Ask HN: Simple projects to implement machine learning

11 points by fjellfras 14 years ago | 12 comments

I have been working through Tom Mitchell's book on machine learning and I want to supplement the theory with a practical project.

Does anyone have experience with small projects which helped them in a similar situation? The most I have come up with is building a news classifier which will use the feedzilla and NY Times api but if anyone has other good ideas please let me know.

I have reasonable programming experience with python and have access to a linode so I will be using these to implement the project.

Thank you

asit 14 years ago |

I used lua to create a simple AI. This AI can learn words and store them in a dictionary. It can link one word with another, having some resemblance or relationship. Also I implemented a "emotional link" between them; one word can be either in a GOOD or BAD relationship with another. Like the word "rat" will have a BAD relationship with "cat". It is based on meaning of words, rather than acting like a mere processor of vocabulary crawler! This AI can communicate with me.. asks me questions that it generates from meaningful concepts. Ofcourse, it is very basic right now and makes errors. But it will grow. The bottomline is that you must implement "from the roots". Perception and recognition are built over these roots. Hope you got on idea what I am attempting to build. One thing most required right now is adding a solid grammer engine to it. If I were not a lazy programmer, I would have done it already. :D

jfaucett 14 years ago |

I built an image interpretation app once that was really interesting from a machine learning perspective. Using image magick I extracted shapes and based on the layout I was able to feed that into a "composition" algorithm that judged whether the image layout was good or not. I also implemented a "plant detection" feature to find out whether a plant was in the image or not. I think images is an area that could still benefit a lot from machine learning, i.e. facial recognition, image search (extracting meaning from images). thats just some of my ideas. machine learning is awesome I wish you all the best!!

fjellfras 14 years ago | |

Thank you, that sounds very interesting. I assume you would use known images with plants in them as the training set.

anujkk 14 years ago |

How about creating a recommendation engine for HN articles? It should recommend articles upvoted by users that have similar likes/dislikes as me.

I would also like to browse HN articles/comments on basis of topics. For example, right now if I want to find HN posts related to "machine learning" I use HN Search. Can you make it better by using machine learning techniques so that I can get all relevant HN articles on a given topic sortable on basis of upvotes and time?

fjellfras 14 years ago | |

Thank you, that sounds like a good idea. I may start looking into the HN api. I have been working to build a search engine for reddit using their api for a week now (not very mature, I have only started to build a database of tags for the posts) but I believe this idea will work here as well.

anujkk 14 years ago | | |

May I know which books/courses/blogs you used to learn machine learning? I'm learning it myself and I have just finished reading "Programming Collective Intelligence".

tzm 14 years ago |

How about analyzing a database of approx 400k car parts requests? Analyze key words, year-make-model, geographic location, etc?

I'm currently doing this to generate a chloropleth map for PartsLine.com. https://skitch.com/tzmartin/eydmm/partsline.com-rfq-to-fips-...

ig1 14 years ago |

Why not sign-up to kaggle and do some of their challenges, that way you can also benchmark yourself against others.

fjellfras 14 years ago | |

Thanks for the suggestion, I actually did sign up to kaggle a couple of days ago and am working on some of their problems. I was looking for something more internet related as getting in data from the internet and grouping or ranking it feels more immediately fun to me.