Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

What do you think of my startup idea?

news.ycombinator.com

Post reply on HN
Print view View original

1–10 of 44 posts

What do you think of my startup idea?

#1

Post by steffon »

I studied a lot of sociology in college and noticed a way that one could organize communities around their shared preferences of content so that people could discover new content with what should be a much higher degree of relevance then current recommendation systems.

I'm calling it a discovery engine, where the user can enter the name of a specific piece of content they have in mind, or something they are generally interested in, and receive recommendations of new content from like-minded people.

Just like Wikipedia issued a call to all people interested in making an encyclopedia, this would issue a call to all early adopters to be recognized as authorities and trend-setters. Think The Tipping Point by Malcolm Gladwell, but taking place online, efficiently, and transparently.

The project calls for combining social-bookmarking and user-generated media with an algorithm that both aggregates similar collections of content into networks and makes recommendations of content based on the evolving network structure. The ranks of "influence" and "in the know" are measured against networks of users with similar collections of content. These two rankings incentivize users to continually post relevant content because they want to remain "influential" and "in the know" in front of the people that are genuinely interested in the same content. The majority of users coming to Topiat for recommendations receive relevant recommendations fueled by the work of those who are genuinely influential and in the know.

But unlike current recommendation systems, which are domain specific and treat an individual as the sum of all their preferences (e.g. Netflix), the discovery engine would allow users to create networks based on all types of content (any combination of music, products, images, videos, URLs etc.) and enables users to explore different interests they have with the ability to create multiple groups of content on their profile. Each group of content becomes aligned with similar groups of content, from which recommendations are generated and delivered to the user (e.g. my oldies music compared with users with similar tastes in oldies music, my surfing group compared with other users that think of surfer the same way I do).

I'm putting together a Y combinator funding proposal based on this basic idea and am looking for feedback before I send it in. If there are any developers that like the idea and want to know more, let me know. Additionally, if you are good with machine learning techniques (e.g. neural networks) and are interested, let me know.

Re: What do you think of my startup idea?

#2

Post by dhouston »

what's a one or two sentence explanation that would get my little sister to use this? what value would she get within the first 2 minutes (assume early on, before you have a ton of people using it?)

ideas like this are tough because they're only really useful once an incredible amount of content has been submitted. what tricks can you do to make the site be sticky to the first 100 users, when nothing yet is submitted? what are the first concrete things the user sees, or does, when they arrive at the site?

i think that's why these recommendation sites are only successful in niches (e.g. travel (tripadvisor), food/entertainment (yelp), movies (netflix)).

Re: What do you think of my startup idea?

#3

Post by steffon »

dhouston wrote: post #2

what's a one or two sentence explanation that would get my little sister to use this? what value would she get within the first 2 minutes (assume early on, before you have a ton of people using it?) ideas like this are tough because they're only really useful once an incredible amount of content has been submitted. what tricks can you do to make the site be sticky to the first 100 users, when nothing yet is submitted…

I've got a few ideas, such as front-loading the site with content and preference data by using differnt API's, such as Flickr's API, Youtube's, del.icio.us, Last.fm's. This way when you showup, lots of your preferences and rating "work" comes along with you. Additionally, the preference dataset is jumpstarted.

What you would tell your little sister about the site is "find more things you like, and if you're good at finding new content, be recognized for it."

Re: What do you think of my startup idea?

#4

Post by dcurtis »

Here's an example of how I would use something like this: I want to watch a movie that's similar to Jurassic Park. I can go to the site and type that in. The web application picks apart the constituent pieces of "jurassic park," like "action movie" and "dinosaurs" and "made a lot of money," then looks for other movies that match the criteria based on information from IMDB, Google, and MPAA.

Then, it looks for people who have voted/commented on the movie, and extrapolates a probability of you liking other movies that they have rated highly or poorly. If your rating history is similar to theirs, it rates their other movies with similar pieces higher. If your history is very different, or if they voted Jurassic Park very poorly, then the other items they voted highly are placed low on your recommendations.

I think that's awesome.

Re: What do you think of my startup idea?

