In the end how to break the game? Let's talk about the system today. 1. The essence of prediction algorithm Essentially, there are only 2 categories of prediction algorithms: 1. Time series based Smooth: For relatively stationary data. Autoregressive: Used for trending increasing and decreasing data. Autoregressive with seasonal factors: for data with periodic fluctuations. 2. Based on causality Two-category problem: XX will/will not happen in the future, typically such as LR. Multi-classification problem: which case is ABC in the future, typically a decision tree. Continuous problem: what is the value in the future, typically linear regression.
When it is possible to model, it is not a mobile number list model package that dominates the world, but a two-stage modeling. For example, to predict the consumption of a customer group, you can use the binary model to predict whether or not to consume, and then use the continuous model to predict the consumption amount, so that the number of consumers who will consume * the predicted consumption amount, you can get the total consumption. This is a typical handling method. This is what it is taught in books, but why is it hammered into scum when it encounters reality.
Difficulties of Prediction Algorithms Because: In order to highlight the effect of the model, the book deliberately chooses a dataset with good quality and complete data. In reality the trouble is far from constant: No data. Most of the time, the data to be predicted is given in one line "total monthly consumption", and there is no other data fart... Still no data. Many companies cannot get first-hand data through Tmall, Douyin, and Amazon, and can only use a little data exported from the background to fool around... Just no data. Most of the companies are not monopoly companies like Toutengah, and only get very one-sided data. The most common, most of the company's users are paid to attract traffic, users only have a mobile phone number + a preferential order.