Since the commercialization of the Internet, whether it is the news client side, video website or e-commerce platform... all platforms default to themselves as an excellent breeder, and they push (feed) content to users according to their own ideas.

These breeders are trained professionals, as the jargon goes - website editors set the agenda for users, selecting content according to the tastes of most users.

Then the editor was too busy to use the help of the machine - the simplest machine method is "trending recommendations", such as sorting by clicks or other data.

The biggest problem with the breeder model is not knowing the appetite of the diners, which can lead to two significant consequences: first, the diners are not satisfied, and the personalized needs of users cannot be met; second, the waste of their own resources, a large number of long-tail resources are not exposed for a long time, increasing the sunk cost.

Some people have discovered the benefits of machines. Machines can recommend content based on user characteristics. Just as a skilled cook can provide meals according to the taste of each diner, if the machine is smart enough, it can solve the personalized needs of all users to a certain extent. Isn't this the C2M of the content industry?

To be precise, this is the C2M of information delivery, which communicates with a single user and jumps out of the mass communication/focus communication rut. Is it enough to revolutionize the life of all search engines and portals?

This kind of intelligent content C2M has a profound background of the times. Today, you have stood on the edge of the era, watching AI technology ignite the lead of IOT, and then you will find yourself irresistibly entering the next era of information nuclear explosion: information end point explosion, information scale explosion, information platform explosion...

On the information superhighway, the cars you drive and the roads you travel have all changed the rules, and the knowledge framework based on the breeder model that you are familiar with is facing subversion.

In this era, the breeder model has failed, and smart machines will become the greatest variable.

The first scenario that emerges is where humans produce content and machines distribute it.

The next scenario that emerges is where machines produce content and machines distribute it.

The content industry is facing a C2M revolution, is that okay?

"Of course not, the machine is stupid." If you think so, then unfortunately, you are doomed to not see the sun tomorrow.

"Of course." If you think so, congratulations on falling into the pit.

The true situation may surprise you.

First, the essence of the C2M path is towards personalized communication

As an independent research direction, the recommender system can be traced back to the collaborative filtering algorithm in the early 1990s. In the mid-term, it is represented by traditional machine learning algorithms, such as the implicit semantic model promoted by the Netflix competition, and now it is a more complex deep learning model.

In recent years, deep learning has advanced by leaps and bounds, making machine recommendations the sun of the entire Internet. Driven by new technologies, personalized communication has also become more feasible and is getting closer to single-user communication.

(A) collaborative filtering staggering start

According to the encyclopedia entry, collaborative filtering is the use of client-based preferences to recommend information of interest to you. These users either share similar interests or share common experiences. Then the website combines your feedback (such as ratings) to filter and analyze, and then help others filter information.

Of course, user preferences are not necessarily limited to information of particular interest, and a record of information of particular interest is also quite important. Collaborative filtering has shown excellent results and is beginning to dominate the Internet industry.

At first, collaborative filtering was applied to email filtering.

In 1992, Xerox scientists proposed the Tapestry system. This was the first application of collaborative filtering system design, mainly to solve the problem of information overload at Xerox's research center in Palo Alto. The research center employees received so many emails every day that they couldn't sort them, so the research center developed this experimental mail system to help employees solve this problem.

Then, the idea of collaborative filtering began to be applied to content recommendations.

In 1994, the GroupLens project team in Minnesota, USA, created a news screening system, which could help news listeners filter the news content they were interested in. After reading the content, the listeners were given a rating score, and the system would record the score for future reference. It was assumed that the things that the listeners were interested in before would also be interesting to read in the future. If the listeners did not want to reveal their identity, they could also rate it anonymously. As the oldest content recommendation research team, GroupLens created the movie recommender system MovieLens in 1997, as well as the similar music recommender system Ringo, and the video recommender system Video Recommender.

Then came another milestone - the e-commerce recommender system.

In 1998, Amazon's Linden and his colleagues patented item-to-item-based collaborative filtering, a classic algorithm used by Amazon in its early days that became popular.

Does collaborative filtering count as artificial intelligence? From a technical point of view, it also belongs to the category of AI. But it must be pointed out that the collaborative filtering algorithm is relatively weak, whether it is user-based collaborative filtering or item-based collaborative filtering, the recommendation effect is always just passable.

