Lonely points

[shanishalgi]

Hi,
I'm trying to use LargeVis to visualize my doc2vec features of 20NG (7532 test documents, 100 features each). I'm using all the default parameters, and I get the following result.

I was surprised by he lonely points in the data as their corresponding documents were not noticeably different than the others in their category. I tried running the algorithms after taking these documents out of the dataset, but got a similar pattern of results - a few ~6-9 lonely points representing seemingly normal documents. I previously modeled this data using various TSNE methods and none showed such a pattern of results. I am wondering if there is a simple explanation or something I am overlooking?

Also, plot.py only works for me if I change in row 29: vec[1], vec[2] to vec[0], vec[1]

Thanks in advance
Shani

[lferry007]

Hi,

I have fixed the bug in plot.py. Please try it!

Your visualization indeed looks odd. Could you please share with us the feature vectors you were using? Thanks!

[shanishalgi]

Attached is my feature vector file (and their labels)
ng20test_features.txt
ng20test_labels.txt

[lferry007]

Hi,

Using the feature vectors you shared and default parameter settings, we get the following visualization.

Since your data set is relatively small, we try to decrease the parameter "-neigh" to 50 and the parameter "-perp" to 20, and get visulization like this:

It seems to look better. (We don't try other settings further.)

[shanishalgi]

Hi,
Has the code been changed since I uploaded my data? Because I used the
default settings and got a very different result.

On Wed, Sep 14, 2016 at 5:17 PM, lferry007 [email protected] wrote:

Hi,

Using the feature vectors you shared and default parameter settings, we
get the following visualization.
[image: ng]
https://cloud.githubusercontent.com/assets/15796471/18515282/63320bb2-7ac7-11e6-8335-8d4a86925c11.png

Since your data set is relatively small, we try to decrease the parameter
"-neigh" to 50 and the parameter "-perp" to 20, and get visulization like
this:
[image: ng_neigh50_perp20]
https://cloud.githubusercontent.com/assets/15796471/18515328/924aae5e-7ac7-11e6-999f-549cfd2f75ce.png

It seems to look better. (We don't try other settings further.)

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#2 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AK99YzYjoOR-mxMmDSnVMi__A-7aEM-2ks5qqAHngaJpZM4JOiKR
.

[tangjianpku]

We've updated the code and you can have a try. If there is still a problem, it may be the problem of system configuration.

Thanks,
Jian
https://sites.google.com/site/pkujiantang/home