[I Technical Meeting] 2020.03.13 “Common Video Advertising Algorithm Experience Talk” sharing meeting
Today, the main content to share is divided into the following four pieces of content: IQiyi effect advertising product introduction, mainly introduces the main resources of iQiyi effect advertising and the corresponding sales form; The second is the practice of advertising sorting algorithm, which mainly introduces the fine model and coarse model used in the sorting algorithm process; The third is the practice of effect advertising business strategy. This module mainly introduces the cold start strategy for new advertising and the intelligent bid strategy and double target bid strategy for bidding. Finally, on the basis of the above three points, the summary and outlook of the effect advertising are carried out.
Effect advertising product — oCPX
In CPX placement, bidding is ranked by ECPM, which is equal to bid multiplied by PCTR. In CPX placement, advertisers encounter three pain points in the upper right corner. First, the conversion cost is not stable. Most of the advertisers of effect advertising are small and medium-sized advertisements, which mainly focus on the final conversion of users, including download, installation and payment. CPX billing mode is based on click, which cannot guarantee the subsequent conversion rate and the corresponding conversion number of advertisers, resulting in unstable conversion costs. Second, manual optimization is cumbersome. Advertisers need to screen out the optimal exposure quality in the case of CPX billing mode, so as to ensure a high conversion rate from click to follow-up. Advertisers ensure exposure quality by setting user orientation conditions, including user gender, age, level and channel, and so on
So as to achieve a good conversion effect under the CPX advertising mode. The third is the extension difficulty. In the case of CPX mode, advertisers can only carry out the operation of click bid. When the click bid increases, it can indeed achieve the effect of exposure and click number increase, but usually the side effect is the decrease of conversion rate and the increase of conversion cost.
Therefore, we try to introduce oCPX mode. The charging formula of oCPX mode is bid multiplied by PCTR multiplied by PCVR multiplied by intelligent bidding factor. Advertisers only need to set the target cost when placing oCPX, and the media side will implement intelligent placement of oCPX advertisements through the algorithm model. Next, we will show you our attempts and practices in PCTR, PCVR and intelligent bidding module in order to realize the automatic driving of advertising.
Sorting funnel logic
Before introducing the practical sorting algorithm, we will first take a look at the sorting funnel logic. When a request occurs, it goes through the following three steps to reach the AD display. The first layer is the recall link. The recall link includes directional recall and intelligent directional recall. Directional recall is mainly based on the directional conditions set by the advertiser. Directional orientation is based on breakthrough advertisers set of screen for higher similarity or forecast conversion/estimated click-through rates higher advertising order, to participate in the subsequent modules, recall elected into the coarse part, after the order of the coarse including creative optimization, lightweight clickthrough rate forecast, forecast and conversion and the corresponding cold start. The creative optimization module is mainly to select the best ideas under the same advertiser, which can realize the novelty of the platform and improve the transformation effect of the platform.
Lightweight CTR and conversion rate estimation solves the problem that a large number of ads in recall need to be refined. Lightweight model calculation is used for advertising screening logic. Cold start is to solve the problem that in the process of new advertising, the model cannot make accurate prediction because it does not get the data generated by this part of new advertising.
After entering the rough arrangement, a small number of advertisements are screened into the fine arrangement, which includes three modules: budget smoothing, click rate estimation and conversion rate estimation. Budget smoothing is mainly to solve the problem of advertisers in the process of daily budget delivery. The budget can be smoothed in every hour to solve the problem of short-term explosion caused by some advertisers due to the good effect of bid or click rate estimation, thus affecting the final cost of advertisers. Click rate estimation and conversion rate estimation use a more complex estimation model to optimize a small number of advertisements, and select TOP1 for advertising display after sorting according to ECPM. Intelligent bidding module is to cover the coarse and fine two links, intelligent bidding is mainly to ensure that advertising in the process of cost stability, the second is to be able to expand in time, so it will cover the coarse and fine two links.
An overall framework for algorithmic practice
Starting from the data stream at the top, the advertising engine produces Tracking data corresponding to the site feature after advertising. The site feature is mainly to record the log of the specific feature used in the sorting process in real time. Field characteristics and Tracking are implemented by Kafka streams and Hive tables generated by the AD data team. Part of the data is used to generate HTTP interface real-time data for the calculation of intelligent bidding parameters, and finally the intelligent bidding module is transmitted to the engine for use through the database. Feature monitoring is to monitor the coverage and fluctuation of field features to ensure the stability of online services. Feature production is mainly for the output of user portrait features, the user portrait here includes the user’s basic attributes, the user’s business interest portrait and the user’s advertising interest portrait, which will be transmitted to the delivery engine through the historical and real-time way. The feature analysis module mainly produces the data needed in model training after correlating the field features with Tracking logs. The above is the data flow module.
