A new research paper published by Google describes a dramatically new way to improve how web pages are ranked. This algorithm claims significant improvements to deep neural network algorithms that calculate relevance.
The new algorithm discusses a method of ranking web pages called, Groupwise Scoring Functions.
Without confirmation from Google we cannot know for certain if it is in use. But because significant improvements are claimed by the researchers, in my opinion it is not far-fetched to consider that this algorithm may be in use by Google.
Does Google Use Published Algorithms?
Google has stated in the past that “Google research papers in general shouldn’t be assumed to be something that’s actually happening in search.”
Google rarely confirms which algorithms described in patents or research papers are in use. That is the case with this algorithm.
Is this Algorithm Part of the March 2019 Core Update?
This research paper shows how Google is focusing on understanding search queries and understanding what web pages are about. This is typical of recent Google research.
Google has recently introduced a broad core update that is reported to be among the biggest in years. Is this algorithm a part of that change? We don’t know and we will likely never know. Google rarely discusses specific algorithms.
In my opinion, it’s possible that something like this could be one part of a multi-part update of Google’s search ranking algorithm. I don’t believe it’s the only one. I believe the March 2019 Core Ranking Algorithm consists of a series of improvements.
Why this Algorithm is Important
The research paper begins by noting that machine learning algorithms label and give values to web pages individually, each web page in isolation from other web pages. Then the algorithms score the web pages in competition with the other web pages to find out which web page is most relevant.
Here’s how the research paper describes how current algorithms work:
“While in a classification or a regression setting a label or a value is assigned to each individual document, in a ranking setting we determine the relevance ordering of the entire input document list.”
The research paper then proposes that considering the age of all of the relevant web pages can give a clue as to what users want. So instead of ranking all the web pages one against the other, by reviewing the age of the web pages first, the ranking algorithm can better understand what a user wants and choose a better web page.
This is how the research paper describes the new algorithm:
“The majority of the existing learning-to-rank algorithms model such relativity at the loss level using pairwise or listwise loss functions. However, they are restricted to pointwise scoring functions, i.e., the relevance score of a document is computed based on the document itself, regardless of the other documents in the list.
…the relevance score of a document to a query is computed independently of the other documents in the list. This setting could be less optimal for ranking problems for multiple reasons.”
Cross-document Comparison
The research paper then shows how the current method of ranking web pages is missing an opportunity to improve the relevance of search results.
This is the example the research paper uses to illustrate the problem and the solution:
“Consider a search scenario where a user is searching for a name of a musical artist. If all the results returned by the query (e.g., calvin harris) are recent, the user may be interested in the latest news or tour information.
If, on the other hand, most of the query results are older (e.g., frank sinatra), it is more likely that the user wants to learn about artist discography or biography. Thus, the relevance of each document depends on the distribution of the whole list.”
In this example, the age of the web pages that are relevant to the search query can help to refine which answer is the best answer.
Modeling Human Behavior for Better Accuracy
The research paper then notes that search engine users tend to compare search results relative to other web pages. They then suggest that a ranking model that does the same thing is more accurate.
Leave a Reply