How are you targeted by ads?

Nowadays, enterprises are paying more and more attention to using artificial intelligence technology to gain insights into data value and develop value-added services. The traditional marketing method that relies on business rules to screen target users has the problems of long data acquisition cycle and low marketing contact rate. This article will introduce the currently popular lookalike expansion technology, which can effectively support business parties in precision marketing.

1. What is Lookalike technology?

Lookalike technology, or similar population expansion technology, is a method of finding a larger group of similar users from a small group of target users through an algorithm model. The core idea is to first build a profile of a small number of target users (seed users) based on their characteristics, such as demographic data, network behavior, consumer preferences, etc. Then, find users who are more similar to these target users in a larger user population. These newly found users are the "similar" group of the target users.

Figure 1 What is Lookalike

It should be noted that lookalike technology is not a specific type of algorithm, but a general term for a series of algorithmic methods. These algorithms use a variety of technical means such as machine learning, statistical analysis, and data mining to achieve the purpose of expanding from a small-scale target user group to a larger similar user group through the evaluation and modeling of user characteristics.

2. How to use Lookalike technology

The core value of Lookalike technology lies in achieving accurate expansion of small-scale users. Specifically, in actual business, advertisers will first select a small number of high-quality typical users as seed users based on their own rules and experience. These seed users can accurately represent the target customer group, but due to the strict selection rules, the number of accurate target customers is limited. At this time, the Lookalike algorithm is used to analyze the characteristic attributes of the seed users, and then find similar users from the full user database. These newly matched users can meet the marketing needs of the business because they are similar to the original seed users in some dimensions. In this way, the Lookalike algorithm can significantly expand the user base based on limited seed users to continuously meet the needs of advertisers to obtain more high-quality customers.

Take car sales as an example. A car brand wants to promote its new SUV in an emerging city, with the goal of covering 1 million drivers in the city who are interested in buying a car. However, based on the historical purchase data of this model, the target car buyer population is only 200,000. In order to reach more potential customers, the car brand uses Lookalike technology, first selecting customers who have recently inquired about or visited this model as seed users, and then matching 800,000 similar users from the entire driving population in the city based on the characteristics of these seed users, such as region, income, lifestyle, etc. In this way, this promotion can reach more target groups that meet brand values ​​and car purchase preferences.

Let's take another example of education and training. An online education platform is preparing to launch a Python programming course, with the goal of pushing the course information to 5,000 students. However, according to the platform's registered user data, there may be only 1,000 people who are interested in Python. In order to expand the influence of the course, the platform decided to use Lookalike technology to first select registered users who have signed up for programming courses in the past year as the seed group, who represent the attributes of users with demand. Then, the platform can match 4,000 potential interested users similar to the seed users from all registered users based on the characteristics of these seed users, such as educational background, career direction, learning preferences, etc. In this way, the new course can reach more target students who have a demand for programming.

3. How to implement Lookalike technology

Figure 2 How to implement Lookalike

Lookalike similarity group expansion methods mainly include the following methods:

(1) Explicit Lookalike Based on User Attributes

This method finds similar users based on user attribute characteristics. Specifically, demographic characteristics such as age, income, education level, region, etc. can be used. When using it, you can first define the core attributes to be matched, and then extract users with similar attribute values ​​from the user group. This method is simple to implement, but it is limited to a few pre-defined attributes and may miss other implicit characteristics that describe the user.

(2) Implicit Lookalike Based on User Behavior

This method will comprehensively analyze various types of user behavior data, such as page browsing time, click frequency, search terms, consumption amount, interests, etc., use seed users as positive samples, and downsample random users as negative samples, use machine learning algorithms to model user interests and consumption preferences, and then match users with similar preferences and behaviors. The advantage is that it can make full use of rich user behavior data, but the disadvantage is that it needs to collect a large amount of behavior data for modeling.

(3) Implicit Lookalike Based on Graph Database

This method builds a network structure of user nodes and user relationship edges in the graph database. After identifying the seed users, it propagates the features and labels to the adjacent nodes based on the edge relationships between the nodes. It continuously finds new similar users in the network graph through recursive iteration, and repeats the process until a sufficient number of people are found, thereby achieving efficient expansion of target users. This technology makes full use of the modeling and computing advantages of the graph database for network data, and can effectively discover more similar users that are difficult to observe, but it needs to deal with the storage and computing problems of large-scale relationship graphs.

This article introduces the relevant concepts and application methods of Lookalike technology. However, in actual business applications, how to scientifically select features to describe users to ensure the effectiveness of subsequent algorithms, and how to collect marketing feedback to continuously optimize machine learning models and form a closed loop are key issues that need to be solved. In short, no algorithm exists in isolation and must be reasonably applied in combination with data and business scenarios to maximize its value.

References

【1】Alchemy Notes. Search promotion meets user portrait: Lookalike similarity group expansion algorithm | CSDN, 2021

Author: Zhang Hao

Unit: China Mobile Smart Home Operation Center