Hi Rishabh,<br><br>I had this feeling before. This is a really nice idea BUT we can not go ahead with the project which is still not tested in the experimental settings. Though it may be a wonderful research exercise, I would still vote to go for the state-of-the-art methods which are completely published with full details and experimental results. Hope you get my point.<br>
<br>Regards,<br>Parth. <br><br><div class="gmail_quote">On Wed, Apr 4, 2012 at 1:55 PM, Rishabh Mehrotra <span dir="ltr"><<a href="mailto:erishabh@gmail.com">erishabh@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Hi Parth,<div>Please find below the conversations I had with Hang Li sir and Prof. Liu regarding my proposed methodology.</div><div>Regards,<br><br><div class="gmail_quote">---------- Forwarded message ----------<br>From: <b class="gmail_sendername">Tie-Yan Liu</b> <span dir="ltr"><<a href="mailto:Tie-Yan.Liu@microsoft.com" target="_blank">Tie-Yan.Liu@microsoft.com</a>></span><br>
Date: 2012/4/3<br>Subject: RE: Doubt regarding Feature selection for 'Learning to Rank' algorithms<br>To: "Hang Li (MSR)" <<a href="mailto:hangli@microsoft.com" target="_blank">hangli@microsoft.com</a>>, Rishabh Mehrotra <<a href="mailto:erishabh@gmail.com" target="_blank">erishabh@gmail.com</a>><br>
<br><br>
<div link="blue" vlink="purple" lang="EN-US">
<div>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">In my opinion, features definitely play an important role in learning to rank. Combining ListMLE with deep learning sounds interesting. We are eager to see
the performance of your implementation. <u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<div>
<p class="MsoNormal" style="text-align:justify;text-justify:inter-ideograph"><span style="font-size:10.5pt;font-family:"Calibri","sans-serif";color:#1f497d">Thanks<u></u><u></u></span></p>
<p class="MsoNormal" style="text-align:justify;text-justify:inter-ideograph"><span style="font-size:10.5pt;font-family:"Calibri","sans-serif";color:#1f497d">Tie-Yan</span><span style="font-size:10.0pt;font-family:"Calibri","sans-serif";color:#595959"><u></u><u></u></span></p>
</div>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<div>
<div style="border:none;border-top:solid #b5c4df 1.0pt;padding:3.0pt 0cm 0cm 0cm">
<p class="MsoNormal"><b><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif"">From:</span></b><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif""> Hang Li (MSR)
<br>
<b>Sent:</b> 2012</span><span lang="ZH-CN">年</span><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif"">4</span><span lang="ZH-CN">月</span><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif"">1</span><span lang="ZH-CN">日</span><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif"">
10:33<br>
<b>To:</b> Rishabh Mehrotra<br>
<b>Cc:</b> Tie-Yan Liu<br>
<b>Subject:</b> RE: Doubt regarding Feature selection for 'Learning to Rank' algorithms<u></u><u></u></span></p>
</div>
</div><div><div>
<p class="MsoNormal"><u></u> <u></u></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">Rishabh<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">Thank you for your interest.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">I am not quite sure whether it is easy to make improvement by combining deep learning and ListMLE. But you can try.
<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">To me, ListMLE is a very elegant model and since it is a log linear model, it appears to have a good match with deep learning.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">Maybe you can also get some comment from Tie-Yan.<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d">Hang<u></u><u></u></span></p>
<p class="MsoNormal"><span style="font-size:11.0pt;font-family:"Calibri","sans-serif";color:#1f497d"><u></u> <u></u></span></p>
<p class="MsoNormal"><b><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif"">From:</span></b><span style="font-size:10.0pt;font-family:"Tahoma","sans-serif""> Rishabh Mehrotra
<a href="mailto:[mailto:erishabh@gmail.com]" target="_blank">[mailto:erishabh@gmail.com]</a> <br>
<b>Sent:</b> Saturday, March 31, 2012 1:03 AM<br>
<b>To:</b> Hang Li (MSR)<br>
<b>Subject:</b> Doubt regarding Feature selection for 'Learning to Rank' algorithms<u></u><u></u></span></p>
<p class="MsoNormal"><u></u> <u></u></p>
<p class="MsoNormal">Hello sir,<u></u><u></u></p>
<div>
<p class="MsoNormal"><u></u> <u></u></p>
<div>
<p class="MsoNormal">I attended your talk on Learning to Rank at MLSS 2011 at NUS Singapore last year in June. I was going through various Listwise approaches for ranking and <u></u><u></u></p>
</div>
<div>
<p class="MsoNormal">the various features used to represent the documents.<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal"><u></u> <u></u></p>
</div>
<div>
<p class="MsoNormal">Recently Deep architectures have been used to learn feature representations in an unsupervised manner and have outperformed the state-of-the-art algorithms for various classification tasks. I am planning to work for an open-source organization
and help them in implementing an algorithm for their Learing to Rank module. <u></u><u></u></p>
</div>
<div>
<p class="MsoNormal"><u></u> <u></u></p>
</div>
<div>
<p class="MsoNormal">I have decided to implement the <b>ListMLE algorithm</b> which you proposed in 2008. Instead of applying the conventional IR features I am thinking of using
<b>Autoencoders</b> or other deep learning algorithm for feature extraction. It would really help me if you could comment on this decision: will it help me improve the performance or is it that features do not play that major a role in this sub-field of IR.<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal"><u></u> <u></u></p>
</div>
<div>
<p class="MsoNormal">Any suggestions form your side would really help me a lot. Thank you for your time.<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal"><u></u> <u></u></p>
</div>
<div>
<p class="MsoNormal">Best regards,<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal">Rishabh Mehrotra,<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal">BITS Pilani,<u></u><u></u></p>
</div>
<div>
<p class="MsoNormal">India.<u></u><u></u></p>
</div>
</div>
</div></div></div>
</div>
</div><br><font color="#888888"><br clear="all"><div><br></div>-- <br>Rishabh.<br><br>
</font></div>
</blockquote></div><br>