Comments & Feedback on Cut Once

The authors would like to thank all comments and suggestions.
Most bugs reported here have already been fixed.
From: Le Zhao
To: "Vitor R. Carvalho"
Date: Wed, Jan 30, 2008 at 6:18 PM
Subject: Re: Do you use Mozilla Thunderbird?

Hi Vitor,
Very nice toolkit!
Here are some experience I had when doing some quick testing: during training, I used about 1100 emails in sent box (the using inbox warning is helpful), processing of the first 1000 is fast, although the "script is slow, do you want to stop or continue" message box popped out once and interrupted the processing, I have to click continue. Then, there seems to be some large? email or some email that caused the DF and TFIDF processing becoming very slow, I have to click continue several times. One solution would be to disable the message box. Also, it would be nice if the two processings (DF and TFIDF) are merged, so that the emails are processed only once.

The buttons have cute icons. I find it difficult to press the "mail statistics" button, because I don't know what will happen after clicking it, will there be a confirmation? will there be explanations of what will be sent out?

The timer during training is helpful, but a tiny problem with it, it seems it does not update the estimate of the finishing time, so when the "stop or continue" message box popped up, interrupting it, postponing the finishing time, the timer at last goes to negative seconds, and decreasing.

The recommendation interface is very easy to use - just one click, but maybe sometimes I would like to select multiple recipients with ctrl+click. Also, after clicking one recommended recipient, the prediction scores do not update, which is helpful sometimes, but with this one more recipient added, the prediction will be more accurate, so it's worthwhile to update the predictions, and the predictions are very fast.

Some observations of the recommendation algorithm, you probably have already dealt with the problems, but it doesn't hurt to repeat :-), some highly ranked recommendations don't co-occur with the listed recipients at all in history, but they still get a very high rank, maybe because their prior is too high? if they don't co-occur in training, then there should be some large penalization. I think the co-occurrence in history is actually more important than predictions based on content. That's all I've tested now. It's difficult to test information leak in a short time..

Cheers,
-Le
From: Hideki Shima
To: email.research.cmu@gmail.com,
Date: Wed, Jan 30, 2008 at 2:38 PM
Subject: Anonymous Data to Email Research in Carnegie Mellon - Version 0.0.1
Add your comments here:

I learned this tool from William's weblog and just tested on my machine. This is a very cool software!
I especially like the 2nd part of the tool to prevent leaking, as I often see a private message going to multiple recepients because of the "reply all" button or "Reply-To: " in email header.

Some bug reports: At first, I installed the tool on Thunderbird 1.5x but the tool bar did not appear, so I got Thunderbird 2.0x and it worked. From my past user experience, I guess there's a way to warn when installing an extension on 1.5x.

I attached screenshots. Training took me a lot of time (about 10 mins) as I have a lot of emails in the Sent directory. 2nd dump shows when the tool suggested me good recepients!
-Hideki
From: Melchert Fruitema
To: email.research.cmu@gmail.com,
Date: Wed, Feb 27, 2008 at 7:22 AM
Subject: Anonymous Data to Email Research in Carnegie Mellon - Version 0.0.1
Add your comments here:

I got the invitation to share data just now, but the compose was not
filled in except for the subject. So, I hit Einstein on the main window
for this to show up. (Note: I tried Einstein yesterday, with same
result. Cancelled it.) I will send it now.
...
I believe this data was created on my sent folder containing news
messages, mostly. It mayay be interesting to see how this evolves, but I
think the model is not intended for newsgroups correspondence.

a new .dat file was created, probably replacing the previous.

(By the way, when I did send the previous share on e-mail, I saw an
error message about not able to perform an OS x close for a log, or
similar wordings.)
From: Le Zhao
To: "Vitor R. Carvalho"
Date: 3/20/2008 5:21 PM
Subject: Re:  Mozilla Thunderbird users needed for User Study

Hi Vitor,

I have an impression that people who emails different people a lot will most likely benefit from Cut Once.  For example, administrators, secretaries, etc..  I think it will be great if you could get Stacey or Constance to use your tool.

Cheers,
-Le
From: Melchert Fruitema
CC: email.research.cmu@gmail.com,
Date: Sun, Apr 06, 2008 at 3:02 PM
Subject: Re: Emails going to the wrong recipient
On 06-04-2008 20:27 CET, anonymous@anonymous.net composed this enchanting statement:
Since thunderbird handles email and newsgroups in one tree it is very easy to accidentally address an email to a newsgroup by mistake (I've done it once). If you start to compose an email to someone with a newsgroup open, you may have this problem. The only solution to prevent this may be to use another program for usenet. Any advice on this?
 
 
It happened to me too: I filed my tax form on the internet. :-[ :-$

Since then, I take care to scrutinize the addresses before submitting my messages.

There is very interesting experiment setup by a group in at the CMU University. They are developing an extension (.xpi) for assisting you in addressing the right people in a message by recommending recipients based on an analysis of your sent folder(s).

Eureka, this changes my view on having this tool operate on news messages. (Note to CMU!)

They have announced that about a month(?) ago in this newsgroup. I think a try would be worthwhile. Moreover, you can influence the development of the tool as the team is very interested to take your feedback. (I did copy them on this message.)