Google android-app's titles and soft clustering

Question

I have a no-trial question: I want to soft cluster the apps from Google Store. Most of the parameters are numbers, so no big clue. There are also "tags" but this is like using categorical features. I can use a one-hot encoding.

The pain is about the titles, which still contain a lot of useful info.

To keep the feature-vector small, I'm attempting several solutions, but nothing really satisfying.

I did a previous work using a sentence embedding solution from Facebook IA, which works great, however, the 8K vector does not seem a good idea for soft-clustering.

How would you "convert" the title string into a vector suitable for a later soft clustering algorithm?

Answer 1

There are two approaches:

• The "classic" one, where you use something like bag-of-words, maybe with manual features (e.g. detect that something is a "brand" by matching the names of the brands of products), etc.
• The "modern" approach would be to use a language model (e.g. BERT) to transform the sentences into a latent representation. You can think of the language model in this scenario as using an encoder in autoencoder, that takes the text as input and returns a representation in the form of a numeric vector of a pre-specified size. Those embeddings are known to encode the meanings in the sentences, though the results would be better if the model was trained on similar data as yours, so using an off-the-shelf pre-trained model may not create great features.