We have talked about so many recommendation algorithms before. How should we apply them specifically in data products? Can these recommendation logics only be applied in recommendation systems?

I don't think so. User interests can permeate almost every aspect of work in business logic. Simply put: no matter what business it is, whenever you come into contact with a customer, isn't it the best strategy to give him his favorite things?

The answer may not be, but it is still very valuable in most fields. In this article, I will share with you two common application directions:

1) Activity popularity, best crowd/activity recommendations:

• We designed a sparkling event based on our experience, but does it really suit the company’s current customer base?
• Recently, the company wants to give back to old customers, and has identified a group of high-value customers. What types of activities are suitable for these customers or what types of coupons do they like?
• The company has added a new partner and discussed a batch of new coupons. We want to know which customers these new coupons are suitable for?

2) Targeting unusual people, push coupons that they like to them:

The company's recent goal is to promote activation and conversion. Faced with the abnormal customers screened out (for example: active but not converted, dormant customers, and those waiting to be lost), what kind of coupons should be given to activate them?

How to solve the above problems?

Millions of people have millions of ideas. This article explores the algorithmic solution from an algorithmic perspective. This algorithmic model is: user interest model - also called marketing response model.

Therefore, before solving the problem, let's take a look at the interest model construction process:

1. Marketing Interest Model

When exploring customers’ interests, we drew on a commonly used model in recommendation systems, DeepFM. This model is widely used in click-through rate prediction and recommendation ranking because it can effectively and deeply integrate high-dimensional and low-dimensional features.

If we think about this model carefully, click-through rate prediction and recommendation ranking problems are essentially based on the user's interactive behavior towards the product, mixing the basic attributes of the user/product, calculating the probability value of the user liking a certain product, and then inferring whether the user will click on it.

There is a basic assumption in the process of building this model: users are more likely to click on something if they like it. Regardless of whether this assumption is necessarily true, just looking at the first half, the model predicts whether the user likes a certain product, which is enough for our application.

During the model construction process, we used activity data as the entry point to obtain the basic attributes of users, the basic attributes of activities, and the behavioral interaction data of users on activities; we integrated these three types of data and fed them into the DeepFM model to calculate the user's preference for the activity, that is, the interest level.

There is an interesting point here, that is, the feature library simplifies the difficulty of feature engineering.

Under normal logic, algorithm engineers need to conduct detailed feature screening, list the existing features, and determine which features are correlated with the target value through calculation methods such as correlation or entropy, and then screen out strongly correlated features and correlation weights.

This process often takes a long time, but the emergence of feature libraries has simplified this work. I will describe the working principle of feature libraries in detail in the following articles.

Here, we simply understand that we put the y value and the key experience value of x (some designs will omit the x key value and only input the y value, depending on the design completeness of the feature library) into the feature library. The feature library will return to you other features that are strongly correlated with the y value and the x experience value and the corresponding correlation weights, as follows:

Engineers only need to perform simple business-related processing such as missing values ​​and discretization on these features, and then they can directly feed them into the model.

Do you find it interesting?

The power of science and technology will gradually replace manual labor, just as cars replaced horse-drawn carriages and machines replaced labor.

Let’s get back to the topic. I described the DeepFM model in the series on recommendation algorithms.

Since this article is about the application of algorithms in product design, in terms of algorithms, the style of the model is simply posted. Interested students can explore it in depth:

The whole operation process is:

After the above discussion, we have obtained a comparison table of users' interest in a certain activity:

Our next task is to use these three tables to design corresponding application scenarios.

2. Strategy Operation Sandbox

During the operation process, event design often encounters a problem, namely: the "curse of knowledge". The event designers' rich experience makes them believe that the event should be designed in this way and what the customer groups at a certain stage should like, and they gradually ignore the exploration of new solutions and the real preferences of customers.

Can this "curse" be broken?

Today we try to provide a new method from the perspective of user interest.

The design ideas for many activities come from past experience and the experience of other companies. This is actually our main method of continuous learning, that is, learning from history and getting inspiration from other people. There is nothing wrong with this in itself, and the model construction itself also follows this idea, that is, summarizing and inferring what our current customers like from the activity data of the past n years, and then designing what kind of activities for them. The problem is that people cannot judge user preferences and design activity plans from the perspective of the entire data set. On the one hand, it is because the amount of calculation is too large, and on the other hand, it is difficult to extract effective information from too complex data.

So, there is an interest model~

The value of the interest model lies in its ability to leverage its powerful computing power to traverse n years of valid data of the company and use the past behavior of existing customers to judge their current preferences. New customers can also roughly judge their current preferences based on their similarity with existing customers, thus indirectly solving the human limitations mentioned above.

