Shulou(Shulou.com)06/03 Report--

First, why do we need a recommendation system?

1. The core value of intelligent recommendation

The essence of recommendation is to let people find what they need in the vast amount of information, which is the product of the era of information explosion. Since the reform and opening up, the material and cultural life has been greatly rich, and the threshold of content and commodity production has been constantly lowing. in the face of information overload, the acceptance speed of users with limited energy is in contradiction with the production speed of content and goods. In this case, the core value of the popularization and application of intelligent recommendation system is to solve the problem of information overload.

2. What kind of enterprises need intelligent recommendation?

Intelligent recommendation to enhance product intelligence and user experience. Take the content distribution industry as an example, the editors of the portal era distribute the content within the channel manually. Take the sports channel as an example, the content granularity that editors focus on may be NBA, but editors are not satisfied with a finer-grained user group such as James.

Intelligent recommendation reduces operating costs and improves operational efficiency. As the flow dividend disappears, enterprises pay attention to reducing costs and increasing efficiency. Traditional portals may require hundreds of editors to distribute content, but still can not meet the more fine-grained content needs of users. With the existence of intelligent recommendation, several operators can be responsible for the distribution of a client's content, and some recommendation interventions can be made on a daily basis.

Intelligent recommendation to improve core business indicators. Jinri Toutiao has built intelligent recommendation distribution content and achieved the rise, which proves that intelligent recommendation can help business growth. Jinri Toutiao previously announced that its users use it for 75 minutes a day, which is much longer than that of portals.

Proven value and industry-wide spread. The success of Jinri Toutiao verifies the value of intelligent recommendation, which is no longer exclusive to Jinri Toutiao and has been expanded into a mobile Internet infrastructure.

What kind of recommendation system do we need?

1. Commonly used recommendation types and scenarios

The commonly used recommendation types include personalized recommendation, related recommendation and popular recommendation.

Personalized recommendation: Feed stream of content media, e-commerce "guess you like", video APP rotation, product details page "guess you like" …... These scenarios are personalized recommendations, through thousands of people, so that users can directly see the content or products of interest.

Figure 1 Personalized recommendation

Related recommendations: according to the similarity of calculated content or goods, some similar content or products are recommended to meet the needs of users for extended reading or extended purchase or repurchase.

Figure 2 related recommendations

Hot recommendations: Weibo's topic list, Zhihu Hot list, etc., are through clicks, collections, comments and other user behavior to calculate the hot and ranking, provide users with popular content to meet the herd mentality.

Figure 3 popular recommendations

2. The enterprise should build itself, and OR should choose a third-party manufacturer.

(1) suitable for self-built enterprises and self-built risks

It is suitable for self-construction for companies that are in line with large companies, profound technology accumulation and strategic depth. Compared with small and medium-sized companies, large companies have sufficient resources, time and manpower reserves to meet their own needs. Large companies have a deep accumulation of technology, and have enough cash cow business, there is no urgent need for growth, they can build on their own. However, self-construction will face higher recruitment cost, time risk and effect risk.

(2) citing third-party manufacturers and risks

Compared with self-building, the introduction of third party does have the advantages of low investment and quick results, but it will face the risk of the precipitation of enterprise intelligent recommendation ability and whether the system can be integrated with the business.

After comprehensive measurement, we think that the best practice for the intelligent recommendation construction of most enterprises is that professional companies do professional things and choose the most professional manufacturers that best meet their needs, rather than the cheapest ones, so as to focus on their core business.

3. How to achieve excellent business goals?

First, data governance intensive cultivation: data burying point and verification. If there is something wrong with embedding and checking itself, no matter how powerful the algorithm is, it won't help. Shenze intelligent recommendation attaches great importance to data embedding and verification. Each customer will have professional analysts to analyze customer recommendation requirements, determine the burial scheme, and work with customer technicians to ensure data quality from the source. In order to achieve better than expected recommendation results.

Second, deep learning is applied to recommendation recall and sorting. Collaborative filtering has existed for more than 20 years, Shenze intelligent recommendation mainly uses deep learning, compared with collaborative filtering, deep learning recall model has more advantages, as follows.

(1) more comprehensive behavior expression. The combination of click, collection, search and other behaviors in the model can more comprehensively express the user's behavior preference, which can not be done in collaborative filtering.

(2) Portrait features can be added, and user portrait related features such as gender and region can be added. If you have some additional tags or information that occurs, the model is compatible and can mix all the information into the same model, which cannot be introduced in the collaborative filtering model at all.

(3) consider the order of users' behavior. For example, the usual behavior order of users is to buy a mobile phone first, and then to buy a mobile phone case; after buying a car, they may buy a car cushion. If the order of recommendation is: the user bought a mobile phone case and was recommended a mobile phone, this logic is obviously wrong. Collaborative filtering does not have the learning ability of sequential relations, it regards all behaviors as a completely equal relationship, while deep learning can learn sequence, which is very cutting-edge in the expression and tuning of the whole model.

(4) combine complex features. Neural network can combine more complex features and mine deeper relationships.

Third, comprehensive behavioral characteristics and third-party portrait data. Shenze intelligent recommendation comprehensive collection of user behavior characteristics, not only click exposure, including sharing, collection, order conversion and other data will be collected, based on the construction of the recommendation system to optimize the recommendation effect. If the customer provides the third party with portrait data, the problem of cold start of the intelligent recommendation system can be better solved.

