First, judge whether the enterprise needs “thousands of faces”
“A thousand faces” is easy to understand. It means finding the right person and delivering the right message in the right form. At present, almost all Internet head products are recommended by the rules of “thousands of faces”, so that sometimes when we see the content we don’t like, we will doubt whether we really don’t like it or whether we have unexplored interests and preferences.
In such a market environment, whether to go with the flow, all “bring”? In my opinion, enterprises first need to consider the value that “thousands of faces” can bring to them. The core lies in the following two points:
First, improve the distribution efficiency of content to users.
If manual policies are used or no policies are used, users can also see their favorite content, but the search cost is high and needs to be borne by users. Intelligent strategies such as machine learning can reduce users’ search costs and improve the efficiency of content distribution to users.
Second, improve the execution efficiency of internal workflow.
Some enterprises often use manual way to run the closed loop operation, from activity planning to execution, to monitoring, and then to the review, if the use of data intelligence in this process, then improve not only the work efficiency, but also have a positive driving force for the effect. In other words, organizations need to spend more time on policy tuning than on implementing the landing process.
Two, how to achieve a thousand people thousand faces
As shown in the figure below, a thousand faces can be divided into three stages:
1. Low order: artificial decision and intelligent execution.
Scenario 1: Planned marketing, usually in the form of a one-off, cyclical operation strategy.
For example, in the first 7 days or 1 day of the promotion of e-commerce enterprises, the one-off information push of enterprises to users belongs to the single planned marketing. Monthly salary day, repayment day of the message reminder needs to be repeated, regular implementation, this is cyclical marketing.
Scenario 2: Hierarchical recommendation is displayed based on user layers.
When users arrive at the product environment, enterprises can effectively use layered recommendation to achieve effect improvement. The common forms are: startup chart, Banner chart and rotation chart. These three recommended items tend to have a small weight, and the update and iteration speed is fast. In addition, the target audience has been basically defined in the design of these three recommended rules. In this scenario, the goal of tiered operations is basically achieved by employing human decision making.
2. Advanced: Artificial and intelligence make decisions together.
At this stage, we sort out two scenarios of touch marketing and refined layered recommendation.
(1) Contact marketing usually refers to the decision of strategic direction by manual, machine-assisted computing to determine the trigger time.
For example, when the user browsed the product for many times but failed to convert, the machine can be set to trigger the coupon push strategy in time when the user has no purchase behavior after browsing the product for 30 minutes to improve the conversion efficiency of the user.
Another example is that for new customers, we hope to deepen their experience of product value step by step, and we usually make user contact on the first, seventh and 30th day of new customers’ entry. If we only rely on manual pull list to complete push, it will take time and energy, and the behavior tracking of new customers will be realized through machines. You can easily automate push on a specific date.
(2) Refined stratification is a process of refined recommendation for users’ personalized behaviors within the product.
For example, the function recommendation menu of the banking industry generally contains more than 100 functions. When users enter the product, it is difficult to determine which function they really need in the first time. At this time, enterprises need to sort out these 100+ functions and which users are suitable for each function. Then the function display is sorted based on the user’s access frequency path in the past period of time, that is to say, artificial + intelligence jointly realizes the refined recommendation.
3. Advanced: Intelligent promotion of the whole process from decision making to implementation to feedback.
When it comes to “a thousand faces”, most people immediately think of whole-process intelligence, which can also be understood in accordance with the current popular machine algorithms and deep learning concepts. The application scenarios are as follows:
First, intelligent marketing, that is, automated and personalized marketing based on algorithmic programs, relies on marketing push triggered by machine recognition. At present, this scenario is not widely used.
Second, intelligent recommendation, personalized recommendation based on algorithm model, is mainly used for information flow, relevant recommendation, popular recommendation, etc.
The core of the “thousands of faces” is ROI, and there are some hard and fast conditions under this principle:
(1) User magnitude and item magnitude. In the product solution of Shence intelligent recommendation, we have a certain number of requirements for the customer’s inventory. When the magnitude of items is lower than 5000 or the daily activity does not reach a specific level, it is not suitable for algorithm recommendation.
