All major search engines use machine learning in one or more ways. With a market share of 84%, Google is known to be the largest. And uses a variety of machine learning algorithms to improve search results for users. We want to introduce you to some of the most important algorithms Google uses for Machine Learning:
- Natural language processing (NLP): Google uses NLP technologies to better understand the meaning of search queries and deliver more relevant results. This includes natural language query processing, misspelling detection, and multi-language query support.
- Recommender Systems: Recommender systems analyze user behavior and preference data to identify preferences. Google uses these recommender systems to deliver personalized search results.
- Clustering and classification: Google uses clustering and classification of (unsupervised) machine learning to rank how relevant a website is to a search query.
- Anomaly detection: Identifies unusual activity from normal within large data sets. For example, Google recognizes spam results and can remove…
- Google FLAN accelerates the transfer of learning from one domain to another and makes it less computationally intensive. Important to know: In machine learning, a domain does not refer to a website, but to the task or group of tasks it performs, such as sentiment analysis in natural language processing (NLP) or object recognition in computer vision (CV).
- V-MoE’s sole purpose is to enable training of large image processing models with fewer resources.
- Sub-pseudo-labels improves the recognition of actions in videos and helps with a variety of video-related insights and tasks.
In addition, Google still uses some core algorithms to optimize the search results of its users:
Introduced in 2015, this algorithm contributed to the fact that Google no longer considers the world as strings of keywords and characters, but as things (entities). Thus, keywords, characters, and words combined into sentences became so-called entities. An entity is a thing or concept that is unique, well-defined, and distinguishable. Previously, Google considered an input like (Frankfurt, Hesse) as two words. They regularly occur together, but just as often they occur separately. With RankBrain, the two keywords Frankfurt and Hesse became one entity. This means that if there is a context, Google recognizes the context and thus the entity (Frankfurt, Hesse) for both keywords.
BERT (Bidirectional Encoder Representations from Transformers) is somewhat younger. Launched in 2019, it provided Google with a transition from a unidirectional understanding of concepts to a bidirectional one. While previous models could only gain knowledge from words in one direction (unidirectional), it now gains contextual understanding based on words in both directions (bidirectional). The order of the words thus plays an important role in the unidirectional model; in the bidirectional one, information is also sent back.
MUM stands for Multitask Unified Model and is multimodal. That is, the model extracts information from different modalities such as texts, images, videos, etc. But what is much more interesting is that MUM can gather information in different languages, which it translates into the user’s language for an answer. If you want to prepare sushi, for example, a Japanese sushi master will probably have the best tips for you – but in Japanese. Through MUM, the user receives the Japanese preparation tips in his language. This also improves access to information for those languages that are not often considered on the Internet.
We have only touched on a few of the key algorithms that can have an impact on organic search. But there is more to what machine learning is being used for.
In Ads, what drives the systems behind automated bidding strategies and ad automation?
How does News know how to group stories?
How are specific objects identified in images?
How are video recommendations made?
These questions can all be answered with machine learning.