I read about Recursive Neural Networks that they can convert Documents to distributed word representation.

In the context of new article recommendation, I am thinking to use this model to convert all news articles to vectors and then recommend to a particular user, news articles similar to ones he browsed.

In vector space this will boil down to finding 'similar' vectors to a given vector(user's news read).

How likely is that this model will work well in practice? Any comments and/or suggestions?


There are many ways to calculate similarity between articles. I have not seen anybody doing a vector conversion and running comparisons. However, there is a text mining strategy called "Term-Frequency / Inverse-Document-Frequency" which is a clever way to find unique words and phrases in documents. You can run this on multiple documents, and compare the extracted keywords to match them for recommendations.

Check out my ebook for more details: https://lizrush.gitbooks.io/algorithms-for-webdevs-ebook/content/chapters/tf-idf.html

If you want to leverage this technique on web documents (like a blog), there is a free service to do so: https://algorithmia.com/recommends

  • 1
    $\begingroup$ But there are problems with these kinds of models. Specially since they work in discrete space, so it does not gives a good similarity measure between documents. The main reason I read was due to the use of discrete entities (bag-of-words model) any small changes in the doc will skew the distribution of doc a lot. While using Neural Embeddings it gives a continuos vector representation of a doc so the similarity measure works well in this space. $\endgroup$ – SHASHANK GUPTA Jan 7 '16 at 5:19
  • $\begingroup$ Excellent point. I'm sure there are shortcomings to TF-IDF, just as there are to any model. Hopefully sharing different ways of tackling this problem will help readers find the best solution for their document set. $\endgroup$ – sheldonkreger Jan 7 '16 at 20:28

You've just described the matrix factorization model, which works well. In fact, it even works without explicit features; i.e., any properties of the items. But you can introduce prior information by augmenting the feature vectors, e.g., with the document embedding, as you describe. I do not see any reason why it should not work, esp. since I've seen it done.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.