(2) the construction degree of user labels and item labels. If there is not enough data to support users’ stratification and identification of object features, it will be difficult to carry out the refined operation.
(3) Real-time behavioral data flow. Data intelligence applications at all stages rely on real-time collection of data, and then it is possible to realize personalized push based on users’ browsing behavior. When all three of the above are met, the enterprise can decide on scenarios and priorities based on its current ROI.
For algorithm implementation of “one thousand”, “it is usually the process from the data system, user behavior data and instill into algorithm recommendation system, through a series of processing, recommend the most appropriate results, and then returns the result the user front-end do show, monitoring the efficacy of users to click on at the same time, in order to determine the recommended effect the quality of the continuous optimization, Form a complete closed loop recommended by the algorithm. As shown below:
1. Item recall, personalized display of items suitable for users to see.
2. Sort, sort the selected items based on a variety of judgment conditions to ensure that they can produce good exposure and conversion effect.
3. Reorder, which requires a lot of service input. Operating personnel, for example, tonal and diversity have certain requirements for products, such as a user associated with pets tend to see a short video, but it is hard to have a product can only play pet class video, which requires enterprises to realize the long-term clear judgment should depend entirely on the algorithm of user behavior output short-term effect.
To make good use of algorithm recommendation, algorithm-based data intelligence depends on the dual escort of technology and business:
First, an intelligent recommendation system that operates efficiently.
First, it presupposes that accurate user behavior data can be obtained.
Secondly, the algorithm model itself. The algorithm function of Shence is better than the recommendation effect written by the self-recommendation or other models used by some customers. We did not even carry out in-depth tuning, which fully reflects the superiority of the algorithm model itself.
Second, system tuning based on business logic.
Algorithms can’t help us solve all problems, and the typical scenario is cold start. In view of this, I put forward two suggestions: (1) Undertake information and content from the channels before new users enter the product to ensure that users can see the content that meets their expectations and needs after entering the product; ② A well-designed strategy proactively collects information from users.
Most of the time, the algorithm does not rely on the construction of user tags and item tags, so some business people will ignore the importance of tag construction; However, the effect of the algorithm to solve the recommendation problem needs to be judged by data analysis, which is essentially a process of layer by layer disassembling an index. If the construction of user portrait label/item label is not perfect, the efficiency and quality of recommendation will be affected to some extent. Therefore, although algorithms can help us solve many problems, they also require us to pay attention to the construction of our own data.
For manual intervention, it usually involves some specific links, such as banning behavior for specific items, parameter adjustment, reordering strategy and so on.
Based on the above, we can understand that the algorithm is essentially realized in combination with business, and there are certain thresholds, which are embodied in business, technology and people:
1. Whether the business model is suitable for algorithm to solve the problem of “thousands of faces”. First of all, the biggest value of the algorithm lies in the efficiency of the distribution of content to users, so whether the enterprise attaches importance to the efficiency of the distribution of content to users under the business model and how influential it is.
Then, in the stage of business development, we are bound to give priority to the construction of content and pull new actions, so do we need to invest a lot of energy to do the algorithm system at this time, to achieve “thousands of thousands of faces”?
Finally, whether the number of items and the number of active users are up to par.
2. Whether there are enough technical resources and big data foundation to support enterprises to do a good job in algorithms, or achieve “thousands of faces” through other forms.
3. Whether the participants have the theoretical and practical abilities of artificial intelligence.
** point, break point marketing. ** Sort out the reasons for the loss of users in the business process, and take targeted measures to recover them.
For example, the user visits the financial details page but does not complete the transformation within the specified time, or the login fails due to some reasons, leading to the loss of users… After sorting out the user behavior process, formulate specific strategies for the risk of loss in each link. Push push to recommend more appropriate financial products for users who have not purchased them. For users who submit orders and fail to pay, push in the station will cooperate with SMS touch and send coupons to facilitate user conversion.
** surface, layered operation. ** by stratifying all users and planning the operation mode suitable for each stratification, so as to achieve thousands of faces. Business people often do not know how to go about designing an operation plan. We suggest that a common model can be used to diagnose the existing problems of the product and then design a high priority operation strategy for the problem areas.
