Seven elements of data precision marketing

Seven elements of data precision marketing

When it comes to precision marketing of big data, we have to mention personalized user portraits first. For each kind of data entity, we further decompose the data dimension that can land, depict every feature of ta, and gather together to form a portrait of the crowd.

0 1 user portrait

User portrait is a tagged user model abstracted from the information of users' social attributes, living habits and consumption behaviors. Specifically, it includes the following dimensions:

Fixed characteristics of users: gender, age, region, education level, date of birth, occupation, constellation.

User's interest characteristics: hobbies, using APP, website, browsing/collecting/commenting content, brand preference, product preference.

Social characteristics of users: living habits, marriage, social/information channel preferences, religious beliefs, family composition.

User consumption characteristics: income status, purchasing power level, commodity types, purchase channel preferences, and purchase frequency.

Dynamic characteristics of users: How to generate accurate portraits of users in current time, needs, places to go, surrounding businesses, surrounding people and news events can be roughly divided into three steps.

1. Collect and clean up data: use the known to predict the unknown.

First of all, we must master the complex data sources. Include user data, various activity data, e-mail subscriptions, online or offline databases and customer service information. This is a cumulative database; The most basic thing here is how to collect the user behavior data of the website //APP. For example, when you log on to a website, its Cookie will always stay in the browser. When a user touches actions, clicks on locations, buttons, likes, comments, fans, and access paths, you can identify and record all his/her browsing behaviors, and then constantly analyze the keywords and pages you browse, and analyze his/her short-term needs and long-term interests. You can also have a very clear understanding of each other's work, hobbies, education and other aspects by analyzing the circle of friends, which is more comprehensive and true than the forms filled out by individuals.

We use known data to find clues and constantly dig up materials, which can not only consolidate old members, but also analyze unknown customers and needs and further develop the market.

2. User grouping: classification label.

Descriptive analysis is the most basic method of analysis and statistics, which is divided into two parts: data description and index statistics. Data description: used to describe the basic situation of data, including the total number, scope and source of data. Indicator statistics: model the distribution, comparison and prediction indicators. There are often some mathematical models of data mining, such as response rate analysis model and customer tendency model. This kind of grouping uses the Lift diagram to tell you which kind of customers have higher contact and conversion value through scoring.

In the analysis stage, the data will be transformed into an impact index, and then you can do "one-on-one" precision marketing. For example, a post-80s customer likes to order food in the morning 10, go home to cook at 6 pm, and go to eat Japanese food nearby on weekends. After the collection and transformation, some labels will be generated, including "post-80 s", "fresh", "cuisine" and "Japanese cuisine", which will be attached to consumers.

3. Formulate strategies: optimize and readjust.

With user portraits, we can clearly understand the needs. In practice, we can deeply manage customer relationships and even find opportunities to spread word of mouth. For example, in the above example, if there is a fresh discount coupon and the latest recommendation of a Japanese restaurant, the marketer will accurately push the relevant information suitable for the product to the consumer's mobile phone; Send the recommendation information of different products, and at the same time, know the customer's behavior and preferences in time through satisfaction survey and tracking code confirmation.

In addition to customer grouping, marketers also observe the growth rate and success rate at different stages, and compare them before and after to confirm whether the overall business strategy and direction are correct; If the effect is not good, what strategies should be used to deal with it. Repeat trial and error, adjust the model and realize circular optimization.

The purpose of this stage is to refine the value, then accurately market according to the customer's needs, and finally track the customer's information feedback to complete the closed-loop optimization.

We start with data integration and import, summarize data, analyze and mine data. There are still some differences between data analysis and mining. The focus of data analysis is to observe the data, make simple statistics and see the reasons for the rise and fall of KPI. Data mining studies data from the perspective of subtlety and model, and discovers knowledge rules from learning sets and training sets. In addition to some commercial software SAS and WEKA's powerful data analysis and mining software, R and Python are recommended here, because SAS and SPSS are expensive and it is difficult to make page and service-level APIs, while Python and R have rich libraries, which can be similar to WEKA's modules and interact seamlessly with other APIs and programs. Here, you need to be familiar with the database.

02 data segmentation audience

An example mentioned in the book Subversive Marketing can be quoted. Let's think about a question: If you plan to collect 200 valid questionnaires, according to past experience, how many questionnaires do you need to send out to achieve this goal? What is the estimated budget and time for implementation?

