Summary of project failure in conversion rate prediction

  Our project goal is to predict the conversion rate of each user's app game and e-commerce after clicking the ad. Period of two semi-development, from the beginning to complete a total of 3 months, completed the first version of the conversion rate prediction. There's really a lot to learn from it. Learn not to say, just mention the feeling to do not good place.       Error 1: You should not mix e-commerce and app data. It is even possible to do a category of more balanced app data, the consequences of this error is that the model to predict the two tasks, the model of all the parameter adjustment to consider the impact of these two tasks, often a good one bad. This creates a lot of stress for our first edition, especially when it comes to handover and KPIs. Even worse is that the app can be considered a sort of balanced data, and the e-commerce is a serious category of skewed data, now I wonder if the e-commerce data we really trained.       Error 2: Features begin with complexity. We have doc features, user features, crossover features. Colleagues who do the previous version have shown that using only doc features can improve conversion rates. Instead of using DOC features to run the system first, we added too many features, and I had more than three weeks to write the Hadoop job to handle the feature logic. If we only use Doc features to train, parallel to the work of adding other features, the progress will not be three months before a version.       Error 3: Experiments with features and parameters do not leave a result that ordinary people can understand. First of all, the feature, the addition of features, there is a very obvious randomness, not through the experiment to explain the importance of the feature, nor seriously through the feature selection algorithm to confirm. So when our first version was released, we were surprised that we didn't have a version, and why the user feature is worse than no user feature. Besides the parameters, the hidden layer of the node, the learning rate, the results of the time window can not be expressed by the graph, and then I write PPT when very depressed, no these pictures, I how to engage in an upgrade interview, can not temporarily create a few pictures it. Even if you do not upgrade, explain to others, can not be arrogant to say, I experiment out, but the data is chaotic.       Error 4: Contempt for online engineering. Originally thought that the work of the line works is relatively few, the algorithm finishes can also, but in fact the online project also has many things. I wrote the engineering part of the code, only to find that even hash algorithm can not solve the binarize speed problem, but also only a binarize task to move to the line. and also to write some tools, data synchronization, monitoring tasks, in fact, spent a lot of time. These can be done synchronously at the time of the algorithm development. Finally, it becomes the bottleneck of the project completion time,It was two weeks late.       Error 5: External causes. After the search was sold, searching ads were merged with us, and the final owner of our conversion rate forecast was not sure. Gradually, I have lost interest in the conversion rate prediction because the project has gradually left me. I am more practical at work, can do not do things, I certainly do not do, unless the work is very interesting. In fact, this is a reason why I often advise colleagues who want to resign because of their job, usually, the change of work is bigger than what you think, and it's faster. Think of how many companies have disappeared, how many departments dissolved, how many leaders left, really can not endure it?       Error 6: Lack of support from higher-level leaders. I have never heard the leader of my team leader asking about the progress of our project. Without the support of the leader, this is not an important project, there will be resistance to push, you can not let colleagues to risk the release of the issue to send a thing he did not feel heard of the project. and other people's cooperation mood is not high, everyone listens to so many projects every day, to others, just another unintelligible project.

Conversion rate Prediction Project Failure summary