How can we guide the continuous optimization of the recommender system through a systematic methodology? How can we integrate complex realistic factors into the recommendation results? The city lions were once very big, and there must be brave men under the reward. Later, someone finally discovered a more flexible way of thinking.

(2) Traditional machine learning begins to accelerate

In 2006, Netflix announced the Netflix Prize. Netflix, an established online movie rental website, held the competition to solve the machine learning and data mining problems of movie rating prediction. The organizers paid a lot of money for this, claiming that the prize will be $1 million for individuals or teams who can improve the accuracy of Netflix's recommender system Cinematch by 10%!

Netflix has revealed a lot of huge data on its blog, for example:

• 我们有几十亿的用户评分数据，并且以每天几百万的规模在增长。

• 我们的系统每天产生几百万的播放点击，并且包含很多特征，例如：播放时长、播放时间点和设备类型。

• 我们的用户每天将几百万部视频添加到他们的播放列表。

Obviously, in the face of these massive amounts of data, we can no longer rely on classification standards established by purely manual or small systems to standardize user preferences across the entire platform.

A year after the competition began, Korbell's team won the first stage prize with an 8.43% improvement. They put in more than 2,000 hours of effort and fused 107 algorithms. Two of the most effective algorithms were matrix factorization (often called SVD, singular value decomposition) and restricted Boltzmann machines (RBMs).

As a complement to collaborative filtering, matrix factorization is to decompose a very sparse user rating matrix R into two matrices: the matrix P of User characteristics and the matrix Q of Item characteristics, and build these vectors with known data and use them to predict unknown items. While effectively improving the computational accuracy, the algorithm can also add various modeling elements to integrate more diverse information and make better use of large amounts of data.

However, matrix factorization also has shortcomings. The disadvantage is that matrix factorization, like collaborative filtering algorithms, belongs to the category of supervised learning, which is crude and simple, and is suitable for small systems. The problem facing the network giants is that if you need to build a large recommender system, collaborative filtering and matrix factorization will take a long time. What should I do?

Therefore, some siege lions have turned their attention to unsupervised learning. The essence of clustering algorithms in unsupervised learning is to identify groups of users and recommend the same content to users within that group. When we have enough data, it is best to use clustering as a first step to reduce the selection of relevant neighbors in the collaborative filtering algorithm.

The implicit semantic model uses the cluster analysis method, and one of its major advantages is that it can not only make scoring predictions, but also model text content at the same time, which greatly improves the effect of recommending through content.

The traditional analysis method is not very accurate in the two steps of labeling users and mapping them to the results according to the labels. For example, the age filled in by the user is not necessarily true, or not all teenagers like comics. The core of the implicit semantic model is to go beyond the dimensions of these surface semantic labels, and use machine learning technology to mine deeper potential correlations in user behavior to make recommendations more accurate.

Under the order of the Netflix Prize Million Dollar Martial Arts Competition, there are frequent talents in the world. In 2009, it reached a peak and became the most prestigious event in the field of recommender system. This competition attracted many professionals to devote themselves to the research of recommender system, and also allowed this technology to penetrate from professional circles to the commercial field. It triggered a heated discussion and gradually aroused the coveting of mainstream websites. Content-based recommendations, knowledge-based recommendations, hybrid recommendations, and trust-based network recommendations have embarked on a rapid development channel.

These recommendation engines are different from collaborative filtering. For example, content-based recommendations are made based on the content information of the project, without the need to rely on the user's evaluation opinions of the project, and more need to use machine learning methods to obtain the user's interest data from the description of the characteristics of the content. Content filtering mainly uses natural language processing, artificial intelligence, probability statistics and machine learning technologies for filtering.

Is the million dollars worth it? According to 2016 Netflix subscriber data: 65 million registered members, the total time spent watching videos per day is 100 million hours. Netflix says the system saves $1 billion a year.

