1
$\begingroup$

I'm working on a project that involves building a news recommendation system. I've come as far as quantifying user interaction with different articles on the site into user's affinity towards atopic using a bayesian function. I also have quantified the recent articles using LDA into the proportion an article talk about each topic.

my users topic-affinity for a user x looks like this(target-x):

 user_id  interest-topic-0  interest-topic-1  interest-topic-2  interest-topic-3  interest-topic-4  interest-topic-5  interest-topic-6  interest-topic-7  interest-topic-8  interest-topic-9 
       0            0.0257            0.2956            0.0386            0.0643            0.1285            0.0000               0.0            0.0257            0.0386            0.1671  

My quantified articles looks something like this(vectors-v):

post_id   topic-0   topic-1   topic-2   topic-3   topic-4   topic-5   topic-6  topic-7  topic-8   topic-9
      x  0.055048  0.000000  0.742544  0.032286  0.059630  0.000000  0.000000  0.01173      0.0  0.095441
      y  0.000000  0.051172  0.000000  0.000000  0.158314  0.042632  0.022281  0.00000      0.0  0.720676
      z  0.028615  0.000000  0.020919  0.000000  0.000000  0.018940  0.882862  0.00000      0.0  0.046078

The shape of target will always be (10,)

The shape of vectors will always be (num_articles, 10)

Both vectors do not follow the same distribution.

Now I'm trying to figure out the best way to recommend articles from vectors v, for a target user x, given x and v. I've tried distance similarity functions like cosine similarity to find the distance between vectors. The results are satisfactory but I'm looking for a better function/metric to pick out top n recommendations for a user.

$\endgroup$

1 Answer 1

0
$\begingroup$

Other ways to measure real-valued vector distances:

  • Euclidean distance
  • Manhattan distance
  • Higher order Minkowski distance
  • Chebyshev distance

Beyond just trying different distance metrics, you can try re-framing the problem or using A/B testing to see which recommendations are most empirically useful for users.

$\endgroup$

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.