The information explosion has brought nectar to hungry netizens, but the proliferation of massive information has also caused netizens to suffer. Today, a number of Internet companies have begun to attempt to present genuine "appetite" messages to netizens through editorial selection and smart recommendations. As Chris Anderson, author of The Long Tail Theory, puts it, we are leaving the age of information and entering the age of recommendation.
In mid-September, ACM RecSys 2012, the top international conference in the field of recommended systems, was held in Dublin, Ireland. Prior to this, RecSys was successfully held five times in Minneapolis, Lausanne, New York, Barcelona and Chicago. .
RecSys 2012 attracted the participation of top scholars from universities in the world and well-known companies in the Internet industry, such as LinkedIn, Yahoo!, Microsoft, Facebook, etc. The topics covered include recommendation algorithm, social recommendation, user modeling, machine learning and human-machine Interactive and other frontier areas. At this top international conference, Chinese Internet companies also appeared. Baidu, from mainland China, was the only domestic company that participated in the conference and was the first domestic company to participate in the conference as the author of the paper.
At the conference, Baidu released the paper: Enlister: Baidu's Recommender System For The Biggest Chinese Q&A Website (Baidu Recommended System Service on China's Largest Q&A Platform). This paper was unanimously approved by foreign counterparts and was finally accepted by the General Assembly. It is reported that RecSys 2012 received a total of 24 long papers, admission rate of 20.2%; received 21 short papers, acceptance rate of 31.8%.
This topic of Baidu is actually an incidental result of Baidu's product R&D. It was completed by Baidu's front-line engineers, mainly from the Department of Recommendations and Personalization and Baidu's knowledge of product R&D. The main achievement of the paper is based on Baidu's problem recommendation system. Now it is providing problem recommendation services for 200 million users Baidu knows every day. At the same time, the follow-up of these technologies will also be used as a common basic technology and applied to other products in the recommendation and personalization department, such as Baidu's new home page navigation, Baidu Post Bar personalized posts, Baidu video personalized videos and other products.
In the R&D process, in the face of worldwide recommended technical problems, the personnel of the two departments reached consensus at the beginning of R&D, and prepared to use a series of innovative strategies to solve problems. First of all, they conducted user behaviors and analyzed the interests, states, and behaviors at multiple levels after the privacy process was performed. A user model was created for each user to give a recommendation result that belongs to an individual, achieving “one person, one worldâ€. User experience to improve the user model.
Second, they innovatively translate the task of ranking in recommendations into click-rate estimation, using machine learning frameworks to solve the industry-recognized problems and build a machine learning ranking model.
In addition, they also use a streaming computing framework that reduces the time from when a question is presented to the user who is interested in the question to 10 minutes, ensuring that the right question can be quickly presented to the right user to quickly resolve it.
It turns out that the general user model, machine learning sequencing, and streaming calculations they use have achieved very good results in the project. After the project went online, Baidu knew that the number of responses increased from 84,000 to 102,000, an increase of 21.4%; the conversion rate increased from 0.148% to 0.179%, an increase of 21.0%.
One of their achievements is to prove the importance of the application of machine learning strategies in the ranking of recommendations, follow-up will continue to optimize and promote to more products; second, streaming computing architecture can give users a good experience, making it Will be used as the core structure of the follow-up recommended products and promotion; third is to prove Baidu recommended technology R & D at the leading level of the recommended industry, Baidu follow-up development plan has a certain guiding significance.
R & D is not easy. According to Baidu engineers, during the R&D process, they encountered difficulties in the selection of samples and features of machine learning sequencing problems.
"According to the more popular CTR prediction method, a negative sample will select content that users have not clicked on. This approach requires very high sample sizes and feature sizes. The industry's usual advertising CTR estimation system usually has ten Billions or billions of samples, billion-level features, often require hundreds of machines to do model training work. Such machine budgets are not affordable to the product line."
According to a Baidu engineer who is involved in R&D, this problem has plagued them for some time. Later, in the group discussion and research of the machine learning group of the Natural Language Processing Department and Baidu, they gradually came up with a small sample selection and feature extraction method. They extracted mega-samples and 100-level features so that they can complete model training with a few machines without significantly reducing the accuracy of the model. "Because of the rich machine of the machine learning team Learning experience and product line colleagues' insights on the product, and ultimately everyone together beautifully solved this problem."
In recent years, with the sudden emergence of social networks represented by Facebook and Twitter, user-contributed content, and socialized approaches to dissemination, the information volume has exploded geometrically, and the user-centric “recommendation†era of information has arrived, regardless of the availability of the Internet. There will be great changes. Those Internet companies with technical strength and forward-looking will surely lead the next wave of the Internet.
