Notes on codes, projects and everything
So apparently Annoy is now splitting points by using the centroids of 2 means clustering. It is claimed that it provides better results for ANN search, however, how does this impact regression? Purely out of curiosity, I plugged a new point splitting function and generated a new set of points.
(more…)After a year and half, a lot of things changed, and annoy also changed the splitting strategy too. However, I always wanted to do a proper follow up to the original post, where I compared boosting to Annoy. I still remember the reason I started that (flawed) experiment was because I found boosting easy.
(more…)While following through the Statistical Learning course, I came across this part on doing regression with boosting. Then reading through the material, and going through it makes me wonder, the same method may be adapted to Erik Bernhardsson‘s annoy algorithm.
(more…)Traversing a tree structure often involves writing a recursive function. However, Python isn’t the best language for this purpose. Therefore I started flattening the tree into a key-value dictonary structure. Logically it is still a tree, but it is physically stored as a dictionary. Therefore it is now easier to write a simple loop to traverse it.
In the previous post, I re-implemented Annoy in 2D with some linear algebra maths. Then I spent some time going through some tutorial on vectors, and expanded the script to handle data in 3D and more. So instead of finding gradient, the perpendicular line in the middle of two points, I construct a plane, and find the distance between it and points to construct the tree.
Recently I switched my search code to Annoy because the input dataset is huge (7.5mil records with 20k dictionary count). It wasn’t without issues though, however I would probably talk about it next time. In order to figure out what each parameters meant, I spent some time watching through the talk given by the author @fulhack.
So I first heard about Panda probably a year ago when I was in my previous job. It looked nice, but I didn’t really get the chance to use it. So practically it is a library that makes data looks like a mix of relational database table and excel sheet. It is easy to do query with it, and provides a way to process it fast if you know how to do it properly (no, I don’t, so I cheated).
I need a slide show script for my portfolio pages but couldn’t find a good one anywhere so I decided to write one myself. The slide show script will be able to display image and the respective description in a predefined order. However, in this version, visitors would not be able to directly jump to a particular slide yet. The script is written in prototype‘s object-orientation approach hence you need to have prototype called.
This post is purely based on my own speculation as there’s no experiment on real-life data to actually back the arguments. I am currently trying to document down a plan for my experiment(s) on recommender system (this reminds me that I have not release the Flickr data collection tool :/) and my supervisor advised to write a paragraph or two on some of the key things. Since he is not going to read it, so I might as well just post it here as a note.
With most of my stuff more or less set, I guess it is time to start documenting the steps before I forget. So I heard a lot of good things about docker for quite some time, but haven’t really have the time to do it due to laziness (plus my relatively n00b-ness in the field of dev-ops). Just a few months ago, I decided to finally migrate away from webfaction (thanks for all the superb support) to a VPS so I can run more things on it.
Folksonomy is a neologism of two words, ’folk’ and ’taxonomy’ which describes conceptual structures created by users [4, 5]. A folksonomy is a set of unstructured collaborative usage of tags for content classification and knowledge representation that is popularized by Web 2.0 and social applications [1, 5]. Unlike taxonomy that is commonly used to organize resources to form a category hierarchy, folksonomy is non-hierarchical and non-exclusive [3]. Both content hierarchy and folksonomy can be used together to better content classification.
This update took me quite a bit more time than I initially expected. Anyway, I have done some refactoring work to the original code, and thought it would be nice to document the changes. Overall, most of the changes involved the refactoring of function names. I am not sure if this would stick, but I am quite satisfied for now.