#5

Post by anamax »

What determines "groups of content" (people? algorithms? either way, the value is in figuring out how)

It isn't clear why "domain-specific" recommendation systems don't work within a group.

Note that it isn't obviously true that separating by group actually produces better recommendations. In fact, the (inadequate) evidence that I'm aware of indicates the exact opposite.

BTW - If you don't have the algorithms mentioned, how much do you think that you actually have? (I can imagine lots of wonderful things that would come from a personal transportation device that got 100mpg, but if I don't know how to make one....)

Re: What do you think of my startup idea?

#6

Post by oditogre »

Sounds like IRC. Once you find a chan you like and kind of settle in, you realize that the actual convo rarely relates to the title of the chan, but is still of interest to the people who regularly go there.

IRC is pretty popular, I don't see why your idea wouldn't be among the less BitchX-inclined internet crowd. ;-P

You'd have to be brutal with the 'influence' system, though, to keep the trolls out of the 'ponies and barbie dolls' group.

The killer feature, to me, of an idea like this would be something like musicovery.com, where overlapping things lead you down new paths.

Re: What do you think of my startup idea?

#7

Post by steffon »

anamax wrote: post #5

What determines "groups of content" (people? algorithms? either way, the value is in figuring out how) It isn't clear why "domain-specific" recommendation systems don't work within a group. Note that it isn't obviously true that separating by group actually produces better recommendations. In fact, the (inadequate) evidence that I'm aware of indicates the exact opposite. BTW - If you don't have the algorithms mention…

"groups of content" are collections of content that users create themselves (the social bookmarking aspect of the site). Users that want to build a reputation for being in the know and influential are incentivized to make these collections relevant. The groups also become lists users can form to bookmark content they find around the web and want to save it one spot.

Based on what you have put into your collection of content, and collections others have made, the algorithm aggregates those similar people into the same network. People in the same network get content recommended to each other from their like-minded peers that they have not discovered on their own.

With regards to the feasability of such an algorithm, I've talked about it with many mathematicians and machine learning programmers and the wheel does not have to be reinvented for this application. The tools already exist, and just have to be customized and tweaked for this application.

Re: What do you think of my startup idea?

#8

Post by anamax »

steffon wrote: post #7
anamax wrote: post #5

What determines "groups of content" (people? algorithms? either way, the value is in figuring out how) It isn't clear why "domain-specific" recommendation systems don't work within a group. Note that it isn't obviously true that separating by group actually produces better recommendations. In fact, the (inadequate) evidence that I'm aware of indicates the exact opposite. BTW - If you don't have the algorithms mention…

"groups of content" are collections of content that users create themselves (the social bookmarking aspect of the site). Users that want to build a reputation for being in the know and influential are incentivized to make these collections relevant. The groups also become lists users can form to bookmark content they find around the web and want to save it one spot. Based on what you have put into your collection of…

In other words, I make a "group of content" that I call "movies" and it gets compared with other peoples' "group of content" that they called "movies".

Why isn't the netflix recommendation system useful for generating recommendations from "groups of content" labelled "movies"?

If it works for movies, why can't it combine "groups of content" labelled "science fiction"?

Re: What do you think of my startup idea?

#9

Post by steffon »

oditogre wrote: post #6

Sounds like IRC. Once you find a chan you like and kind of settle in, you realize that the actual convo rarely relates to the title of the chan, but is still of interest to the people who regularly go there. IRC is pretty popular, I don't see why your idea wouldn't be among the less BitchX-inclined internet crowd. ;-P You'd have to be brutal with the 'influence' system, though, to keep the trolls out of the 'ponies a…

Here are two large differences between IRC:

1.) One of the main problems I am trying to solve is in your second sentence: "Once you find a chan you like..." It's time consuming to find new content you like, especially in a setting as nebulous (for most people) as IRC. With the discovery engine, all a user needs to do is rate recommendations based on what they are interested in, and they are immediately connected with groups of like-minded people and what they know. Fast and easy.

2.) In addition to connecting to groups of like-minded people, the user gets recommendations from those who are most influential and most in the know. All content is not equally desirable, and users using the discovery engine will only get recommended that content which has empirically shown is wanted by people that think like them (relevant).

Post reply on HN