The previous method was as follows: the recovery rate of online questionnaire was around 5%. To ensure that 200 questionnaires are received, it is necessary to send 20 times as many questionnaires, that is, 4,000 questionnaires. If you can recover within one month, it is a good performance.

But it's different now. Within 3 hours of performing big data analysis, you can easily achieve the following goals:

Select 1% VIP customers accurately.

A total of 390 questionnaires were sent out and all of them were recovered.

35% of the questionnaires were returned within 3 hours after mailing.

Within five days, more than 86% of the questionnaires were returned.

The required time and budget are lower than the previous 10%.

How did this achieve 35% recovery within 3 hours after the questionnaire was sent out? That's because the data has been customized one to one. Using the data, it can be concluded that when Mr. A is most likely to open the email, he will send out the questionnaire at that time.

For example, some people will open their mailboxes on their way to work, but if they are drivers, they won't have time to fill in the answers, while people who take public transportation will play with their mobile phones on their way to work, so the probability of filling in the answers is high. These are the benefits of data segmentation.

03 pre-test

Forecasting allows you to focus on a small group of customers, but this group of customers can represent most potential buyers of a specific product.

When we collect and analyze user portraits, we can achieve precise marketing. This is the most direct and valuable application. Advertisers can post advertisements to users who want to reach them through user tags. In this case, they can improve their ROI through the aforementioned back-end CRM/ supply chain system, multi-channel marketing strategy, marketing analysis, marketing optimization and one-stop marketing optimization.

Let's talk about the changes in the marketing era. Most traditional enterprises still stay in the era of "marketing 1.0", focusing on products to meet traditional consumption needs and entering "marketing 2.0", taking social value and brand as their mission, and unable to fully and accurately meet individual needs. In the data age of Marketing 3.0, we need to personalize each consumer, conduct one-on-one marketing, and even accurately calculate the transaction conversion rate to improve the return on investment.

Marketing under big data subverts the classic marketing 4P theory, and is replaced by new 4P, people, performance, process and forecast. In the era of big data, the competitive boundaries of offline regions have long since disappeared. It is better than an early prophet's ability to predict the next purchase time from the real transaction data of customers by using big data. The key word in the era of marketing 3.0 is "prediction".

Predictive marketing allows you to focus on a small group of customers, but this group of customers can represent most potential buyers of a specific product. Take the above picture as an example. You can target your marketing activities to 200,000 potential customers or existing customers, including most buyers (40,000 people) of specific products. You can also allocate part of your budget to attract a smaller customer base (such as 20% customers) instead of the whole customer base, thus optimizing your expenses.

In the past, we may look at data passively, but predictive marketing emphasizes decision-making value, such as purchase time. What you want to see is not the date of her last purchase, but the time of her next purchase, and the survival probability in the future, which will eventually produce the customer lifetime value (CLV). Predictive marketing has given birth to a new data-driven marketing method, which is customer-centric. The core is to help enterprises complete the transformation from product or channel-centric to customer-centric.

04 accurate recommendation

The greatest value of big data is not ex post analysis, but prediction and recommendation. I take e-commerce as an example. "Accurate recommendation" has become the core function of big data to change the retail industry. Take Stitch fix, a clothing website, as an example. In terms of personalized recommendation mechanism, most clothing ordering websites adopt the mode that users submit body and style data+edit manual recommendation. Stitch Fix is different in that it also combines machine algorithm recommendation. These customers provide body proportions, subjective data, cross-check sales records, and mine each person's unique clothing recommendation model. This one-on-one marketing is the best service.

Data integration has changed the marketing model of enterprises. Now experience is not accumulated in people, but depends entirely on consumer behavior data to make recommendations. In the future, salespeople will no longer be just salespeople, but will be able to recommend products with professional data prediction and humanized friendly interaction, and upgrade to consultative selling.

05 technical tools

There are several options to predict the technical ability of marketing:

1. Use the forecast analysis work platform, and then input the model into the activity management tool in some way;

2. Analysis-driven forecasting activities are outsourced to market service providers;

3. Evaluate and purchase predictive marketing solutions, such as predictive marketing cloud and multi-channel event management tools.

