Reporting by XinZhiyuan
Editor: Yuan Xie La Yan
From automatically identifying street signs to automatically removing scam information, the Google Maps project is now increasingly relying on machine learning tools.
The future of curbing bad network dynamics is in the hands of machine learning.
In the hands of search giant Google, machine learning tools got a real workout in 2021 by regulating violations on Google Maps.
The Google Maps team said: "Our team is committed to making the content posted by real users on maps as reliable as possible and based on real-world experience. This work helps protect businesses from bullying and scams and ensures that reviews are helpful to users. This content policy is designed to prevent misleading, false and abusive comments on our platforms."
Google used machine learning to remove hundreds of millions of fraudulent edits and nearly 200 million offending images in a year
In a recent official blog post on how to keep maps information reliable, Google said it combined machine learning and human operators to block more than 100 million attempts to fraudulently edit Google merchant page profiles on maps apps in 2021.
This is thanks to the continuous improvement of machine learning models, which have improved Google's efficiency in identifying malicious screen swiping and suspicious behavior of robots.
In the same way, Google also removed more than 7 million fake merchant pages on maps apps, of which 630,000 were based on user reports.
In addition, Google said it blocked 12 million attempts by fraudsters to impersonate other companies and 8 million scams that claimed permission to other companies' merchant pages on the Maps app.
Machine learning tools also helped the Google Maps team remove nearly 200 million "low-definition or rule-breaking" photos and videos.
On top of that, as a result of these breaches, Google deleted 1 million user accounts that were used for fraud.
The post, about maintaining the reliability of Google Maps, also mentions how Google has taken steps to protect merchants from fake comments on the Maps app after the COVID-19 pandemic eases in 2021 and business activities reopen.
More than 95 million map page comments that violated Google's relevant rules were removed, of which 60,000 were related to COVID-19-related issues to varying degrees.
Google Maps receives about 2 million messages from user contributions every day. This information includes up-to-date merchant hours, contact details and customer reviews.
Throughout 2021, the information provided by users helped Google Maps improve the merchant page service. 30% more businesses appear on Google Maps in 2021 than in 2020.
Google Maps, a hugely popular app, will undoubtedly become a magnet for misconduct. Therefore, Google has made great efforts to maintain a clean and clean application environment, which is really gratifying to users.
Extra training makes the audited AI smarter
However, although Google Maps has achieved good results using machine learning, manual participation in this process is rather more reassuring to users, after all, AI intelligence has its limits.
For example, the garlic bread in a pizzeria tastes so bad that the user gives a joke comment of "bursting", which is deleted by the AI as a threat of violence. Artificial intelligence really has no sense of humor.
On the Google Maps team, Ian Leader, a product manager responsible for user posting, described how machine learning can be combined with human review in a post on the official blog:
Machine learning plays an important role in the audit process. A large number of audit systems that apply machine learning are Google's "first line of defense because they are good at pattern recognition."
The system checks each comment for possible violations of the rules. For example, the system determines the language pattern of comments, the history of users or business accounts, and whether there is any unusual activity related to a particular place of business (such as a sudden increase in the number of one- or five-star reviews).
The leader said the machine learning model eliminated "the vast majority of fake and fraudulent content" before any user saw it. This process may only take a few seconds, and if the model doesn't find any problems in the comments, it will pass through it as quickly as possible for other users to read.
However, these systems are not perfect. Leader example: "Sometimes the word "gay" is used as a derogatory term, which is not allowed in the comments bar of Google apps.
But if the machine learning model is trained using only datasets of hate speech, we may mistakenly remove ads from gay business owners or comments about safe spaces for sexual minorities."
As a result, the Google Maps team often performs quality tests on AI and conducts additional training to teach the system a variety of contextual contexts for specific words and phrases, in order to refine machine learning models, reduce bias values, and guarantee a balance between removing harmful content and protecting useful comments.
Google Maps also has a group of comments that people manually evaluate businesses and user tags. In some cases, in addition to deleting offending comments, Google suspends user accounts and files lawsuits.
Machine learning reads street signs for Google Maps
Google Maps' business relies on machine learning algorithms in part that goes far beyond auditing. It can be said that without machine learning, most of Google Maps' business cannot be carried out now.
Images and censored data are static and can't keep up with the ever-changing world around users. Machine learning algorithms can analyze instant images and data and identify changes in new data.
This allows the map app to update only based on the latest changes in the real environment. This improves the speed at which map content is generated and ensures that the generation process is automated while maintaining accuracy.
The Google Maps project uses deep neural networks to automate the process of reading image information. The algorithm is publicly available on GitHub through TensorFlow, Google's own open source machine learning software library.
The Google Maps project has long used machine learning to identify car license plates, and is now using the same technology to obtain information from street signs.
Google aims to use the technology to improve the location data of about one-third of the world's addresses in the Maps app. When tested on several street signs in France that are more difficult to identify, the latest machine learning algorithm achieved an accuracy rate of 84.2%, which is better than before.
The Google Maps project now applies machine learning tools that improve the software suite that used to read street numbers and street names. The new algorithm removes any irrelevant text from the image and replaces previously unlearnable abbreviations with full names.
The algorithm identifies building outlines for Google Maps
Buildings are landmarks and are a key part of how users know where they are when they view a map.
Old algorithms in the past used to generate irregularly shaped spots when trying to guess whether part of a picture was a building. When these pictures are superimposed on the map, they don't look like real buildings at all.
To solve this problem, Google's data operations team continued to manually label common building outlines, and then used the annotated data to train machine learning algorithms to let the AI learn which images corresponded to the edges and shapes of the building.
Relying on Google's technology, money, and manpower, the process now allows AI to map as many buildings as it has in the past decade.
Now, when new buildings or shops appear in an area, Google's machine learning algorithms recognize changes and update existing maps instead of redrawing the entire area. This saves a lot of time and effort for both the supply and demand sides of the service.
The algorithm updates real-time transit data for Google Maps
Google is conceiving new ways to give users real-time visibility into the status of their bus rides.
Google Maps will enable predictive power through machine learning, informing users in advance of whether the public service they are going to take will encounter obstacle delays. Google Maps' algorithm now has real-time access to tracking data, and in the test run it can predict delays in hundreds of cities around the world.
In short, Google's machine learning model uses standard traffic data as the baseline truth value, and then adjusts for the particularity of bus travel and route.
The Google Maps team extracted training set data from the bus's location sequence, which came from real-time feedback from the bus agency, and then aligned them with the speed of the bus in the trip to produce a highly confident training dataset.
In today's ever-changing world, the most up-to-date information provided by Google Maps is invaluable. Without machine learning, Google Maps is also unsustainable.
Resources:
https://www.androidpolice.com/google-maps-machine-learning-block-100-million-abusive-edits/
https://blog.google/products/maps/how-we-kept-maps-reliable-2021/
https://www.engadget.com/google-maps-review-bombing-machine-learning-153740932.html
https://blog.google/products/maps/how-google-maps-reviews-work/
https://www.springboard.com/blog/data-science/machine-learning-google-maps/