Finally introduced online reasoning module, put in the engine in the request arrives, will get the user portrait respectively, user vector price factors, contextual features and intelligence, and recall the information need to sort of advertising into a thick row of modules, thick row request thick row ranking with users and advertising vector product advertising screen, A small number of ads are filtered and then entered into the refinement module. Refinement transmits user characteristic data and some additional information obtained from the delivery engine to the refinement ranking. Inference of FM model is performed while advertising business data is obtained through advertising business data, or TFserving request is called to get estimated conversion rate. After the corresponding results are obtained, they are returned to the refinement module of the advertising engine to screen out the optimal advertising for advertising display.
Model training module
Including rough and fine row model training. The main derivative reason of coarse layout is that the refined layout model is relatively heavy and the online reasoning performance is poor, so coarse layout module is derived from the inability to deal with a large number of advertisements transmitted by recall. The main goal of coarse layout is reasoning efficiency. We tried to adopt FM model at the beginning of rough layout, but with the development of advertising business, FM model could not meet the needs of online reasoning efficiency, so we introduced the twin tower model as shown in the following figure.
The two-tower model produces the user vector and the creative vector respectively and realizes the efficient online reasoning by the product of the two vectors. After solving the reasoning, we need to ensure that the rough layout module can predict more accurately, mainly through three ways to improve the prediction ability of the model.
The first is multi-dimensional features, adding user features and creative features used in model training as much as possible. The second is the feature Embedding mode. We try to implement Embedding operation for multi-value user portrait features, continuity features and advertising side features to improve the ability of feature characterization. Finally, MLP logic with different structures is used for the user vector and the creative vector to improve the generalization ability of the model, so that the product of the user vector and the creative vector can well represent the CTR and conversion rate of the final estimate.
After solving the problem of prediction accuracy, we need to consider for the new AD for processing, requires model updated in a timely manner, we level by day full amount plus hours training level of creative reasoning logic, implemented to support the new advertising business, it is mainly refers to an hour regularly use logical reasoning on the right side to get the creative vector, change into are being used on the engine, This makes the coarse layout better able to support new ads.
Click through rate estimates used in the refinement module
Until now, we have carried out four major iterations in the process of iQiyi effect advertising. Respectively as follows, the first version for the logistic regression model, using commercial users portrait features, context, and features of advertising business fitting forecast clickthrough rate, but the logic in the process of back to assess because the model is relatively simple, we want to improve forecast accuracy, only through the way of characteristics of the engineering, characteristics of the project is relatively cumbersome, In order to solve this problem, we try to introduce FM model. The advantage of FM model is that low-dimensional vectors can be constructed to achieve the combination effect of feature crossing. In addition, on the basis of FM, we construct the characteristics of user advertising portrait.
After the completion of FM iteration, we encountered the problem of data and model timeliness, so we tried to introduce FM with online learning to realize real-time model update and training, and build real-time user advertising portrait features on the basis of FM, so as to improve the prediction accuracy of the model.
FM model is combined by low-dimensional vector in pairs. Subsequently, we find that the representation ability of model and feature is insufficient. Therefore, we try to introduce deep learning and use Embedding in deep learning to improve feature representation ability. Use MLP logic to improve model generalization.
The introduction of depth model is first faced with the selection of models. Here, we compare wide Deep model, DNN model and DCN model. Compared with THE DNN model, the wide deep model can improve the memory ability of the model through the left wide part, and the left part can also carry out rapid iteration. Compared with DCN model, wide Deep can achieve AUC consistent with OFFLINE evaluation with DCN, and its implementation complexity and online reasoning efficiency are far lower than DCN, so we finally choose Wide Deep model. After model selection, the optimization of wide deep model is mainly aimed at learning rate, optimizer, statements of hidden layer and the number of neural units in each hidden layer. As shown in the figure on the right, the detailed model structure will not be described here.