But what are the limitations of the interest model itself?

It’s very simple. The limitation is that the model cannot read the news. What does that mean?

That is, the model cannot get inspiration from other people and lacks imagination, unless you feed other people's experience data into the model; and this is often impossible to do because other companies cannot provide you with the data you want; therefore, relying solely on interest models is often unsuccessful because its ability to explore and inspire unfamiliar things is far inferior to that of humans.

So, there is a combination of man and machine~

Human-machine integration mainly refers to the process in which people draw on the interest model's ability to mine historical data during the plan-making process, integrating their own ideas and inspirations to form a more complete plan.

Strategic Operation Sandbox came into being~

This sandbox has three main capabilities:

1. Do existing activities match customer groups?

If the activities you are running come from your own ideas or past experiences, are they suitable for the company's current customer base?

We can break down the activities, identify recognizable activity type features, and feed them into the interest table above to identify a group of active customer groups. We can then compare this group of customer groups with the customer groups originally determined for the activity, take the intersection to view the distribution of customer interests, and then determine whether the activity design is suitable for the current customer groups.

2. When facing new customer groups, which types of activities have a high degree of match?

This question has some shadows of recommendation. According to the new customer group, we circle the corresponding activity characteristics in the interest table; and sort them according to the interest level of each activity characteristic, and check the interest distribution of each activity characteristic. The activity characteristics with generally high interest levels are the activity characteristics that meet the customer group, and we can use our imagination around this activity characteristic.

3. In the face of new coupons, which customer groups should we design activities for?

One of the main activity characteristics for evaluating interest in the interest model is coupons. When a new coupon is generated, we can filter out coupon customer groups similar to existing coupons in the activity type list in the interest table; and sort them according to the interest value, filter out customer groups with interest levels greater than a certain threshold, and use them as target customer groups to design new activities.

If the number of screened customers is not enough, you can use this part of the customer base as the seed customer base and use the method of population diffusion to expand the number of customers.

At this point, the three application directions of the strategic operation sandbox are clear. Welcome friends who are interested to chat~

3. Promotion/recall engine

We have talked about the event design process before. Next, let’s talk about the issue of promotion/recall.

The problem of activation/recall is essentially the problem of stimulating abnormal customer changes, which is mainly divided into two steps:

1. Identify which customers are abnormal customers

As operators, we need to pay attention to the following anomalies: active customers who do not convert, inactive customers, and customers who are about to churn. These three types of customers are relatively theoretical. When combined with specific business scenarios, there are often more detailed divisions, such as customers who frequently browse the homepage without in-depth visits, customers who only make inquiries but do not make specific transactions, customers who frequently use one function without involving other functions, etc. These abnormal behaviors usually reflect the psychological contradictions of customers.

In a complete life cycle, a normal customer visit often goes from initial visit to in-depth browsing, from just making inquiries to completing specific transactions, from completing a single transaction to completing multiple types of transactions, from low-frequency visits to high-frequency visits, and so on.

As customers become more familiar with the APP, their use of the APP will gradually deepen. Otherwise, there will be anomalies, and this anomaly is something we should explore repeatedly.

In this article, we will take three categories as examples: active but not converted customers, inactive customers, and customers to be lost, to explore the application of interest models.

2. Determine what coupons abnormal customers prefer

As we mentioned above, the most distinctive activity type is coupons, which is consistent with our understanding; therefore, we can extract the user's interest in coupons from the interest table and apply it here.

Now that you have identified the abnormal customers and figured out what kind of coupons these customers prefer, what do you need to do next?

I think it's clear now, right?

Send these customers their favorite coupons in an attempt to maximize the activation of these customers and achieve the goal of promoting changes in abnormal customers.

In terms of engineering, we can present it in the form of a flow chart:

From the above information flow, we can basically see the usage logic of this activation/recall engine, namely:

1. Identify unusual customers and coupons of interest and configure them into the tag library;
2. Connect the tag library to the event center and trigger the push cancellation platform according to certain rules;
3. The promotion platform obtains abnormal customers and corresponding coupons and sends them in a personalized manner;

There are two highlights of this idea:

• The interest model will adjust the coupons of interest based on changes in user interests. This ensures that the same customer receives different types of coupons at different times, and prevents customers from being disgusted by the same coupons.
• The event center sends coupons in real time, which not only ensures real-time performance, but also the rule configuration can increase the flexibility of the coupon sending platform (sending in the morning or afternoon has different impacts on different customers), satisfying customers in terms of time dimension.

Well, the article ends here. For the interest model, maybe you will have better application ideas. Welcome to chat with us~

Author: livandata

Source: livandata