Auction AD system-Dynamic features

Dynamic features

Features have multiple options for CTR Prediction, we now have three axes, ad,user,context. user has a lot of tags,cookies, age,gender is obtained from the targeted information, theURL has some domain names and analysis of the topic,AD hierarchy category, Advertiser,campaign,solution,creative,URL. Clearly, if only as a learning problem, any two-axis or three-axis combination of features can be used as a feature in the logistic regression to learn, but also to analyze the strength of each feature is how much, But it will also produce the many features mentioned above, and it is not surprising to combine billions of features.

        I'm introducing another aspect of thinking, a choice of intuitive features, for this Span lang= "en-US" >ctr prediction problem, or for some Internet problems, dynamic features are quite effective. Dynamic features are relative static features, static features for example, the above mentioned static characteristics can be obtained by combination, such as age equals user and Span lang= "en-us" >ad satisfies both conditions, the characteristic value is 1, otherwise 0. and dynamic optimization is the statistics of this combination in the history of what its performance is, that age equals 15 years old and for e-commerce ads in the history of how the click-through rate, the history of the CTR Show is a particularly strong indicator, or it is more than 1 or 0 represents more information. Dynamic characteristics are the characteristics of aggregated click-Feedback statistics as ctr predictions on the label combination dimension. Can also be understood in a different way, we can think that this aggregation dimension of the ctr is a weak classifier, such as the previous example only know age and ad type to predict the CTR. We can train thousands of weak classifiers for thousands of combinations, and this weak classifier is the input feature of the last learning model.

1.  Engineering architecture Extensibility is strong, because the generation of dynamic features, only need to make unified mining traffic on historical data, only needs to be configured to realize the mining of arbitrary features , compared with online learning, as mentioned above, any one of the features of the model corresponds to a scheme on the models, if the model quickly change weight and the feature set quickly variable features, variable features will be much simpler, Because the variable model in the advertising process also involves multiple machine communication problems, variable features do not have this problem, even if you need to calculate very fast, can also be implemented with streaming computing platform. 2.  for the new (A, u, c) combination has a strong back-off capabilities, such as there is a kind of advertising material has not appeared, If you want to add this feature in the model and update the weights, the process is more complicated. If it is only in the characteristic end of the description of it, the material to different groups of ctr as a dynamic feature in the system, as long as the flow of computing to do fast enough, as long as the ad on-line, you can quickly in the model to make the right decision. The disadvantage of the dynamic characteristics is that because the features are dynamic, it is necessary to have a large cache to store, the amount of storage is huge, because all the combinations can generate dynamic features, in addition, the speed requirements of the update is also relatively high.

I'll cite a couple of examples, such as a cookie (u) and Creative (a) that can be combined, but this group is not too much of a problem, because cookies and creative, whether you do dynamic or static features, can get too little statistical information, Again such as Gender (U) and topic (c), this combination can get more statistical information, and its conclusions are relatively stable, but not enough fine. Location (u) and advertise (a) are also statistically larger, but less granular. There are category (a) and category (U),cookie (u),Creative (a),gender (U) so characteristic or static, or dynamic, We use models to learn the most valuable features for the final decision.

Auction AD system-Dynamic features