Hi,
Lukas Vlcek wrote:
Enis,
Thanks for your time.
I gave a quick glance at Pig and it seems good (seems it is directly based
on Hadoop which I am starting to play with :-). It obvious that a huge
amount of data (like user queries or access logs) should be stored in flat
files which makes it convenient for further analysis by Pig (or directly by
Hadoop based tasks) or other tools. And I agree with you that size of the
index can be tracked in journal based style in separated log rather then
with every since user query. That is for the easier part of my original
question :-)
The true art starts with the mining tasks itself. How to efficiently use
such data for bettering user experience with the search engine... one potion
is to use such data for search engine tuning, which is more technical
oriented application (speeding up slow queries, reshaping the index, ...).
But I am looking for more IR oriented application of this information. I
remember that once I read on Lucene mail list that somebody suggested
utilization of previously issued user queries for suggestions of
similar/other/related queries or for typo checking. While there are well
discussed methods (moreLikeThis, did you mean, similarity, ... etc) in
Lucene community I am still wondering if once can use user search history
data for such purpose and if the answer is yes then how (practical examples
are welcomed).
Well, the logs of the search engine is used to improve the search engine
in several ways. All the major search engines including google, yahoo
and ms uses the logs to deliver more relevant results to the user. How
they do this is a big research area. Although google tends to not
publish it's algorithms, yahoo and MS do.
The logs can be used to improve various components of the search engine.
For example the suggesters (google suggest, yahoo suggest ) and the
spell checkers. Most of the spell checkers uses the noisy channel error
model. In this model, the user emits word w with a probability p(w). But
due to the error in the channel (that is misspelling) the user emits w'
instead. Spelling correction deals with correcting w' to w. The
calculation involves edit distance, prior probabilities of words,
probabilities of errors(prob. of writing "spel" instead of "spell" )
and a predefined lexicon. But in the word of internet, there in no
lexicon, so given a word w' you may not know exactly if w' is misspelled
or not. The logs can be used for calculating the probabilities, building
the lexicon, and generating possible suggestions(in case of a misspelled
word).
For spell checking you can consult :
An improved Error model for noisy channel spelling correction, Brin et.al.
Learning a spelling error model for search query logs, Ahmad et.al.
spelling correction for search engine queries. Martis. et.al
techniques for authomatically correcting words in text. Kukish (an
excellent review of topic)
Spelling correction as an iterative process that exploits the collective
knowledge of web users. Brill et.al. (This MS paper is very good)
For the query suggestion, yahoo has a paper about using Machine learning
methods using the log data, for query suggestion. The basic idea behind
is that, the user submits a query, then if not satisfied with the
results, refines the query and resubmits it. Looking at all the
collective data we can find possible better suggestions for a submitted
query, classify them as useful or not useful or irrelevant, then display
the useful ones. You can find the paper at yahoo research site.
using the web log for improving search engine quality is a much harder
problem. Unfortunately, i could not find time to read more about this
topic yet. I know that MS uses a method based on neural nets, called
ranknet, which ranks the search results. The net is trained with the
server logs. Below i list some papers, but i have not read all of them
so i cannot say anything further about them :
Accurately Interpretting Cilckthrough Data as Implicit Feedback
A Simulated Study of Implicit Feedback Models
Identifying "Best Bet" Web Search Results by Mining Past User Behavior
Improving Web Search Ranking by Incorporating User behaviour Information
Learning User Interaction Models for Predicting Web Search Result
Preferences
Optimizing_search_engines_using_clickthrough_data
Query Chains: Learning to Rank from Implicit Feedback
Lukas
On 8/10/07, Enis Soztutar <[EMAIL PROTECTED]> wrote:
Lukas Vlcek wrote:
Hi Enis,
Hi again,
On 8/10/07, Enis Soztutar <[EMAIL PROTECTED]> wrote:
Hi,
Lukas Vlcek wrote:
Hi,
I would like to keep user search history data and I am looking for
some
ideas/advices/recommendations. In general I would like to talk about
methods
of storing such data, its structure and how to turn it into valuable
information.
