Thursday, April 28, 2011
Wednesday, April 27, 2011
Classifier Showdown « Synaptic
Saturday, April 23, 2011
Friday, April 22, 2011
Tuesday, April 19, 2011
math - Mathematics for AI/Machine learning ? - Stack Overflow
- Logic - An Investigation of the Laws of Thought (Boole) and Set Theory and Logic (Stoll)
- Computation - Introduction to the Theory of Computation (Sipser)
- Probablility -
algorithm - Help Understanding Cross Validation and Decision Trees - Stack Overflow
Monday, April 18, 2011
Notes on Path Finding problem
astar c
.)matrix[u][v] = 1
denotes an edge between u and v, and matrix[u][v] = 0
denotes no edge between u and v. Then (matrix)^3
(just simple matrix exponentiation) is 'magically' the path matrix of exactly length 3.- An efficient implementation of Dijkstra's algorithm takes O(Elog V) time for a graph with E edges and V vertices.
- Hosam Aly's "flood fill" is a breadth first search, which is O(V). This can be thought of as a special case of Dijkstra's algorithm in which no vertex can have its distance estimate revised.
- The Floyd-Warshall algorithm takes O(V^3) time, is very easy to code, and is still the fastest for dense graphs (those graphs where vertices are typically connected to many other vertices). But it'snot the right choice for the OP's task, which involves very sparse graphs.
Raimund Seidel gives a simple method using matrix multiplication to compute the all-pairs distance matrix on an unweighted, undirected graph (which is exactly what you want) in the first section of his paper On the All-Pairs-Shortest-Path Problem in Unweighted Undirected Graphs [pdf].
(This is not exactly the problem that I have, but it's isomorphic, and I think that this explanation will be easiest for others to understand.)
Suppose that I have a set of points in an n-dimensional space. Using 3 dimensions for example:
A : [1,2,3] B : [4,5,6] C : [7,8,9]
I also have a set of vectors that describe possible movements in this space:
V1 : [+1,0,-1] V2 : [+2,0,0]
Now, given a point dest, I need to find a starting point p and a set of vectors moves that will bring me todest in the most efficient manner. Efficiency is defined as "fewest number of moves", not necessarily "least linear distance": it's permissible to select a p that's further from dest than other candidates if the move set is such that you can get there in fewer moves. The vectors in moves must be a strict subset of the available vectors; you can't use the same vector more than once unless it appears more than once in the input set.
My input contains ~100 starting points and maybe ~10 vectors, and my number of dimensions is ~20. The starting points and available vectors will be fixed for the lifetime of the app, but I'll be finding paths for many, many different dest points. I want to optimize for speed, not memory. It's acceptable for the algorithm to fail (to find no possible paths to dest).
Update w/ Accepted Solution
I adopted a solution very similar to the one marked below as "accepted". I iterate over all points and vectors and build a list of all reachable points with the routes to reach them. I convert this list into a hash of <dest, p+vectors>, selecting the shortest set of vectors for each destination point. (There is also a little bit of optimization for hash size, which isn't relevant here.) Subsequent dest lookups happen in constant time.
Sunday, April 17, 2011
Regularization for high dimensional learning Course
www.disi.unige.it/dottorato/corsi/RegMet2011/
Thursday, April 7, 2011
Welcome to Social Web Mining Workshop, co-located with IJCAI 2011
International Workshop on Social Web Mining
Co-located with IJCAI, 18 July 2011, Barcelona, Spain
Sponsored by PASCAL 2
News: the submission deadline has been extended to 20 April 2011.
Introduction:
There is increasing interest in social web mining, as we can see from the ACM workshop on Social Web Search and Analysis. It is not until recently that great progresses have been made in mining social network for various applications, e.g., making personalized recommendations. This workshop focuses on the study of diverse aspects of social networks with their applications in domains including mobile recommendations, service providers, electronic commerce, etc.
Social networks have actually played an important role in different domains for about a decade, particularly in recommender systems. In general, traditional collaborative filtering approaches can be considered as making personalized recommendations based on implicit social interaction, where social connections are defined by some similarity metrics on common rated items, e.g., movies for the Netflix Prize.