Conversion estimation
First consider whether you can directly reuse the Widedeep model used in CTR estimation. The answer is definitely not possible. The main difficulties in the conversion rate forecast is a transformation goal, the current iQIYI whole effect advertising platform support including download, install, focus on WeChat ID, activation, registration, etc., 2 it is to convert data, into several thousand level every day, and mainly concentrated in the registration and attention WeChat ID, less the rest into the target data, Click and transformation of three is used as the positive and negative samples, but to all ae request forecast conversion rate, resulting in the training sample distribution are biased and reasoning, is four different resources, transformation of various types corresponding conversion rate differences are large, similar to the one example, download the attention WeChat ID conversion rate can reach 5% 10%, The conversion rate of paid and submitted forms reaches 1/1000 or lower, and the four problems encountered in the conversion rate estimation process are solved in the following way.
Additional module calibration
Calibration exists because in the process of ECPM sorting, the bid multiplied PCTR and PCVR are used. PCTR and PCVR need to be consistent with CTR and conversion rate to make the ECPM relatively stable. But is it really possible to ensure that the CTR or THE CONVERSION rate is consistent with the CTR or the conversion rate during model estimation? The answer is usually no, mainly for the three reasons shown on the left. First, there is a big deviation in the estimation of new advertisements. We also mentioned before that in the process of new advertisements, due to the lack of accumulation of historical data, the estimation deviation is usually relatively large, which shows that some advertisements cannot be exposed and some advertisements are estimated at a high value. The second positive and negative sample attribution problem, take the conversion rate as an example. When the conversion rate generates training samples, due to the conversion delay, the training data set usually cannot completely cover all the converted samples, resulting in the final conversion rate we estimate is much lower than the real one. Accuracy and stability of inconsistent problem, we found that in the process of practice model in the process of forecast, usually PCTR or PCVR forecasts, the higher the flow rate, the greater the variance of the corresponding, here will be PCTR or PCVR and stability, resulting in overall model forecast is not accurate, finally introduced the calibration logic.
Effect advertising carries out the practice of business strategy
Cold start module mainly to solve the problem of exploits and Explore, but also try to solve the problem of model forecast is not accurate, inaccurate forecast model is mainly due to the real distribution of hits after bidding and there is a difference between estimated click-through rates for distribution, model training model is used to screen out the training samples, thus forming the Matthew effect.
Strategy practice in the process of the most important module – intelligent bid
According to the ECPM calculation formula, ECPM is equal to PCTR times PCVR times transformed bid times intelligent bid factor. The main objective of intelligent bidding factor is to control cost and increase consumption. When the advertiser sets the target conversion cost, when the target cost is higher than the conversion cost in the process of advertising, we increase the advertising consumption by raising the price, when the target conversion cost is lower than the actual conversion cost, we reduce the intelligent bidding factor to achieve cost control. Next we will introduce to you in order to achieve the intelligent bid refined operation to try to various processing.
Considering the cost dynamic adjustment strategy, we consider the following problem, advertisers in the process of expanding the volume of high quality flow can be radical volume, for inferior flow to minimize the opportunity to get the amount. We introduce a traffic optimization strategy. The traffic optimization strategy is mainly compared between the PCTCVR value estimated by the fine layout module and the historical PCTCVR mean value. When PCTCVR is greater than the historical mean value, we consider the flow as a high-quality flow and continue to improve the intelligent bidding factor on the basis of the previous step. When PCTCVR is lower than the average of previous times, we think it is inferior traffic to reduce the intelligent bidding factor.
Points resources control strategy, in the process of iQIYI advertising has a number of different advertising resources, different advertising resources corresponding transformation cost and conversion rate is not consistent, in order to realize intelligent channel of fine on the bid, we a split for resources, calculate different resource intelligent bid factor, So as to realize the intelligent bidding factor can carry out fine mediation of advertising, as shown in the right figure in the process of advertising when the target conversion cost is greater than the actual conversion cost, the increase of intelligent bidding factor leads to the increase of intelligent bidding price, can obtain more exposure, so as to achieve the smooth expansion of advertising.
Finally is introduced in the process of business strategy practice using double target bid logic, advertisers in the process of putting, usually focus on front-end and back-end costs, front-end usually include download and activation, etc., are relatively more of the target transformation, the back-end cost is including pay, leave, orders, etc., into less data cannot be directly model training, In order to ensure that the advertiser backend difference is large, according to the front-end conversion cost and back-end conversion cost set by the advertiser, we respectively set the calculation of its corresponding intelligent bidding factor and back-end intelligent bidding factor, back-end intelligent bidding factor is in the state of volume to choose the maximum value between the two for volume, Back-end cost does not meet/large deviation when we add the full operation of the two, the weight is determined by the back-end conversion cost, the back-end conversion number, the higher the proportion in the final intelligent bid factor, the final realization of the advertiser back-end cost deviation is small.