But either way, we must determine three basic abilities:

1) Connect customer data from different sources, including online and offline, and prepare data for forecasting and analysis;

2) Analyze customer data, use the system and customize the prediction model to do advanced analysis;

3) Start the right behavior at the right time, in the right customer and in the right scene, and may cross-sell across different marketing systems.

06 prediction model

The industry standard for predicting customers' purchase possibility is RFM model (latest consumption R, consumption frequency F and consumption amount M), but the application of this model is limited, and it is essentially a tentative scheme without statistical and forecasting basis. "Past achievements cannot guarantee future performance", RFM only focuses on the past and does not compare the current behavior of customers with that of other customers. This makes it impossible to identify high-value customers before buying products.

The prediction model we focus on is to have the greatest impact on customer value in the shortest time. Here are some other model references:

The participation tendency model predicts the possibility of customers participating in a brand, and the definition of participation can be diversified, such as attending an activity, opening an email and clicking on a page. This model can determine the transmission frequency of EDM. And predict the trend, whether to increase or decrease activities.

Wallet model is to predict the maximum possible expenditure of each customer, which is defined as the annual maximum expenditure of a single customer to buy products. Then look at the growth model, if the current total target market is relatively small, but it may be large in the future, we need to find these markets.

The price optimization model is a structure that can maximize sales volume, sales volume or profit. Pricing for each customer through price optimization model. Here, you need to develop different models for the products you want, or develop a general and predictable customer price sensitivity model to determine which quotation has the greatest impact on customers.

Keyword recommendation model can predict customers' liking for a certain content and what hot spots and explosions customers are interested in according to their online behavior and purchase records. Marketers use this prediction result to decide the content marketing theme of a specific customer.

Forecast aggregation model, forecast aggregation model is to predict which category customers will belong to.

Application of 07AI in Marketing Field

Last year, artificial intelligence was particularly hot, especially the rapid development of deep learning such as machine vision, language recognition and game AI, which made people start to panic whether artificial intelligence can take over human work. I personally have a strong interest in new technologies, and I am very optimistic about the relationship between new technologies, data and reality.

I used to be asked "Do you have a shopping card" when I paid the bill in a foreign retail store? When I said there was no cashier, I would quickly persuade me to open it for free, and there was a discount. I just need to fill in my mobile phone number and email address, and then I can do marketing activities for my purchase records. When I come next time, they ask me to quote my phone number as a consumer identification. At that time, I thought it would be more convenient to do face recognition, and I could pay the bill by brushing my face. And this scene also had an experiment last year. Ant Financial has developed a biometric robot named Mike Mark, which is said to have exceeded the ability of human eyes to recognize other people's faces. There is also VR shopping, Amazon Go, a cashier-free store launched by Amazon, which realizes the shopping experience through gesture recognition, Internet of Things and subsequent data mining.

For the marketing field, there are mainly the following three forecasting marketing technologies:

1, unsupervised learning technology

Unsupervised learning technology can identify hidden patterns in data without clear prediction results. For example, finding interest groups among a group of customers, perhaps skiing or long-distance running, is usually put into clustering algorithm to reveal the real potential customers in the data set. The so-called clustering is to automatically find important customer attributes and classify them accordingly.

2. Supervised learning technology

Train the machine through the case, learn the identification data, and get the target result. This is generally a prediction of given input data, such as predicting the life cycle value of customers, the possibility of interaction between customers and brands, and the possibility of future purchases.

3. Strengthen learning skills.

This is to use the potential patterns in the data to accurately predict the best selection results, such as which products should be provided for a user's promotion. This is different from supervised learning. Reinforcement learning algorithm not only needs input and output training, but also completes the learning process through trial and error.

From the technical point of view, the recommendation model uses collaborative filtering, Bayesian network and other algorithm models. Reinforcement learning is considered by Jeff Dean, the head of Google's brain team, as one of the most promising AI research directions. Recently, DeepMind, Google's AI team, published a paper entitled "Learning Reinforcement Learning".

According to the team's words, it is called the ability to "learn to learn" or the inductive ability to solve similar related problems. In addition to intensive learning, it is also transfer learning. Transfer learning is to transfer a general model to a small data to make it personalized, and it can also produce effects in new fields, similar to human analogy.

Reinforcement learning and transfer learning can also use small data, which I think is very exciting. Creating AI through AI can also make some work of data scientists realized by machines.