0
$\begingroup$

I've been searching for a model that can take as input an arbitrary number of feature vectors. As in the title, wanted to build a model that would be able to pick the best candidate from a set. For each record the size of a set could differ (so the number of classes would differ as well).

Since NLP deals with sets of arbitrary length, my first attempt was to use something like RNN but I wanted it to be insensitive to order. I'm aware of bidirectional RNNs. I've also found a 'bag of words' implementation in rnn and ffn.

bag of words implementation in RNN

neural bag of words model


An example (an artificial one)

We've got a number of movies.

For each movie we've got a set of reviews (the number of reviews varies). Only one of them is positive, the rest is negative.

Of course, one could create a model for each individual review but one can do better. The idea is that the information from the other reviews might significantly improve the model.

Zero padding won't probably do the trick, the number of reviews varies greatly.


Update & a possible solution

There's a well known paper about rankNET, where the authors use shared weights across the network for a learn-to-rank problem (check out Siamese networks). One could use the idea (shared weights) to build a model for the multiclass classification I've described above.

While I was not able to implement it in keras, it is quite easy to derive the backpropagation algo for these kinds of networks.

It's easy to teach the network, even if there are different sizes of input vectors for every record. I'm still unsure how it will perform though.

$\endgroup$
2
  • 2
    $\begingroup$ What do you mean by sensitivity to order? Did you actually mean direction? Because RNNs are pointless if there's no order, they're designed to tackle sequences. Sequences are ordered by definition, otherwise they would be sets or bags $\endgroup$
    – Aksakal
    May 14, 2019 at 17:08
  • $\begingroup$ It was a poor choice of words, sorry. Indeed, sets/bags are better, thank you. $\endgroup$ May 14, 2019 at 19:01

1 Answer 1

1
$\begingroup$

It sounds like might want to look into Pointer nets - the output is a set of pointers to different locations in the input. They’ve been used for things like finding boundary points for convex sets and ordering sequences of numbers.

Even with Pointer nets (and basically any type of architecture, as far as I know) require sequence inputs to be padded to one uniform length, so “arbitrary length” doesn’t totally apply but you can get away with pretty long padding with Pointers.

$\endgroup$
0

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