Hi Srinath, Thanks for the response. On Jun 23, 2014, at 10:48 PM, Srinath Perera wrote:
> Hi Lahiru, > > Sorry for not responding earlier. I was traveling last week. > > I guess you know frequent item set != clustering algorithms. +1 > > Can we have a chat sometime to discuss details about the implementation. Yes sure, that would be very useful. Thanks Lahiru > > --Srinath > > > On Sun, Jun 22, 2014 at 6:25 PM, Lahiru Gunathilake <[email protected]> > wrote: > Hi All, > > I have implemented another frequency counting algorithm[1] which is a classic > algorithm for mining frequent items in a data stream. Basically users can > specify a minimum average value of frequent items with an error value. > > This algorithm will accept two user-specified parameters: a support threshold > s [0-1] and an error parameter e [0-1] such that e << s and recommended > number for e is normally s/10 or s/20. Let N denote the current length > of the stream, i.e., the number of tuples seen so far. At any point of time, > this algorithm can be asked to produce a list of events along with their > estimated frequencies. The answers produced by this algorithm will have the > following guarantees: > 1. All item(set)s whose true frequency exceeds sN are outputs. > There are no false negatives . > 2. No events whose true frequency is less than (s - e) N > is output. > 3. Estimated frequencies are less than the true frequencies > by at most eN. > . > The incoming stream is conceptually divided in to buckets of width w = > ceiling(1/e) transactions each. Buckets are labeled with bucket ids , > starting from 1.We denote the current bucket id by bcurrent whose value is > ceiling(N/w). For an element , we denote its true frequency in the stream > seen so far by "fe" . Note that e and w are fixed for a data stream while N, > bcurrent and fe are the variables whose value changes when the stream > progress. > Here the data structure , is a set of entries are tuples with the form of > (event, f, delta), where event is the actual event and "f" is the is an > integer representing its estimated frequency, and delta is the maximum > possible error in "fe". > > In this algorithm every new event will anyways added in to the > data-structure optimistically and there is no initial condition to enter in > to the data-structure but iteratively if certain events are not match to the > condition provided > by the user those events will be removed when N%windowSize = 0, during this I > output those events as expired-events in the window. For the incoming events > if event is already exists I output those as current-events and if its a new > event I just add it to the data-structure and iterate through all the events > available and find the matching events based on s and e values and only those > events will be output. > > I am not sure based on the window definition this is the correct approach or > may be windows aren't the best way to associate this algorithm in to siddhi. > I have attached my patch to jira[2] and if you can look that would be great. > > Siddhi Query will looks like below, > from cseEventStream#window.lossyFrequent(0.1,0.01) " + > "select symbol, price > " + > "insert into > StockQuote; > > > [1]Gurmeet Singh Manku and Rajeev Motwani. 2002. Approximate frequency counts > over data streams. In Proceedings of the 28th international conference on > Very Large Data Bases (VLDB '02). VLDB Endowment 346-357. > > Regards > Lahiru > On Jun 19, 2014, at 2:36 AM, Lahiru Gunathilake wrote: > >> Hi All, >> On Jun 17, 2014, at 1:49 AM, Lahiru Gunathilake wrote: >> >>> Hi All, >>> >>> I am planning to evaluate different event stream clustering algorithms as >>> part of my studies(I am a graduate student at indiana University). I think >>> Siddhi is a good place to experiment this, As per my understanding based on >>> the docs Siddhi doesn't have a stream clustering interface I can use >>> directly to plug my own algorithm. So I am thinking of first come up an >>> interface for different clustering algorithms and add implementation of >>> algorithms for each event stream by invoking an operation like >>> SiddhiManager.addQuery. Or I can make the algorithm configure as part of >>> query language. If the second option is more consistent with current model >>> I can wrap-up the work in that way but initially focussing on first >>> approach will be easier for me. So each algorithm can be associated to a >>> desired event Stream or can be associated globally. If its associated with >>> each stream algorithm will run local to each stream otherwise it will run >>> in global context. Based on the algorithm I can provide a way to configure >>> it with parameters. >>> >> I am sure I have confused with above implementation details, after looking >> in to Siddhi extension points I figured out I just have to implement a new >> window type. I have implemented one algorithm to keep the most frequent >> events >> came in a event stream. So queries can looks like below, >> >> from cseEventStream#window.frequent(2) " + >> "select symbol, price >> " + >> "insert into >> StockQuote; >> >> There are multiple algorithms to keep the most frequent events in a given >> window size for now I just implemented a simple algorithm[1] with the >> processing complexity of O(1) and space complexity O(n) where n is the limit >> of the most frequent items. I have created a patch and attached it to >> jira[2]. >> >> [1] Jayadev, and David Gries Misra, "Finding >> repeated elements," in Science of computer >> programming 2, no. >> 2 (1982): 143-152. >> [2]https://wso2.org/jira/browse/CEP-877 >> >> Thanks >> Lahiru >>> To start this I hope to implement a frequent item set mining algorithm >>> which can be used to find out most frequent items of an event stream. >>> Search engines use these kind of data to find out most frequent searches in >>> a given time window and optimize the search queries. I can start with some >>> algorithms like Misra-Gries algorithm[1] and Manku and Motwani [2] and >>> then move towards more of data clustering algorithms. For the time being I >>> will write the clustering results in to a file and later I think I can use >>> more stable storage (either wso2 registry or other prefered way in wso2 >>> product stack). If Siddhi or WSO2 CEP already have the capability of >>> frequent item mining I will start with a more classification type algorithm. >>> >>> Your feedback will be very useful for my work. If you have requirement for >>> any specific type of algorithms based on the real client interactions you >>> have, I would like to know them and implement them with Siddhi and do the >>> comparison. >>> >>> Thanks >>> Lahiru >>> _______________________________________________ >>> Architecture mailing list >>> [email protected] >>> https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture >> >> _______________________________________________ >> Architecture mailing list >> [email protected] >> https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture > > > _______________________________________________ > Architecture mailing list > [email protected] > https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture > > > > > -- > ============================ > Srinath Perera, Ph.D. > http://people.apache.org/~hemapani/ > http://srinathsview.blogspot.com/ > _______________________________________________ > Architecture mailing list > [email protected] > https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture
_______________________________________________ Architecture mailing list [email protected] https://mail.wso2.org/cgi-bin/mailman/listinfo/architecture
