Using Deep Learning at Scale in Twitter’s Timelines | Hacker News Reader

Using Deep Learning at Scale in Twitter’s Timelines

blog.twitter.com

62 pointshunglee227 comments

It does not work for my timeline at all. Maybe the idea of algorithmic newsfeed is ok, I dreamed of it in the times of RSS and Livejournal. But last few days twitter became almost unusable. It shows "You might like" almost every time I open it, and it is filled with content that I hate. Shady political posts, glamour chicks posting cans of cosmetics. It shows tweets liked by someone as if it's retweeted instead of liked.

Maybe it's optimised for rage which drives more page refreshes? Twitter started to look like it's really dying, desperately trying to squeeze last ad juices from its userbase.

It is a disappointment that this article merely scratches the surface and didn't give even slightest details. It is more like PR than a tech blog because it merely shares anything substantial.

"We have really cleverly done a thing that no one actually wanted" is the summary

But noone actually wants a "relevant timeline". They want all the posts by their friends most recent first, and that's it. Twitter's contributions to the creation and reinforcement of echo chambers has done the entire world an enormous disservice.

This is such a backwards-looking point of view.

0. The fact is, I want a algo timeline. There are far too many posts in a day for me to read everything.

1. People made the same complaints about Facebook's algorithmic feed. Very few users complain about that anymore (outside of HN, I guess). If anything, without FB's algo feed, users would be overwhelmed with a flood of posts. FB will surface very granular posts like a friend liking something, so we are talking about multiple thousands of potential news feed posts per day. And if FB were to take out that granularity, that means potentially missing tens of highly relevant posts a day.

2. Obviously I don't have any inside info, but I guarantee you Twitter split tested the algorithmic timelines and is also measuring user engagement and they wouldn't have kept this highly complex feature (with probably a whole engineering team dedicated to it) unless there was a big boost.

3. Q1 of 2017, the first quarter they rolled out the algo timeline, user growth accelerated.

4. I don't know a single person with experience growing/measuring consumer web products who thought the algo timeline was a bad idea. It is such an obvious piece of low-hanging fruit.

My first thought when I saw what twitter was doing with my timeline was, "oh man, now they've gone ahead and ruined my twitter feed just like they ruined my Facebook feed."

A similar timeline thing recently appeared on G+.

Not a Twitter or FB user, but do visit G+ once a week or so for a few minutes. Or did. Now everything is out of order and a jumble of crazy including surfacing my own long past contributions over more recent ones. I doubt I'll be back. I can't speak directly to Twitter timelines but if it's anything similar I wouldn't use it.

There may be a point to reducing noise and promoting relevant feeds. Maybe it has to happen to keep the view-able volume to a manageable level. Maybe. Or maybe the strategy was get everyone in the door and hooked, then start controlling what they see. And that is hugely valuable to both commerce and government.

A/B testing aside, my prediction is that removing user control over what is seen backfires hugely and may provide an opening for competing services. Until they manage to lock the internet down to 4 approved services anyway. Which probably has at least been considered.

Point 2 - Are you serious? People are not complaining because they're not on HN, if they were this kind of people they would be here downvoting you. Facebook users don't even remember they had a different timeline before, they don't understand and don't want to understand why some post shows and some other doesn't, they just accept it as a fact of life.

Point 2 - Yes, the original article actually confirms that they did A/B testing

"Impact on people using Twitter is typically measured by running one or more A/B tests and comparing the results between experiment buckets. The set of metrics we use here usually relate more directly to usage and enjoyment of Twitter."

Those people think it is a good idea because they can sell the service of gaming it!

People have accepted these timelines as a fait accompli but that's all.

It's hard to think of a recent "improvement" by Twitter that hasn't made the user experience worse. The overarching goal seems to be to reduce the user's control and increase Twitter's control over what users see. I personally use Twitter much less as a result.

Likewise. I don't actually want to see what I like, I want to see what's really happening. I can't do that if their algorithm is distorting my view, so I just don't use it anymore.

> But noone actually wants a "relevant timeline"

You are deeply incorrect about that, and the attachment to chronological timelines is probably what doomed Twitter for so long.

I absolutely want a relevant timeline. So do most people. Most people don't want to spend all day glued to Twitter to keep up with ongoing conversations. They want to be able to quickly open it and find content which intrigues them and doesn't confuse them.

Twitter listening too much to power users for years and years is what kept them from building products the masses could love.

It's insane to me that people on HN think chronological timelines are good. Should I have to read hundreds or thousands of posts a day (which is definitely how much content my friends produce) just to find the 5 or 6 which matter to me. Why wouldn't I want a machine to do that sorting for me?

Exactly. Why should Twitter choose which of my friends' posts I'm allowed to see ?

> But noone actually wants a "relevant timeline".

I want that

That's exactly what I was thinking!

I thought that until I started checking it once a week or so, then I really liked it

Why isn't a working method of disabling these sorts of sorting 'features' ever provided?

Seems like they could please the "I want to see relevant" and the "I want to see most recent" camps.

Settings > Timeline > uncheck "Show me the best tweets first"

Does that not work?

I do not see that option anywhere.

On web it's under Account > Content (the bottom section).

> We measure a model’s quality in two ways. First, we evaluate the model using a well-defined accuracy metric we compute during model training. This measure tells us how well the model performs its task – specifically, giving engaging Tweets a high score.

Very curious to know what this metric would be. Maybe some combination of likes, retweets, and viewing time? Or maybe including a general liking term (i.e. liking a person's tweet gives a small boost to all their tweets).

Anecdotal experience with twitter's timeline suggests it's viewing time - and that they haven't considered that people are just spending more time finding the tweets they're looking for...

It's amazing how many bright minds are paid to sort lists in a particular way.

s/sort lists in a particular way/find the particular way to sort lists//.

Please don't do this, Twitter. This is a step away from growth. There are so many problems you should fix. A ranked feed will ruin the quality of content and it will become another mindless Facebook/Instagram clone. I don't want this and I love Twitter.

Ironically, with people using its new tool to make Big Text On A Gradient Background built into the post box, Facebook is now looking more and more like Playskool Twitter.

Reply on news.ycombinator.com