directory that took a serious approach to using automation to curate the directory. It's probably more feasible than people think because it's a matter of classifying links up or down (relevant or not) which is a much more ontologically and mathematically tractable problem than learning a ranking function or trying to classify things into one of N>2 categories. I think the problem may more have been the lack of a sustainable revenue model. Getting volunteers to curate the directory is particularly destructive because the people who most want to volunteer either (1) want to promote something or (2) want to get paid so somebody else can promote something.
- davemp 4y agoI consider Reddit to be a form of web directory in a way.
- danans 4y ago> I never saw a web directory that took a serious approach to using automation to curate the directory. That's what search indexes are. Automatically curated directories. It's just that the UI to the index is natural language queries instead of clicks through a graph, because the latter isn't scalable to large corpuses.
- PaulHoule 4y agoNo, I think a modern directory would be a set of topics in an ontology with links. If I had to seed one I would suck all the external links out of Wikipedia, impose some organizing structures (probably overlapping trees or dags) and then build a set of classifiers for nested topic relevance, spaminess, etc. There are certain sites which have a landing page for every topic in some set of topics, you could add a lot of links quickly if you built rules for importing links from particular sites. Adding 10 sites a day with 1000 links each would be very possible, in 100 days you could build out a million links.
- ctippett 4y agoIt's not quite the same as the web directories of old, but I find the links and content contained in the "Awesome [topic]" lists on GitHub to be a useful resource. The curation aspect is managed via pull requests just as any other open source project, although this doesn't make it immune from going stale.
- PaulHoule 4y agoThat's a good example and they aren't overrun with spam.
- drc500free 4y agoI'd say rather that it wasn't designed for modern adversarial scenarios. Google won the search engine wars because PageRank did better in the late 90s adversarial environment, where every page was packed with white-on-white keywords. It was significantly harder to create a counterfeit influence graph than to keyword spam with hidden text elements. That all seems pretty quaint these days, but even worms and viruses at the time were versions of "my goodness, who would ever write a script that emails itself to your address book, or copies itself across the network to all the other unsecured PCs with passwordless full hard drive access?"
- jsemrau 4y agoWas researching Ad spend as the ad-tech industry is falling apart. https://imgur.com/gallery/1Aw3t4i https://imgur.com/gallery/1Aw3t4i This is not a search result anymore. This is an outrage.
- rgbrgb 4y agoThere's maybe a good argument there for blackbox (non-interpretable) ranking algos.
- ketzo 4y agoHow would you really blackbox a search ranking, though? A website owner can always just search for themselves, and see how their ranking changes. When there’s so much money at stake, people will go to great lengths to reverse-engineer the things that put them higher in the page, no matter how you try to hide the levers that make the rankings work.
- leobg 4y agoPersonalized results. If you Google yourself, you just see what you see. But you can never know what others will see.
- nordsieck 4y ago> Personalized results. If you Google yourself, you just see what you see. But you can never know what others will see. That just turns it into a statistical problem. I don't think you can escape the problem of a search engine being used as an oracle.
- pornel 4y agoRanking has many many signals, which are mixed in non-linear ways, and are dampened and vary over time. You only get low-resolution low-frequency sampling of the result.
- rgbrgb 4y agoI mean blackbox in the sense that I can't explain how it works or look under the hood and understand it. Many machine learning models you want to be explainable for UX or debugging (e.g. Netflix's "because you watched X"). This is a rare case where it's better if you can't figure out how it produced a result. Instead, you want to throw queries and user behavior into a blackbox algo and have it tell you a result then give it feedback on whether the result was good (did the user come back and ask the same question? did they click the top result and leave or did they have to come back and click through many more?). I think this is kind of how google works now, though results are frequently meh. Millions of backlinks will get you noticed but your ranking will just keep dropping if users don't appear to find your content useful (e.g. they hit the back button a lot and keep searching).