Incipia blog

Google Play’s App Store Keyword Ranking Algorithm Update

Gabe Kwakyi | November 14, 2016

Feature Image Credit: Google Android Blog

Google, the king of algorithms and algorithm updates (i.e. cute and cuddly animal names, massive disruptions in organic ranks), is cooking up some improvements to its play store keyword ranking algorithm.

The new changes are aimed at improving relevance of apps returning for so-called "broad" searches, or non-app name searches like "horror games" or "selfie apps." Per Google's words, about 50% of play store searches are broad, and:

"Searches by topic require more than simply indexing apps by query terms; they require an understanding of the topics associated with an app. Machine learning approaches have been applied to similar problems, but success heavily depends on the number of training examples to learn about a topic. While for some popular topics such as “social networking” we had many labeled apps to learn from, the majority of topics had only a handful of examples. Our challenge was to learn from a very limited number of training examples and scale to millions of apps across thousands of topics, forcing us to adapt our machine learning techniques."

Google's article explains that their initial attempt at building a machine learning algorithm capable of serving good results for these broad searches used a deep neural network, which didn't quite work as they had hoped for new app discovery, instead producing the same apps over time in response to broad searches, rather than new apps.

Google's new attempt is to make the process more like the way that humans learn and come to understand language and word associations. This new attempt makes use of  the Skip-gram model, which is a model to predict related words given an input word. Google's new model creates so-called "classifiers" for any given word, to create many lists of classifier relationships, and ultimately {app, topic} associations. In its latest update, Google will also be leaning on non-machine learning efforts, by having people rate the quality of results.

skip gram analysis

To the left are some sample relationships between words, as identified by a Skip-Gram analysis, per Tensor Flow documentation.

Image Credit: TensorFlow.org

 

Google's goal is to create an algorithm that could generate a sensible relationship between keywords (e.g. {photo} and {share}), and, through studying app metadata and user interactions, produce the most relevant app for a given keyword, even if the app returned is brand new. Additionally, Google's algorithm must be capable of learning new words (e.g. selfie, fleek, etc.) and building new associations with those words and other words and apps.

It seems that, despite some early issues with over-generalization, Google is pressing on with this effort to improve broad search results for play store users. It will be interesting to see how these changes play out in keyword ranks (and downloads) for all Android apps.

The Bottom Line: With 50% of play store searches classified as "broad" (e.g. selfie apps) rather than app names, Google is using machine learning plus human input to boost its keyword ranking algorithm's ability to return relevant apps when users use broad searches for new app discovery. This likely means big changes are in store for play store keyword rankings.

Original Soure: Google Blog