With the recent development of Web 2.0, there emerges a number of globally deployed applications for explicit social interactions, such as Facebook, Flickr, LinkedIn, Twitter, etc. These applications have been exploited by academic institutions and industries to build modern recommender systems based on social networks, e.g., Microsoft's Project Emporia that recommends tweets to user based on their behaviors.In recent years, rapid progress has been made in the study of social networks for diverse applications. For instance, researchers have proposed various tensor factorization techniques to analyze user-item-tag data in Flickr for group recommendations. Also, researchers study Facebook to infer users' preferences.
However, there exist many challenges in mining social web and its application in recommender systems. Some are:
- What is the topology of social networks for some specific application like LinkedIn?
- How could one build optimal models for social networks such as Facebook?
- How can one handle the privacy issue caused by utilizing social interactions for making recommendation?
- How could one model a user's preferences based on his/her social interactions?
Topics:
The workshop will seek submissions that cover social networks, data mining, machine learning, and recommender systems. The workshop is especially interested in papers that focus on applied domains such as web mining, mobile recommender systems, social recommender systems, and privacy in social web mining. The following list provides examples of the types of areas in which we encourage submissions. The following comprises a sample, but not complete, listing of topics:
- Active learning
- Matchmaking
- Mobile recommender systems
- Multi-task learning
- Learning graph matching
- Learning to rank
- Online and contextual advertising
- Online learning
- Privacy in social networks
- Preference learning or elicitation
- Social network mining
- Social summarization
- Tag recommendation
- Transfer learning
- Web graph analysis
Louhi 2011
Wednesday, April 6, 2011
2011 IEEE GRSS Data Fusion Contest
WorldView-2 multi-spectral multi-angular acquisitions and
participating to the Contest. The deadline for the paper submission is
May 31, 2011. Final results will be announced in Vancouver (Canada) at
the 2011 IEEE International Geoscience and Remote Sensing Symposium.
Check the IGARSS 2011 abstract at http://slidesha.re/gLagLW
About the IEEE GRSS Data Fusion Contest:
The Data Fusion Contest has been organized by the Data Fusion
Technical Committee of the Geoscience and Remote Sensing Society of
the International Institute of Electrical and Electronic Engineers and
annually proposed since 2006. It is a contest open not only to IEEE
members, but to everyone.
This year the Data Fusion Contest aims at exploiting multi-angular
acquisitions over the same target area.
Five WorldView-2 multi-sequence images have been provided by
DigitalGlobe. This unique data set is composed by five Ortho Ready
Standard Level-2 WorldView-2 multi-angular acquisitions, including
both 16 bit panchromatic and multi-spectral 8-band images. The imagery
was collected over Rio de Janeiro (Brazil) on January 2010 within a
three minute time frame. The multi-angular sequence contains the
downtown area of Rio, including a number of large buildings,
commercial and industrial structures, the airport and a mixture of
community parks and private housing.
Since there are a large variety of possible applications, each
participant can decide the research topic to work with. Each
participant is required to submit a full paper in English of no more
than 4 pages including illustrations and references by May 31, 2011.
Final results will be announced in Vancouver (Canada) at the 2011 IEEE
International Geoscience and Remote Sensing Symposium.
2011 DigitalGlobe - IEEE GRSS Data Fusion Contest
--
♥ ¸¸.•*¨*•♫♪♪♫•*¨*•.¸¸♥
Monday, April 4, 2011
Greg Mankiw's Blog: Advice for Grad Students
- Don Davis gives some guidance about finding research topics.
- John Cochrane tells grad students how to write a paper.
- Michael Kremer provides a checklist to make sure your paper is as good as it can be.
- David Romer gives you the rules to follow to finish your PhD.
- David Laibson offers some advice about how the navigate the job market for new PhD economists.
- John Cawley covers the same ground as Laibson but in more detail.
- Kwan Choi office advice about how to publish in top journals.
- Dan Hamermesh offers advice on, well, just about everything.
- Assar Lindbeck tells you how, after getting that first academic post, to win the Nobel prize.
Saturday, April 2, 2011
Friday, April 1, 2011
Books for Reinforcement Learning:
--