(3) deep learning brings "driverless"

In recent years, users' major pain points have emerged. The popularity of smartphones has made the huge amount of information and the small reading screen a pair of difficult contradictions to resolve. Users' reading scenarios are no longer stuck on computer screens, but shift to mobile fragmentation. Search engines fail, manual recommendations are too busy, and machine recommendations are not enough. This transformation is a life-and-death test for large content platforms. If you can meet your needs, you will live, if you don't meet them, you will die.

Faced with this problem, YouTube and Facebook have come up with a new solution: using deep learning to build smart machines. In the past decade, deep learning has made huge leaps and is more advantageous for solving large amounts of data.

If human content recommendation is like driving a car, then the content recommendation brought by deep learning is like driverless cars. In this technology, user data is used to "perceive" user preferences. The recommender system can basically be divided into data layer, trigger layer, fusion filter layer and sorting layer. When the data generated and stored by the data layer enters the candidate layer, it triggers the core recommendation task.

Take YouTube as an example. Its newly disclosed recommender system algorithm consists of two neural networks, one for candidate generation and one for sorting. First, taking the user's browsing history as input, the candidate generation network can significantly reduce the number of recommended videos, selecting the most relevant set of videos from a large library.

The candidate videos generated in this way have the highest relevance to the user, and then further predict the user rating. The goal of this network is only to provide broader personalization through collaborative filtering. The task of the ranking network is to carefully analyze the candidate content and select a small number of the best choices. The specific operation is to use the designed objective function to score each video according to the video description data and user behavior information, and present the highest-scoring video to the user.

In this mode, the machine completely takes over the platform. Under the continuous training of deep learning, the machine will become more and more intelligent, the IQ of dealing with people will gradually increase, and in a sense, it will gradually take on the responsibility of watchdog.

Is the content industry about to be disrupted by C2M?

The world is so big that an ATM at a bank in Corpus Christi, Texas, spat out a note on the 11th that read "save me". The news quickly spread across the Chinese Internet and made headlines on many websites.

Do you need to see the exact same article from N websites?

These redundant messages consume your energy and traffic, just as you can see many instant noodle advertisements on any TV channel, making it difficult to quickly find what you want from the large amount of information.

How to solve the embarrassment of redundant user information?

There have been many unsuccessful technical solutions in the past, personal portals were short-lived, RSS subscriptions were not a climate, and cross-site tracking was not on the table. Only C2M can lead the future.

The C2M model can be applied to the whole network like Jinri Toutiao, or it can be based on a giant platform like Facebook. Its core lies in extracting, sorting and transmitting massive amounts of information to users based on user behavior habits, characteristics and demands, which is the secret to overcoming pain points.

But there are also many voices of doubt. For example, there are opinions that recommendations such as collaborative filtering can easily lead to the formation of information cocoons, inability to recognize reading scenes, poor immediacy, and long time, and Jinri Toutiao is often criticized. It also has to deal with challenges such as difficult-to-capture user interests, privacy, and management of user data.

Support and doubt each end, which is right and which is wrong? Although there are two major opportunities in the future, there are currently three mountains to cross.

1. The reasons for support are as follows:

① Thousands of people have thousands of faces, and the mouth can be adjusted.

The personalized content recommendation mechanism can recommend information to users based on their preferences. Through various algorithms, by analyzing the historical behavior of users, comparing relevant users and related items to guess what users may like, listing candidate sets and verifying, users can get more accurate content, so that the information distribution can be thousands of people and thousands of faces, to achieve accurate connection between content and users, rather than the traditional sense of one-sided delivery.

② Fishing needles in the sea to improve efficiency

Personalized recommendation eliminates the need for users to extract and search in massive information. Users do not need to touch the needle in massive information, which removes some useless information for users to a certain extent, narrows the scope of user information search, and improves user reading efficiency.

③ Do what you like to enhance stickiness

Constantly recommend content suitable for users to increase user stickiness. Personalized recommendation technology makes accurate recommendations of content that users are interested in through algorithms to help users quickly discover content of interest. When you read a piece of content, it will immediately recommend relevant things to you, which can increase user stickiness and improve user experience.

④ Dig the long tail and break the poles

Personalized recommendations can help users mine long-tail content through relevant algorithms to avoid the Matthew effect of polarization. When user A likes unpopular long-tail content, but user B has the same or similar interests and behavioral habits as user A, the system can recommend unpopular content that user A likes to user B, so that unpopular content gets more exposure, helps users discover more long-tail content, and avoids the polarization of content production ecology.

