[Xapian-devel] Some questions about Letor project

[Parth Gupta]

Hi Jiarong,

>
>>    1. In letor.cc, we have two parts of functions: the training part and
>>    the ranking part. I’ll use SVMRanker as an example. The training part
>>    basically uses the libsvm library and training data to train a model, then
>>    save the model file. The ranking part will calculate score for each
>>    document in searching results (MSet) by using the trained model file. My
>>    question is for each of our three rankers: 1) SVMRanker 2) ListMLE 3)
>>    ListNet, do we need three different types of training part? (The ranking
>>    part for each of those have the same form I think) I’m not sure the
>>    parameters for these three different rankers are the same or not (I guess
>>    they’re different). In my understanding, the letor.cc basically just pass
>>    parameters ranker. It’s the ranker will do training and calculating things
>>    actually. So if we can generalize the form for training part, we don’t need
>>    functions like prepare_training_data_for_svm,
>>    prepare_training_data_for_listwise etc. We just need  prepare_training_data
>>    instead. (We can benefit from inheritance of ranker in training part just
>>    like in ranking part)
>>
>> In general, I think we need a different training part for each ranker.
>  There may be some similarities in these existing rankers, and inheritance
> would be a sensible way to avoid duplicating code if so, but we'd like to
> have a framework which we can extend to a completely different type of
> ranker in future.
>

Ideally, we decided to have only a single method like prepare_training_file
and it would be the responsibility of the Ranker's to interpret the data
the way they want for example, pairwise approaches need pairs and  so on.
The data format we have decided is the standard one and commonly used among
Letor community. Example taken from the SVM-rank page (
http://www.cs.cornell.edu/people/tj/svm_light/svm_rank.html). So at this
moment I would say, please focus on only one method and remove the others.
This should also be communicated to Hanxiao.
<line> .=. <target> qid:<qid> <feature>:<value> <feature>:<value> ...
<feature>:<value> # <info>
<target> .=. <float>
<qid> .=. <positive integer>
<feature> .=. <positive integer>
<value> .=. <float>
<info> .=. <string

>
>>    1. There is one thing I have to confirm: once we have the training
>>    model (like model file of SVMRanker), we won’t train that model again in
>>    general. (The behavior of questletor.cc under bin/ confuses me)
>>
>> I'm not familiar with the behaviour of questletor, but I suppose it's
> reasonable to assume that we don't update models after initial creation.
>  It would be nice to be able to do so, but I think many training algorithms
> aren't updatable.  I feel I may be misunderstanding your question here,
> though.  Parth: any comment to add?
>

 questletor is just an example of how the code works. Once the model is
trained, you dont need to retrain it unless you really want to. So for the
better interpretation you can add a condition in questletor that train only
when model does not exist.

>
>>    1. Since RankList will be removed, according to the meeting last
>>    week, its related information will be stored under MSet::Internal. My plan
>>    is to create new class under MSet::Internal. That class will have two kinds
>>    of feature vectors: normalized one and unnormalized one. Since it’s in
>>    MSet::Internal, there is a wrapper class outside it I think. So it also
>>    needs to provide corresponding APIs in that wrapper class. Also, the ranker
>>    will use MSet instead of RankList. Do you have any suggestions for this
>>    part?
>>
>> This sounds like a reasonable approach.  This sounds like something you
> could implement very soon, and that is sufficiently standalone we could try
> and get it merged to master on its own.
>

I am not sure if you really need to store the normalised feature vector,
just a method to normalize should do the job. We should definitely consult
Hanxiao when he comes to a point where he sees storing a normalised version
will help in some way. Btw, which type of normalisation methods are you
talking about? If you refer to QueryLevelNorm (
http://trac.xapian.org/wiki/GSoC2011/LTR/Notes#QueryLevelNorm) then you
that is the standard and your featurevector would be like that. Do you mean
to further normalise it?

>
>>    1. For FeatureVector, I think it could be discarded since it just
>>    stores the information of feature vector of  each document, those
>>    information will be stored in the new class in MSet::Internal mentioned in
>>    3.
>>
>> Sounds right to me.
>

Okay, sounds fair but also please store the additional information such as
score and label as the featurevecor class currently does.

>
>>    1. For FeatureManager, it will set the context for feature
>>    calculation, like set Database, set query and what kinds of features we
>>    want. It provides some basic information like term frequency and inverse
>>    document frequency etc. Also it will have function update_mset to touch
>>    feature information to MSet.
>>
>> Again, sounds plausible.
>

Btw we decided not to categorise features based on types like document
dependant, query dependent etc in the end but we agreed to give user the
power to select a subset of features may be in form of a list<Integer> or
something.

>
>>    1. For feature selection, I don’t know when to apply this selection.
>>    We will provide the features we want to use to FeatureManager. So the
>>    feature selection will provide some information like this feature is better
>>    so it will have larger weight? Or this algorithm will select subset of
>>    features we provide to generate feature vectors?
>>
>> I'd expect the feature selection to select a subset of features: but it's
> also very good for it to be able to return information that a human can
> check over to see if it's making plausible decisions.
>

Both the feature selection algorithms mentioned on the Letor ProjectIdea
page are subset selection based. It is one time and happens before the
training. These algorithms will give each feature a score that how
important each feature is and the user needs to select top N features based
on some educated heuristics presented in the corresponding paper and the
computational power at disposal.


>
>>    1. For automated tests, my idea is to use some data to test the
>>    functionality of letor module. It will also cover different configurations,
>>    like using different rankers, to test the functionality. I think I need
>>    some help for this part. Can someone give me some advice?
>>
>> I'm not sure what advice you need; Parth - any ideas here?
>

The test concerning to xapina-letor would be mainly focused around the
features and the rankers. So what you can do is use a small test collection
with a few documents and check if the features calculated are correct or
not, the ranking using each ranker is acceptable or not etc.

Cheers,
Parth.

>
> _______________________________________________
> Xapian-devel mailing list
> Xapian-devel at lists.xapian.org
> http://lists.xapian.org/mailman/listinfo/xapian-devel
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.xapian.org/pipermail/xapian-devel/attachments/20140523/e467406a/attachment.html>