- The long list of articles at the bottom of the page seems excessive; why not just show 10 most recent? expand when needed; it just clogs the UI
- Good HTML should read like any other good source code, if the logical, visual, and functional pieces of HTML are broken into disorganized pieces it will be hard for a new member of your team to keep up or to maintain it
- CSS doesn't have any comments [EDIT: this is just for showing your code - if you want good feedback let me understand what you are trying to do - geez]
- You should do some simple clustering or at least use some epsilon in you document similarity matrix to distinguish if an article is new or not; if same or very similar article shows on three different web-sites will it have triple effect on the prices? For example:
- Instead of showing 10 most recent articles it's better to show N-most recent articles with S-most recent trend; this might be difficult to implement but seems more helpful for making a decision; otherwise the user has to do extra work to figure out: "this is an old news, I know that two days ago they thought otherwise, but now there is new trend"
- This is not my domain of expertise and I'm not sure how traders work but this might be useful in healthcare so please keep working on it keeping in mind that this could work much better in different domains
[EDIT:] Try shared neighbor distance method. Whatever you use now for similarity feed it into Jarvis & Patrick method and try different NN. This will take care of outliers and variable densities throughout heterogeneous clusters.
I'm no expert in webdev, but while CSS comments may be useful for maintenance and for new team members, they are also a waste of bandwidth. I'm honestly curious, what HN webdevs do with comments in CSS - do you use them, or not, or maybe have some deployment scripts that strip comments off?
Ideally, you would use a tool like less or sass to generate your css (which we don't), so your comments would be in the underlying file but not in the css given to the user.
Maybe it's useful to support the site, while it's small it's ok, wait 2 more years... We are a very Tech centric community you should expect that kind of comment from HN folks.
Actually, we are doing clustering, but as you found out it's not perfect. Like you said, we don't want the same news repeated over and over again to be factored in the bullbear index.
I like this idea: Instead of showing 10 most recent articles it's better to show N-most recent articles with S-most recent trend.
One more idea. You could add a "system error button" or something similar that a user can click and tell you how your NLP/ML algorithms work. Then, use it for improving your models or maybe even implement active learning.
- The long list of articles at the bottom of the page seems excessive; why not just show 10 most recent? expand when needed; it just clogs the UI
- Good HTML should read like any other good source code, if the logical, visual, and functional pieces of HTML are broken into disorganized pieces it will be hard for a new member of your team to keep up or to maintain it
- CSS doesn't have any comments [EDIT: this is just for showing your code - if you want good feedback let me understand what you are trying to do - geez]
- You should do some simple clustering or at least use some epsilon in you document similarity matrix to distinguish if an article is new or not; if same or very similar article shows on three different web-sites will it have triple effect on the prices? For example:
http://www.insidefutures.com/articles/article.php?id=345842
http://www.mrswing.com/articles/Grain_Market_Analysis_from_J...
- Instead of showing 10 most recent articles it's better to show N-most recent articles with S-most recent trend; this might be difficult to implement but seems more helpful for making a decision; otherwise the user has to do extra work to figure out: "this is an old news, I know that two days ago they thought otherwise, but now there is new trend"
- This is not my domain of expertise and I'm not sure how traders work but this might be useful in healthcare so please keep working on it keeping in mind that this could work much better in different domains
[EDIT:] Try shared neighbor distance method. Whatever you use now for similarity feed it into Jarvis & Patrick method and try different NN. This will take care of outliers and variable densities throughout heterogeneous clusters.