⑤ Two-way communication, deep optimization

Personalized recommendations based on users are the result of in-depth analysis and communication with users, which enhances the user's interactive experience. Traditional manual recommendations are recommendations that cast a net everywhere, without detailed division and screening of users. Machine recommendations are based on user characteristics and habits. Users can get two-way communication and communication, and user behavior can also affect the next recommendation. To a certain extent, feedback is obtained, which enhances the user's interactive experience.

(6) Classification and operation refinement

Personalized recommendation also helps the platform to classify content, which is conducive to the fine management and operation of the platform. Information makes the platform continue to emerge, various forms of content are becoming more and more abundant, and the display area of users' mobile phones is limited. Personalized recommendation allows merchants to better classify content for different customers, which is conducive to fine operation.

2. The main points of doubt are:

① Draw dungeons as prisons, set limits on thinking

Personalized news experiences can easily hold back thoughts. The result of personalized recommendations is recommendations based on users' historical data and historical behavior, based on similar users or similar items, to a certain extent, the content that users are interested in is fixed in a specific closed loop, which filters information for users while also cutting off a lot of information for users. The content of personalized recommendations is collected from your interests and determines your interests. Therefore, if you can't get in touch with "new" things, you can't naturally cultivate new interests, and it is easy to make users more and more narrow-minded.

② The human heart changes, what does the machine understand?

Machine recommendations can't recognize changes in demand brought about by changes in reading scenarios, can't perceive why users need to read, and can't match the complexity of human emotions. For example, at a certain stage, we pay attention to something because everyone is talking about it, but that doesn't mean we are all interested in similar things.

③ Aesthetic offline, good or bad is indistinguishable

The difficulty of personalized recommendations poses a challenge to the quality of recommended content. In the past, it was not so easy for editors to evaluate the quality of an article. Now it is easy for machine recommendations to ignore the quality dimension. Inaccurate machine algorithms will make headlines mixed with content. Machine recommendations may recommend a worthless article highly, or they may bury the truly valuable article. Machine recommendations can only measure the value of your article from external data. There is no way to analyze whether it is valuable from the essence of the content.

④ It takes a long time, and the total is half a beat slower.

Personalized recommendation behavior based on massive data takes a long time and has poor immediacy. For example, news recommendations have timeliness problems and need to be updated continuously. Data analytics such as analyzing users' historical behavior and comparing similar users takes a long time, and it is not easy to form recommendation results in the first time. And collaborative filtering and other methods also have the problem of cold start, that is, when there is no mature historical data at the beginning of user experience, it takes a long time to collect user click log data to generate recommendations.

⑤ Common hotspots, individual convergence

Not all users are equal to each other, but collaborative filtering methods do not take into account individual differences between users. For example, we have observed that entertainment news is constantly recommended to most users, even when users do not click on entertainment stories. The reason is that entertainment news is generally very popular, so it is always enough clicks from a user's "neighbor" to recommend entertainment stories.

3. Where are the future opportunities?

The future opportunities lie in two major driving forces: the industry's commercial drive for long-tail gold mines; and the strong personalized demand from users.

① Changwei Gold Mine

Personalized recommendations can help users discover more high-quality long-tail content and improve the commercial value of the platform. Generally, platform users access only about 10% of the popular content, but many niche and unpopular content is not easy to find in the database. We call it long-tail content.

According to the long tail theory, due to cost and efficiency factors, when the venues and channels for commodity storage and distribution are wide enough, the production cost of commodities drops sharply so that individuals can produce them, and the sales cost of commodities drops sharply, almost any product that seemed to be in extremely low demand before will be bought as long as it is sold. Personalized recommendation can spread the long tail content that niche likes through user-based recommendation technology in collaborative filtering, fully tap the long tail content, and generate long tail gold mines.

