Where’s My Depth First Search Machine Learning? | Hacker News Reader

Where’s My Depth First Search Machine Learning?

medium.com

76 pointsgnad13 comments

Three reasons you're not going to see a change like this in the algo:

1) Recommendations are somewhat economic substitutes, not economic complements. Which is to say, most people look at the Amazon page for a given book, then look through the associated recommendations, and (generally) choose only one of the options.

2) If enough people actually bought Silent Spring after reading the Three Body problem, then the Silent Spring would already show up in the recommendations, because Amazon shows you recommendations based on what customers most similar to you also bought

3) If Amazon had shown you Silent Spring BEFORE you read the Three Body Problem, would you have even wanted to read it? Or would you have been like wtf Amazon how is this related? My bet for the average customer is the latter. Amazon can't make money by recommending you things that you won't even want until two weeks later

Right now Amazon is showing two kinds of recommendations: 1) "the customers that bought this also bought…". That one can show you books that are totally unrelated to the book you are buying or even to your actual interests. 2) the "Sponsored products related to this item". I don't see why they couldn't substitute one of them for "books the author of this book recommends", or something like that.

1) I don't see what is immutable about this logic. Next to the "people who enjoyed that also enjoyed..." list there could be a "to go a bit deeper..." or "books referenced by this book or author..."

2) The problem is in "enough". People who will want to read about that book are only a niche and will never be more numerous than people who simply bought the part II and III of the Three Body Problem (it is a trilogy). The author here really points out an unaddressed problem.

3) Another argument to make a separate "book referenced by this book/author"

Article not really related to Machine Learning. Rather clickbaity story about how the author wants to see book referenced in a book to be recommended to him prior to buying any book.

I don't see how that's clickbaity, or unrelated to machine learning. No, it's not a technical post, but it still raises an interesting point about what recommendation systems are actually optimizing for and what's missed.

Additionally, this sort of referential database could be quite useful in other pursuits. Want to map mentions within works to research how culture addresses itself? It may also capture a larger market segments, people who are interested in one topic may find they share interests with the author in another topic.

Regarding “clickbaity”, I clicked it hoping to hear about some interesting way to apply the concept of depth first search to machine learning. The actual subject, however, has nothing to do with depth first search: depth vs. breadth first has to do with the order in which links are traversed, but what the author wants is for machine learning to discover links it currently doesn’t know about at all (i.e. explicit textual references). However, this probably isn’t intentional clickbait; the author probably thought the analogy made sense. (Maybe it does somehow and I’m just not seeing it.)

I see it as the following, he views breadth first as just skimming over each book, seeing the books that are related to it when sitting side by side (perhaps when in another users cart), as opposed to depth, which goes into each item, and then goes deeper picking out books from within that book.

From what I understand, the author needs to brush up his knowledge about depth first search. Scanning the book for references to other books would normally be a part of breadth first search. Also, a bfs would yield books that are more related to the selected book than a DFS.

but those books this new method finds might not be books that customers are likely to buy, even with the new information. perhaps they are just bad books but are still referenced in such a way.

The reason you are not seeing this is that the problem is too hard for and language is very complex.

There are Banks that would pay Millions if somebody could extract business relations between companies. It would probably be enough in a fist step to know what a company name is an see if 2 turn up in a sentence. The author is asking for something that looks even harder to me. First identify book titles. And then identify a recommendation. I think machine learners like to play with images, because training is easy (recapture). But how should they go an tag book contents for recommendations. It takes much more time than identifying vehicles or traffic signs and requires more education.

Before analyzing something like that there will other things be done. Like work on the structure of sentences (who interacts). Or if some statement has a positive or negative tone.

Mine is PageRank

Bookstores are not libraries.

Reply on news.ycombinator.com