Fourth, trained engineers and algorithms. Shenze intelligently recommends engineering and algorithm teams with BAT background and has rich experience in recommendation system construction.

4. Data accumulation and empowerment

In the era of data capital, data is the core asset of an enterprise. Jack Ma once said that Alibaba is a data company, and Alibaba has achieved sustained and rapid growth with data-driven business. As the demographic dividend disappears, fine operation and data-driven become the key to the success or failure of enterprises, and data precipitation is the first step. Shenze intelligent recommendation can not only provide recommendation results that exceed customers' expectations, but also precipitate data assets for customers in the way of private deployment. Customers can use the collected data for scenarios such as refined operations, user portraits or advertising marketing to maximize the value of the data.

It is better to teach people to fish than to teach people to fish. Recommendation system, as the infrastructure of mobile Internet, is related to the success or failure of enterprises. If SaaS services are adopted, SaaS manufacturers no longer provide services or the focus of business changes, which may lead to fatal impact on enterprise business. Therefore, the recommendation system suitable for enterprises needs to be autonomous and controllable, and the ability of the recommendation system should also be precipitable. Shenze Intelligent recommendation adopts privatized deployment, and the recommendation system is deployed to the customer's server. at the same time, there will be training in the use of the recommendation system, and the customer's technical staff can also add features, adjust recalls, train models and adjust parameters. really deliver to the customer independent and controllable recommendation system.

5. Deep integration of recommendation and business

The algorithm has no values, but the enterprise has values. The algorithm serves the business of the enterprise, and in the final analysis, the recommendation result is determined by the will of the enterprise.

The recommendation system is combined with the business to maximize value. If you only pursue the recommendation index, the title party and fake and shoddy low-cost products can meet the short-term recommendation index, but in the long run, it will hurt the user experience and destroy the ecology of the platform. Only the combination of recommendation system and enterprise business can achieve value maximization.

Recommended operation background based on business awareness. Based on the above two points, we provide customers with an intelligent recommendation operation background, and users can weight and reduce the weight according to the quality of content or goods provided by self-media or suppliers. Avoid some self-media authors or merchants who study algorithms to take advantage of platform loopholes. And so on, to help the enterprise recommendation system and business integration, to achieve maximum value. Shenze Intelligence recommends a list of scenarios and an item library as shown below.

Fig. 4 list of scenarios recommended by Shenzhi Intelligence

Fig. 5 the product library recommended by Shenze Intelligence

Through the scenario list recommended by Shenze Intelligence, users can view the basic information and operation entry of the scene from the scene list. The item library can be used for the general library of recommended items in various scenarios, and can be blocked for one or more items. After blocking, the item will not be introduced in any recommended scenarios. Pan-content and pan-e-commerce enterprises are widely used, such as pan-content platforms, which need to be banned for some politically sensitive, negative energy, pornographic, gambling and drug content, and pan-e-commerce platforms. problems involving the removal of certain goods from shelves or excessive complaints about quality service problems.

Third, recommendation system helps business growth practice.

1. The customer dimension of Shenze data service.

Industry dimension: information / forum / short video / long video / novel / online education / radio and television

Business dimension: recommend 0-1 to replace expert rules, have self-built recommendations to help growth, enable multiple product lines (resident)

Growth indicators: average CTR/ consumption (pan-content industry), retention rate, order placing rate / GMV (pan-e-commerce industry)

2. Practice of helping customers' business growth.

Case 1: a content media with high timeliness and personalized recommendation increases the number of page views per capita

There is an algorithm team within the company, but it encounters the bottleneck of recommendation effect and hopes to seek third-party recommendation vendors to improve the recommendation effect. Recommend to improve the core indicator: click-through rate. The end result is CTR: 40.67% higher than self-built recommendations and 11.23% higher than competitors' recommendations.

Fig. 6 the growth practice of a content media business

Case 2: personalized recommendation of an IPTV playback platform to improve the accuracy of film and television recommendation and increase the number of visits per capita

Before cooperating with Shenze, the platform uses manual editing and recommends a full amount of fixed content on a weekly basis. There are three goals: there is an urgent need to introduce an intelligent recommendation engine to achieve personalized recommendation of "thousands of people and thousands of faces"; to improve the accuracy and update frequency of film and television recommendations; and to enhance users' viewing experience.

Fig. 7 Business growth practice of an IPTV playback platform

The effect of intelligent recommendation in this case is that, compared with manual recommendation, the CTR index is increased by 6 times; the per capita browsing times of recommended content is increased by 1.9 times; complete the training of the overall program, empower the customer algorithm team, and extend the program to other locations.

Shenze Intelligent recommendation system is an intelligent recommendation system that enables business growth. Customers currently serving include Huitoutiao, Mom help, Philippine help, Zhongqing watching Point, Vertical and horizontal Literature, Baidu Video, Oriental Pearl, Xinshou, and so on. Based on the powerful data collection ability of Shenze Analysis, it helps enterprises complete the whole recommendation process from user behavior data collection, modeling, mining to effect analysis, and from "data collection + recommendation engine + effect feedback". The following figure is the implementation flow chart of Shenze Intelligent recommendation.

Fig. 8 flow chart of intelligent recommendation landing

Finally, recommending some valuable books will be of great help to you to understand big data and intelligent recommendation.

Follow the official account of Shenze data and reply to "Intelligent recommendation driven Business" to get the full version of PPT.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.