② The rigid demand of the times

The era we live in has changed. After 20 years of development, the Internet has become the mobile Internet, and now it is about to integrate AI into the IoT era. End points and information are expanding rapidly in a nuclear explosion, and it will become increasingly difficult for users to find the information they need in the massive amount of data. In this case, traditional search engines are no longer capable. The most representative of the earlier Yahoo and Google, which are classified directories, have entered a dead end. It is extremely inefficient to learn about knowledge in an unfamiliar field through search engines!

To meet the rigid needs of the times, hope lies in personalized recommendations. Machines need to understand users as much as possible, and based on user data, actively recommend information that interests and needs users. At present, although a little achievement has been made in the past 20 years, Tang Seng has only taken the first step in learning scriptures, and there is still a long way to go.

4. Three mountains to cross now

Personalized recommendation faces many problems in the development process, such as the difficulty of predicting user interests, the privacy of user-related data, and the difficulty of data processing, all of which bring great threats and challenges to personalized recommendation.

The first mountain, precisely.

The user's interest is susceptible to multiple factors and constantly changes, which is an inevitable challenge for personalized recommendation. The basic part of the personalized recommender system is user interest modeling, and the quality of user interest modeling directly determines the quality of personalized recommendation. However, user interests are affected by multiple factors such as social, scene, and environment at any time. The continuous change of user interests makes it difficult to predict users' future tendencies based on past data, and also affects the accuracy of recommendation results.

The second mountain, privacy.

For personalized recommendations based on user data, how to protect user privacy is a big problem. Traditional content recommender systems mine users' page access records to find out their access habits, and then filter information according to user requests on the server side, trying to provide users with information recommendation services and spam filtering services. But how to provide users with more accurate content recommendation services while protecting user privacy is a big challenge.

The third mountain, values.

In addition to the three mountains, there is another problem that deserves attention. The current machine recommendation is equivalent to "no three views" and "no aesthetics". Operating in the Chinese circle will definitely encounter considerable challenges due to well-known reasons.

Traffic fraud and cheating are relatively obvious examples. For example, some netizens told the author: I often see some videos online with tens of thousands or hundreds of thousands of learners, and the numbers are so large that we doubt our lives. After testing, the number of people will increase by three once the page is refreshed, and the number of new courses will increase by dozens, which is instantly clear. Test some video live broadcasts in the middle of the night and shoot against the wall. From the start of the live broadcast for ten minutes, the number of live fans can still rise. When a real fan enters, the number of people rises again. Cheating is cool for a while, but my heart is not at ease.

There have been companies that have cast some very vertical large-scale advertisements on the client side of intelligent recommendations, some of which are very effective, and some of which are too obvious to be fraud---the traffic brought over when the reading volume instantly exceeds 10,000 is not as good as reading the number that breaks thousands. So much, the data is serious, it depends on whether the person using it is serious or not.

In the future, how to continue to innovate in technology and management of personalized recommendation, whether the participation of artificial intelligence factors can improve many existing problems, and generate better recommendation results for users will become an important issue.

III. The technological routes that the giants are pioneering

In fact, no matter how big the support or doubts are, personalized recommendations have attracted countless giants to submit.

At present, in the market, new and old technologies still occupy each other's territory. New deep learning technologies are rising rapidly and aggressively; old-school technologies are also constantly being optimized to prevent accidents. The battle between new and old technologies is a hot spot at the moment, and it is also the two major routes that determine future development.

(1) Old-school technology believes that traditional recommendation technologies can improve themselves

1.Google news routine, continuous optimization

Google News is an online information portal that aggregates the coverage of thousands of sources (after grouping similar news) and presents it to logged-in users in a personalized way. Due to the large number of articles and users, and the given response time requirements, a purely memory-based approach is not applicable and a scalable algorithm is required, so Google News uses a combination of model-based and memory-based techniques.

Google News remains the foundation of collaborative filtering. It uses a combination of model-based and memory-based techniques for personalized recommendations. According to the recommender system, the model-based part relies on two clustering techniques:

① Probabilistic Latent Semantic Indexing (PLSI): The "second generation" probabilistic technique of collaborative filtering, in order to identify clusters of users with similar ideas and related items, introduces hidden variables, corresponding to a finite set of states for each user-item pair, which can adapt to situations where users may be interested in multiple topics at the same time.

