Yea, that Excel-table-like design probably breaks all design rules.
It surprises me most when I occasionally see design professionals complementing it.
I am rather a backend programmer and I have no illusions.
But even the Hacker News audience does not have enough geekiness to understand my previous attempt at a movie recommender system: http://5000best.com/movie-galaxy/
Thank you. I was in top 50 in the Netflix Prize and wrote a book about recommender systems (http://arek-paterek.com/book/) if someone is interested how the bread is made.
Yes, the interface became a little complicated. The previous version was more accessible for a casual user, for example the following article about the website was received pretty enthusiastically (retweeted about 1000 times)
http://gigazine.net/news/20121012-5000-best-movies/
I guess the right way to fix it is to hide some features and add a tutorial.
Thanks. We'll see. Each feature costs developer's time.
Edit: the stack is G-WAN and everything is written in plain ANSI C. I wrote my own HTTP servers in C/C++ for my previous websites, but this time I tried G-WAN. I think it saved me one week of work or two.
The website is running on the cheapest virtual dedicated server from 1&1 with 25% of one processor time. The total cost is less than $10 monthly.
The Netflix Prize winners had a few RBM models that used the dates.
Regarding the DBM - I also tried to use more than one layer, and without success. I tried out 3-layer and 4-layer autoencoders (can be called 1.5-layer and 2-layer DBM), with initialization by stacked RBMs or without it. It did not work well probably because: a) the model was inaccurate, and b) the learning method proposed for DBM was not completely correct. Intuitively, the right DBM-like model with the right learning method should have a chance to improve something on the Netflix task.
I found some improvement though (rather learning time than accuracy) in the standard RBMs. Instead of using CD, I split the weights into two sets, creating a directed RBM version. The "up" weights from the visible nodes to hidden are learned with CD with T=1. The "down" weights are learned to best fit the visible nodes, using the hidden nodes as predictors. The hidden nodes generated by CD T=1 are good enough, and we do not need additional iterations with increased T.
Extending my comment, I am almost sure that passing a small negative argument (which is converted to a large positive number) causes malloc to return NULL on Linux.
This makes me think. In 2012 people still think that genetic programming, a numerical optimization technique (and not the most efficient one), magically solves random machine learning problems.
I had to choose one name per movie.
If a movie is based on a novel, I believe the author of the novel should be credited in the first place. Usually the script degrades the quality of the novel.