从字段库一路做到名单。两个模型:一个回答会不会回来,一个回答以什么方式回来。前者用来排全年,后者用来排单次活动。
这份清单不是从零开始列的。它来自做过的线下召回项目:哪些字段在这个场景下值得放进池子,哪些算得出来,哪些要另外想办法。每做一个项目,清单往回补一次。
长期召回规划用的是回归模型给每位沉睡客打的分。先回第二步把模型训练出来,这里的人数和概率会跟着第二步的人群定义走。
单次活动圈人不训练模型,只取第二步训练好的来打分。先回第二步训练回归模型,再至少训练一个回归特异模型。训练好哪个 label,这里就能选哪类活动。
One line from field library to contact list. Two models: one answers whether she will come back, the other answers how. The first plans the year; the second picks the list for a single campaign.
This list was not written from scratch. It comes from in-store win-back projects we have run: which fields are worth putting in the pool for this use case, which can be computed, and which need another source. Every project adds back to it.
The year-long plan uses the score the return model gives each lapsed customer. Train the model in step 2 first; the counts and probabilities here follow the audience defined there.
This step trains nothing; it scores with models trained in step 2. Train the return model there, then at least one return-type model. Each trained label makes that campaign type available here.