② MinHash: Putting two users into the same cluster (hash bucket) based on the intersection of the items they have viewed. To make this hashing process scalable, a special method is used to find nearest neighbors, and Google's own MapReduce technology is used to distribute computing tasks among several clusters.

Memory-based methods mainly analyze "concomitant views". "concomitant views" refers to an article that has been viewed by the same user over a pre-defined period of time. Forecasting requires traversing the recent historical data of active users and retrieving neighboring articles from memory. At runtime, the comprehensive recommendation score of candidate items in the pre-set set set is calculated as a linear combination of the scores obtained by these three methods (MinHash, PLSI, and concomitant browsing), and then the recommended results are output according to the height of the calculated values.

2. LinkedIn developed a system for four scenarios

Linkedin mainly realizes personalized recommendation through Browsemap, a collaborative filtering recommendation platform developed by Linkedin. Browsemap is a generalization platform developed by Linkedin to implement the item collaborative filtering recommendation algorithm. The platform can support the recommendation of all entities in Linkedin, including job seekers, job postings, enterprises, social groups (such as schools, etc.), search terms, etc. To implement a new entity collaborative filtering recommendation through this platform, the developer only needs to do simple work such as accessing relevant behavior logs, writing Browsemap DSL configuration files, and adjusting relevant expiration parameters.

The paper points out that the Browsemap platform is most commonly used on LinkedIn in four recommendation scenarios: recommending companies to job seekers, recommending similar companies, recommending similar resumes, and recommending search terms.

① Recommend companies to job seekers: implement collaborative filtering based on items through Browsemap, calculate the similarity value between users and potential companies, and obtain relevant company characteristics; analyze relevant company characteristics and user/company content characteristics (including user location, work experience; enterprise products, related descriptions) together to obtain the final preference score.

② Similar company recommendation: There are two differences between recommending companies to job seekers: first, the similarity of content features becomes the similarity between company portraits; second, browsemap is built based on a variety of user behaviors.

③ Similar resume (user) recommendation: This part of recommendation is realized through the company details page browsing behavior and user portrait characteristics. At the same time, the attributes of similar resumes are used to make up for the missing attributes of resumes to get the user's virtual resume.

④ Search term recommendation provides four ways of association: first, collaborative filtering: time and space factors are added when calculating the correlation between search terms; second, the click through rate of search results based on recommended search terms; third, based on the overlap between search terms; fourth, based on the click through rate of recommended search terms. But the experimental results show that the results of collaborative filtering are the best, and even better than the results of combining these four methods.

3. Three stages of Jinri Toutiao

As a popular personalized recommended product in China, Jinri Toutiao technology has undergone three stages:

In the early stage, non-personalized recommendations were the main ones, focusing on hot-article recommendations and new-article recommendations. At this stage, the granularity of the description of users and news was also relatively coarse, and recommendation algorithms were not widely used.

In the mid-stage, personalized recommendation algorithms are the main ones, mainly based on collaborative filtering and content recommendation. The technical idea of collaborative filtering is no different from that introduced earlier. Based on content recommendation, it is to first characterize the news, and then use the user's positive feedback (such as clicks, reading time, sharing, favorites, comments, etc.) and negative feedback (such as disinterest, etc.) to establish the connection between the user and the news tag, so as to conduct statistical modeling.

At the current stage, large-scale real-time machine learning algorithms are the main ones, with hundreds of billions of features used, which can update the model at the minute level. The architecture is divided into two layers: the retrieval layer, which has multiple retrieval branches to pull out news candidates that users are interested in; and the scoring layer, which uses real-time learning to model and score based on user characteristics, news characteristics, and environmental characteristics. It is worth mentioning that the actual sorting is not completely sorted according to the model, and some specific business logic will be integrated together for final sorting and sent to the user.

Why did Toutiao succeed? According to the analysis of the article, many people will say that Toutiao's personalized recommendation technology is doing well, but in fact it is not. The reason is that Jinri Toutiao's personalized recommendation has also undergone a complex evolution process: from manual recommendation to machine recommendation to final iteration of algorithms and technologies, repeated verification, and increasingly perfect.

The new school of technology believes that deep learning is the wise choice

New technology mainly refers to the use of deep learning personalized recommender system.

