Ranking Algorithm

The score is calculated based on a logarithmic scale between 0 and 10 where

10 represents the most trustworthy web source there can be.

What are Ranking Algorithms?

 Ranking algorithms are computational processes used to order items, such as web pages,
products, or multimedia content, based on their relevance or importance to a given query or
user.
 These algorithms determine the priority or sequence in which items are presented to the
user, aiming to maximise utility, relevance, and user satisfaction.

Ranking Algorithms Evolution.

Initially, basic techniques like keyword matching were used to rank content. However, as the volume
and complexity of data increased, more sophisticated algorithms were developed.

 Early Days: relied on basic keyword-matching techniques. These methods ranked pages
based on the frequency and location of keywords within the text.

 PageRank: A significant breakthrough came with Google’s introduction of PageRank in 1998.

Developed by Larry Page and Sergey Brin, PageRank assessed the importance of web pages
based on the number and quality of links pointing to them. This method revolutionised web
search by leveraging the web’s hyperlink structure to improve relevance.
  Modern Algorithms: Today, ranking algorithms have become highly complex, incorporating
machine learning, user behaviour analysis, and contextual signals. Google’s algorithms, for
instance, now use hundreds of factors, including user engagement metrics and AI models like
BERT, to deliver more accurate and personalised results.

Why are Ranking Algorithms Important?

Ranking algorithms are crucial in various industries, influencing user experience, engagement, and
business outcomes.

 Search Engines: The most prominent application of ranking algorithms is in search engines.
Algorithms like Google’s determine the order of search results, directly impacting user
satisfaction and information retrieval efficiency.

 E-commerce: In e-commerce platforms like Amazon, ranking algorithms sort products based
on relevance to search queries, user preferences, and purchase history, thereby enhancing
the shopping experience and driving sales.

 Social Media: Social media platforms like Facebook, Instagram, and Twitter use ranking
algorithms to prioritise content in users’ feeds. These algorithms consider user interactions,
content popularity, and recency to keep users engaged.

 Recommendation Systems: Services like Netflix, Spotify, and YouTube rely on ranking
algorithms to suggest content to individual user tastes, significantly influencing consumption
patterns and user retention.

Ranking algorithms are the backbone of many digital platforms, shaping how information is accessed
and consumed.

PageRank (PR) is a Google algorithm that ranks web pages in search results by evaluating the number
and quality of links to a page. It operates on the principle that pages receiving more high-quality links
are deemed more important and are thus ranked higher.

How it’s calculated

Here’s the original PageRank formula:

PR(A) = (1-d) + d (PR(T1)/C(T1) + ... + PR(Tn)/C(Tn))

where

 A is the analyzed page

 T1…Tn are the pages pointing to the analyzed page

 C is the number of links placed on the analyzed page

 d is a damping factor that corresponds to the probability that a user will abandon a page
(usually set to 0.85)
 1. A page is given an initial PageRank score based on the links pointing to it.
Example
five pages with no links. Each gets a PageRank of (1/5) or 0.2.

2. This score is then distributed to other pages through the links on the page. If some links are added
to the five pages above and calculate the new PageRank for each, then
3. This calculation is repeated as Google crawls the web. If we calculate the PageRank again (called
an iteration), the scores change. It’s the same pages with the same links, but the base PageRank for
each page has changed, so the resulting PageRank is different.

The PageRank formula also has a so-called “damping factor,” the “d” in the
What are the Types of Ranking Algorithms?

Ranking algorithms come in various forms, each designed to address specific needs and contexts.

 content-based ranking
 collaborative filtering
 hybrid models
 learning to rank.

Content-Based Ranking

Content-based ranking algorithms assess the relevance of items based on their content and how well
they match the query or user preferences. These algorithms analyse features such as keywords,
metadata, and other textual or multimedia content to determine relevance.

The core idea is to compare the content of each item to the query or user profile. For example, the
algorithm evaluates how well a web page’s text matches the search terms in search engines. In
recommendation systems, it might analyse product descriptions or movie genres.

 Examples:

 PageRank: Developed by Google, PageRank evaluates the importance of web pages

based on the number and quality of links pointing to them, assuming that more
essential pages will likely receive more links.

 BM25: A probabilistic information retrieval model that scores documents based on

term frequency, document length, and inverse document frequency, improving
relevance ranking in search results.

Collaborative Filtering

Collaborative filtering algorithms rely on user interactions to make recommendations or rankings.

They analyse patterns in user behaviour to predict what items a user might like based on similar
users’ preferences.
Collaborative filtering can be user-based or item-based. User-based collaborative filtering finds users
with similar tastes and recommends items they have liked. Item-based collaborative filtering
recommends items similar to items the user has liked.

 Examples:

 Netflix Recommendation Algorithm: Uses collaborative filtering to suggest movies

and shows based on similar users’ viewing history and ratings.

 Amazon Product Recommendations: Leverages user purchase history and browsing

patterns to recommend products that similar users have bought or viewed.

Hybrid Models

These models often integrate content-based and collaborative filtering methods to provide more
accurate and diverse recommendations.

Combining different algorithms, hybrid models can address limitations such as the cold start
problem in collaborative filtering (where there is insufficient data on new users or items) and the
overspecialisation problem in content-based systems (where recommendations may lack diversity).
  Examples:

 Google’s Hummingbird Algorithm: Integrates various signals, including content

relevance, user intent, and contextual factors, to deliver more accurate search
results.

 Spotify’s Recommendation System: Combines collaborative filtering, natural

language processing of song lyrics, and audio analysis to recommend music.

Learning to Rank

Learning to rank (LTR) algorithms utilise machine learning techniques to optimise ranking. These
algorithms are trained on large datasets to learn the optimal way to rank items based on various
features.

LTR approaches can be pointwise, pairwise, or listwise. Pointwise methods predict the relevance of
individual items, pairwise methods optimise the relative order of item pairs, and listwise methods
consider the entire list of items for optimisation.

 Examples:

 RankNet: A pairwise learning-to-rank algorithm developed by Microsoft that uses

neural networks to predict the ranking order of document pairs.

 LambdaMART: An extension of the LambdaRank algorithm using gradient-boosted

decision trees to optimise ranking based on user interactions and relevance
judgments.