In mid-September, ACM RecSys 2012, the top international conference in the field of recommended systems, was held in Dublin, Ireland. Prior to this, RecSys was successfully held five times in Minneapolis, Lausanne, New York, Barcelona and Chicago. .
RecSys 2012 attracted the participation of top scholars from universities in the world and well-known companies in the Internet industry, such as LinkedIn, Yahoo!, Microsoft, Facebook, etc. The topics covered include recommendation algorithm, social recommendation, user modeling, machine learning and human-machine Interactive and other frontier areas. At this top international conference, Chinese Internet companies also appeared. Baidu, from mainland China, was the only domestic company that participated in the conference and was the first domestic company to participate in the conference as the author of the paper.
At the conference, Baidu released the paper: Enlister: Baidu's Recommender System For The Biggest Chinese Q&A Website (Baidu Recommended System Service on China's Largest Q&A Platform). This paper was unanimously approved by foreign counterparts and was finally accepted by the General Assembly. It is reported that RecSys 2012 received a total of 24 long papers, admission rate of 20.2%; received 21 short papers, acceptance rate of 31.8%.
This topic of Baidu is actually an incidental result of Baidu's product R&D. It was completed by Baidu's front-line engineers, mainly from the Department of Recommendations and Personalization and Baidu's knowledge of product R&D. The main achievement of the paper is based on Baidu's problem recommendation system. Now it is providing problem recommendation services for 200 million users Baidu knows every day. At the same time, the follow-up of these technologies will also be used as a common basic technology and applied to other products in the recommendation and personalization department, such as Baidu's new home page navigation, Baidu Post Bar personalized posts, Baidu video personalized videos and other products.
In the R&D process, in the face of worldwide recommended technical problems, the personnel of the two departments reached consensus at the beginning of R&D, and prepared to use a series of innovative strategies to solve problems. First of all, they conducted user behaviors and analyzed the interests, states, and behaviors at multiple levels after the privacy process was performed. A user model was created for each user to give a recommendation result that belongs to an individual, achieving “one person, one worldâ€. User experience to improve the user model.
Second, they innovatively translate the task of ranking in recommendations into click-rate estimation, using machine learning frameworks to solve the industry-recognized problems and build a machine learning ranking model.
In addition, they also use a streaming computing framework that reduces the time from when a question is presented to the user who is interested in the question to 10 minutes, ensuring that the right question can be quickly presented to the right user to quickly resolve it.
It turns out that the general user model, machine learning sequencing, and streaming calculations they use have achieved very good results in the project. After the project went online, Baidu knew that the number of responses increased from 84,000 to 102,000, an increase of 21.4%; the conversion rate increased from 0.148% to 0.179%, an increase of 21.0%.
One of their achievements is to prove the importance of the application of machine learning strategies in the ranking of recommendations, follow-up will continue to optimize and promote to more products; second, streaming computing architecture can give users a good experience, making it Will be used as the core structure of the follow-up recommended products and promotion; third is to prove Baidu recommended technology R & D at the leading level of the recommended industry, Baidu follow-up development plan has a certain guiding significance.
R & D is not easy. According to Baidu engineers, during the R&D process, they encountered difficulties in the selection of samples and features of machine learning sequencing problems.
"According to the more popular CTR prediction method, a negative sample will select content that users have not clicked on. This approach requires very high sample sizes and feature sizes. The industry's usual advertising CTR estimation system usually has ten Billions or billions of samples, billion-level features, often require hundreds of machines to do model training work. Such machine budgets are not affordable to the product line."
According to a Baidu engineer who is involved in R&D, this problem has plagued them for some time. Later, in the group discussion and research of the machine learning group of the Natural Language Processing Department and Baidu, they gradually came up with a small sample selection and feature extraction method. They extracted mega-samples and 100-level features so that they can complete model training with a few machines without significantly reducing the accuracy of the model. "Because of the rich machine of the machine learning team Learning experience and product line colleagues' insights on the product, and ultimately everyone together beautifully solved this problem."
In recent years, with the sudden emergence of social networks represented by Facebook and Twitter, user-contributed content, and socialized approaches to dissemination, the information volume has exploded geometrically, and the user-centric “recommendation†era of information has arrived, regardless of the availability of the Internet. There will be great changes. Those Internet companies with technical strength and forward-looking will surely lead the next wave of the Internet.
Portable power system ,Portable power storage,Portable solar system, Outdoor power station, Portable power station
SHENZHEN CHONDEKUAI TECHNOLOGY CO.LTD , https://www.szsiheyi.com