Deep learning is a method of machine learning based on the representation of data. Observations (such as an image) can be represented in a variety of ways, such as vectors of intensity values per pixel, or more abstractly as a series of edges, regions of a specific shape, etc. It is easier to learn tasks from examples (e.g. facial recognition or facial expression recognition) using certain specific representations. The benefit of deep learning is to replace manual feature acquisition with efficient algorithms for unsupervised or semi-supervised feature learning and hierarchical feature extraction.

When conventional recommendation algorithms have been unable to analyze and process large amounts of data in a timely manner and accurately make recommendations for independent users, companies with the corresponding technical level have begun to use deep learning to solve the pain points of mass content analysis and recommendation. We take YouTube and Facebook, which have introduced deep learning earlier, as examples for analysis.

1.YouTube neural networks

YouTube's recommender system is one of the largest and most complex recommender systems in the world. YouTube has more than one billion users worldwide, and the length of videos uploaded every second is measured in hours. The growing stock of video "corpus" requires a recommender system to continuously recommend videos of interest to users in a timely and accurate manner.

Compared to other commercial recommender systems, YouTube recommender systems face three main challenges:

① Scale. Most of the existing viable recommendation algorithms are unable to handle YouTube-level massive videos.

② Freshness. The YouTube video "corpus" library is not only huge in reserves, but also a steady stream of new videos uploaded every moment. The recommender system should analyze and model the content uploaded by users in a timely manner, and at the same time take into account the balance between existing videos and newly uploaded videos.

③ Noise. Due to the sparsity of user behavior and unobservable factors, the history of users is inherently unpredictable.

In order to solve these problems, the YouTube recommender system shifted its research focus to deep learning, and the TensorFlow (the second generation artificial intelligence learning system developed by Google) system developed using Google Brain brought flexibility to the recommender system in development and testing.

The YouTube recommender system is mainly composed of two deep neural networks: the first neural network is used to generate a list of candidate videos; the second neural network is used to score and rank the input video list in order to recommend the top videos to users.

Candidate video generation relies on collaborative filtering algorithms to generate a broad list of personalized recommendations for users. Ranking neural networks are based on the first list of candidate generation networks to provide finer differentiation and always achieve a high recommendation hit rate. By defining an objective function to provide a series of characteristics describing the video and the user, the ranking network scores each video according to the objective function. The set of videos with the highest score is recommended to the user.

It is YouTube's massive videos that generate the need for deep learning, effectively compensating for the time-consuming problem of collaborative filtering to process data.

2. A big step for Facebook

Facebook has been using its Newsfeed feature to personalize recommendations for nearly a decade. In September 2006, NewsFeed (information flow) was launched, along with MiniFeed (personal news). NewsFeed is a system that automatically integrates the generated content and information flow, and it decides on its own what news, news, and events we read. Its coverage, the accuracy of its information push, and its influence far exceed our imagination. It can be said that NewsFeed is a big step forward for Facebook in artificial intelligence.

How does Facebook use deep learning to evaluate content and users?

First, when it comes to viewing text, Facebook uses "natural language processing" technology to scan the "status" and "logs" of each person's post in order to "really understand the semantics of the text" and not only to rate them. In the process of scanning the logs, the system will automatically identify "excessive headline party" or "excessive commercialization" content, and such content is increasingly rare in NewFeed.

Second, in terms of content translation, when dealing with non-English languages, Facebook engineers have developed a deep learning platform that analyzes and translates text written in more than 100 languages every day. For example, when a friend publishes a message in German, NewsFeed will be reflected in English to an American friend, creating a digital virtual environment that can overcome language barriers and enable everyone to connect.

Third, in terms of recognizing objects, Facebook is also using deep learning technology to identify objects in photos and videos. Not only that, but it can also further explore who is likely to be interested in these photos, or which users these photos are associated with, so as to recommend them to target users.

The dilemma of deep learning

Can deep learning be invincible all over the world?

At least for now, deep learning is only effective in "shallow" intelligent problems such as Speech and Image, while it is a little less effective for language understanding and reasoning. Maybe future deep neural networks can solve this problem more "intelligently", but it is still almost hot.