As for the structure:
==============
For now I don't have exact idea about what kind of information I
should
keep. I know that this is application specific but I believe there can
be
some common general patterns. as of now I think can be useful to keep
is
the
following:
1) system time (time of issuing the query) and userid
2) original user query in raw form (untokenized)
3) expanded user query (both tokenized and untokenized can be useful)
4) query execution time
5) # of objects retrieved from index
6) # of total object count in index (this can change during time)
7) and possibly if user clicked some result and if so then which one
(the
hit number) and system time
Remember that you may not want to store all the information available
at
runtime of the query, since it may result in great performance burden.
For example you may want to store the raw form of the query, but not
parsed form since you can later parse the query anayway (unless you
have
some architecture change). Similarly 6 seemed not a good choice for
me(again you can store the info externally). You can look at the common
and extended log formats which are stored by the web server.
The problem is that all the information do chance in time. The index is
updated continuously which means that expanded queries and total number
of
documents in index do change as well. But you are right that getting
some of
this info can cause extra performance expenses (then it would be
question of
later optimization of architecture design).
Well i think you can at least store the size of the index in another
file, and log to the changes in the index size from there. The
motivation for this comes from storage efficiency. You may not want to
store the same index size over and over again in n queries before the
index size changes, but store it once, with the time, per change.
As for the information I can get from this:
=============================
Such minimal data collection could show if the search engine serves
users
well or not (generally said). I should note that for the users in this
case
the only other option is to not use the search engine at all (so the
data
should not be biased by the fact that users are using alternative
search
method). I should be able to learn if:
1) there are bottleneck queries (Prefix,Fuzzy,Proximity queries...)
2) users are finding what they want (they can find it fast and results
are
ordered by properly defined relevance [my model is well tuned in terms
of
term weights] so the result they click is among first hits)
3) user can formulate queries well (do they issue queries which return
all
index documents or they can issue queries which return just a couple
of
documents)
4) ...?... etc...
Web server log analysis is a very popular topic nowadays, and you can
check for the literature, especially clickthrough data anaysis. All the
major search engines has to interpret the data to improve their
algorithms, and to learn from the latent "collective knowlege" hidden
in
web server logs.
It seems I have to do my homework and check CiteSeer for some papers :-)
Is there any paper you can recommend me? Some good one to start with?
What I want to achieve is far beyond the scope of the project I am
working
on right now thus I cannot spend all my time on research (in spite of
the
fact I would love to) so I can either a) use some tool which is already
available (open sourced) and directly fits my needs (I don't think there
is
any tool which I could use out-of-box) or b) implement something new
from
scratch but with just very limited functionality.
You do not have to implement this from scratch. You just have to specify
your data mining tasks, then write scripts(in pig latin) or write
map-reduce programs (in hadoop). Either of these are not that hard. I do
not think that there is any tool which may satisfy all you information
needs. So at the risk of repeating myself i suggest you to look at pig
at write some scripts to mine the data.
Coming to literature, i can hardly suggest any specific paper, since i
am not very into the subject either. But i suggest you to skip this
step, first build you data structures (log format), then start
extracting some naive statistical information from the data. For
example, initially you may want to know
1. avarage query execution time
2. avarage query execution time per query type(boolean, fuzzy, etc.)
3. histogram of query types (how many boolean queries, etc.)
4. avarage #of queries per user session.
5. etc.
The list can go on and on depending on the data you have and information
you want. These kind of simple statistical analysis can be very easy to
extract and relatively easy to interpret.
As for the storage method:
===================
I was planning to keep such data in database but now it seems to me
that
it
will be better to keep it directly in index (Lucene index). It seems
to
me
that this approach would allow me for better fuzzy searches across
history
and extracting relevant objects and their count more efficiently (with
benefit of the relevance based search on top of history search
corpus).
I think that more scalable solution would be to keep such data in pure
flat
file and then periodically recreate search history index (or more
indices)
from it (for example by Map-Reduce like task). Event better the flat
file
could be stored in distributer file system. However, for now I would
like to
start with something simple.
I would rather suggest you to keep the logs in rolling flat files. An
access to the database for each search will take lots of time. Then you
may want to flush those logs to the db once a day if you indeed want to
store the data in a relational way.
I infer that you want to mine the data, but you do not know what to
mine, right? I suggest you to look at hadoop and pig. Pig is a is
designed especially for this purpose.
You've hit the nail on the head! I am very curious about how one can use
such data to improve user experience with search engine (given my
project
schedule time constraints).
I know this is a complex topic...
Regards,
Lukas
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
Anyway, thanks for your reply!
BR
Lukas