Generally, there are two scenarios:
Scenario 1: Proactively reach users based on push outside the product
For example, based on the customer life cycle model, we need to screen out the key links from the journey of new users under the condition of “good results in attracting new users but unsatisfactory transformation”, and formulate automatic operation strategies for the whole process of attracting new users, activation transformation, repurchase, loss recall and so on. At the same time, it can also determine whether the operation strategy in the activation link is helpful to transformation and improvement, insight into the operation quality and effect, and improve the operation efficiency.
Scenario 2: Improve user conversion through in-product display
First of all, the operation positions in the product should be sorted out, and the user stratified operation should be reflected in the formulation of strategies, such as setting function menu operation plan and product recommendation plan respectively for new and old users. In the process of setting up and operating the recommendation plan, differentiated content can be displayed for specific groups by combining the user label system, so as to achieve thousands of faces.
Generally speaking, thousands of faces is a threshold thing, specific in the process of construction need to pay attention to the following issues:
1. Data basis and technical support
(1) Promote the construction of user and item labels;
(2) The degree of integration of data flow and business flow, which directly determines the execution efficiency and quality of the machine after we formulate strategies;
(3) Channel construction and management, which is a problem that troubles many operators. In terms of channel construction, high-quality third-party tools and self-built systems can be selected to ensure that push can be successfully delivered. After solving the technical problem of channel construction, it is necessary to select the appropriate scene to request users to open the channel. After getting through the access to the user’s path, we should pay attention to the management of the channel, just like the Shence intelligent operation system, it can help customers to actively choose the number of users in a certain period of time, etc..
2. Business practical experience accumulation and review
(1) Systematic operation thinking. Based on the above methodology, the operation strategy is sorted out so as to have a complete understanding of the global operation.
(2) Effect verification and knowledge precipitation, effect recording and display through the machine, to break the experience loss caused by personnel flow. Its essence is to help us improve the efficiency of online strategy and real-time monitoring of the efficiency of result evaluation.
3. Productivity tools
If the efficiency tool can play its value, it can directly realize low-order operation, that is, after the human decision is made, the machine will execute it to release the human cost and improve the efficiency of the human decision and evaluation. But in this process, the efficiency tool needs to have the following capabilities:
(1) Visual strategy editing ability that integrates with data and products;
(2) Online management ability;
(3) Ability of action effect recovery and evaluation.
Only with the ability of these three aspects, can help operators improve work efficiency and liberate human resources in a real sense.
Three, how to judge the landing effect of thousands of thousands of faces?
From the perspective of index effect evaluation, it can be divided into three levels:
1. Scene transformation, which is applicable to click rate, click number ratio, per capita click times, etc., can intuitively tell me whether the recommendation is effective for user transformation.
2. Content satisfaction, which can be measured in terms of consummation rate, consumer market, retention improvement, etc.
3. Business objectives are achieved, which can be broken down into activities’ participation transformation, business transaction transformation, ROI improvement, etc.
The above three levels actually have a progressive relationship. During the performance evaluation process, we sometimes see an increase in CTR, but no increase in conversion rate. Why is this? In fact, business conversion rates are influenced by many factors, including the content itself, product quality, price, etc.
From the perspective of evaluation methods, it needs to be carried out based on the company’s business performance, which is usually divided into two forms: 1. Version comparison, whether it is grayscale test or the comparison before and after the new strategy, needs to consider the change of the environment and target population.
2.A/B test, which can exclude the influence of external factors at the same time to evaluate the effect. However, if an enterprise wants to conduct A/B test, it must ensure sufficient sample size to ensure that the business effect and data statistical effect are significant enough.Shence data A/B test function has been released, welcome new and old customers experience!
In addition, enterprises should clearly recognize that the performance evaluation always depends on the construction of the data base.
2. Construction of label system.
3. Productivity tools. Detailed effect assessment on a topic often involves constant exploration, hypothesis and verification. If no data analysis platform plays a role in this process, it will lead to inefficient communication and execution across functional lines.
I hope that after this sharing, we can have a further understanding of data intelligence to achieve thousands of faces. Thank you!