The research and application of deep learning in the field of recommender system is still in its early stage. Even though deep learning is considered to be able to solve the problems of cold start and slow data processing of collaborative filtering, it also has its own secrets under the scenery.

First, the cost is too high. Data is critical to the further development and application of deep learning. However, over-reliance on labeled big data is precisely one of the limitations of deep learning. Data collection is expensive, and the cost of labeling has begun to rise, which makes deep learning too expensive. And for many small companies with small volumes and less data, even if they have the ability to use deep learning to improve personalized recommendation results, they face the embarrassment of no data support.

Second, are there any solutions to reduce costs? Yes, but it is difficult to achieve. Deep learning is divided into supervised learning and unsupervised learning, and the cost of acquiring a large amount of unsupervised data is negligible. At present, supervised learning is generally used, but essentially most recommendation models based on supervised learning are difficult to completely avoid existing problems and improve recommendation quality. Unsupervised learning has a lower cost than supervised learning due to the need to label data, but the current deep learning ability to learn from unsupervised data is seriously insufficient, so the application of deep learning in recommender systems is still in its early stages.

The two forces of the new and old schools are fighting against each other, promoting each other but blending with each other. Traditional recommendation technology is constantly improving under the impact of deep learning, and deep learning is constantly innovating with a strong drive to catch up with traditional recommendation technology, but it is also facing a development dilemma. But it is in the process of self-development and innovation on multiple platforms that the boundaries between the new and the old schools have become increasingly blurred and increasingly integrated. Even companies that insist on perfecting traditional recommendation technology have begun to slowly enter the field of deep learning. The more mature new schools of deep learning have not completely abandoned the old-school technology. So, what school is king in the future?

IV. Who will kill the deer in the future?

Content C2M is essentially an insight and prediction of the human heart. The battle between technology and the human heart does not happen overnight. The fundamental feature of human thought is "consciousness", that is, the ability of individuals to understand their own and others' mental states, including emotional intentions, expectations, thoughts, and beliefs, and use this information to predict and explain the behavior of others.

However, there is a serious problem in the current field of artificial intelligence: people misunderstand how deep learning models work and overestimate the capabilities of network models.

With deep learning, we can train a model that can generate text descriptions based on the content of images. This process is seen as the machine "understands" the image and the text it generates. When there is a slight change in an image that causes the model to start generating rather ridiculous captions, the result is very surprising - the model fails. The machine can find a cat, but the machine still cannot recognize all the information related to the cat.

Looking back at history, it is not difficult to find that the goal of technology has been not so much to replace humans with machines, but to create smart machines to improve efficiency. The development of collaborative filtering technology is a clear example.

In recent years, the enthusiasm of Internet Tech Giants to create "smart machines" has been extremely high, which is also due to the efficiency. According to estimates from Microsoft Research, about 30% of page views on Amazon's website come from the recommender system; Netflix's chief product officer claims that more than 80% of movie viewing comes from the recommender system, and claims that the Netflix recommender system is worth up to one billion dollars per year; according to Alibaba's figures, in 2013, the gross merchandise volume directly guided by recommendations was 5.68 billion yuan. Jinri Toutiao has built the company's core business on the recommendation engine, and is one of the most important companies in recommendation technology today....

In the development of content C2M, although deep learning has many shortcomings, it is a high probability that deep learning will dominate the future. We see that the old and new schools representing traditional recommendation technology and deep learning are promoting and merging with each other. In the top 20 platforms of global traffic, although many companies still use collaborative filtering technology, such as Google News, LinkedIn, etc., some of them have also prepared or even adopted deep learning and other technologies to improve their own shortcomings. And pioneers such as YouTube and Facebook have begun to enjoy the dividends of deep learning.

From the breeder model to smart machines, the C2M of the content industry has taken hold, and the day of disruption is not far away.

We can believe that although there are still some constraints in deep learning, with the strong development of AI technology and industry, the technical bottleneck will eventually be broken through.

It is necessary to be vigilant that after C2M crosses the two mountains of accuracy and privacy, humans have mastered new powers through AI, and the desires and ambitions of the masters should also be controlled to a certain extent, especially the issue of values, which will become increasingly important.