Application of DSP core algorithms and data value

fish001

Application of DSP core algorithms and data value [Copy link]

With the rapid growth of programmatic buying, DSP has developed from the conceptual stage to a market standard product service with standardized assessment targets. However, the current DSP market is mixed. When all DSPs can standardize and connect to the traffic trading platform (Exchange) to place advertisements, algorithms and data become the core competitiveness of DSPs. They essentially constitute the logical support points for DSPs to achieve better KPIs under the same resource constraints. In fact, algorithms and data are the underlying structures that DSPs cannot see. Their value is mainly reflected in three aspects: first, accurately identifying and reaching users for advertisers; second, accurately estimating the value of users; and third, selecting marketing information that impresses TA according to their demand recognition stage. Solving these three problems can make it possible to reasonably allocate advertising budgets, thereby achieving the effect of optimizing advertising placement. Scenario application of algorithms and data Usually, the effectiveness of algorithms and data in DSP is mainly reflected in the following scenarios: First, find the target user (TA) in the vast Internet resource market. Taking the maternal and infant industry as an example, there are two difficulties to be solved in the marketing process. The first is to use trusted logic to confirm that the "maternal and infant" users you find are indeed the "maternal and infant" population, and the second is the unified recognition ability of the "maternal and infant" population across different channels and media. Ideally, each tag population cluster is a continuous process in which the DSP algorithm builds a model based on the user's Internet behavior, and then verifies and iterates the optimization model through the trusted Panel library. Therefore, DSP builds a trusted demand judgment model based on complete user behavior and a trusted panel library, which is the key to building a "trusted" tag user. The most practical Panel library is the customer's first-party data, including offline membership information or conversion information collected from online advertising (in theory, the customer's purchasing user groups on various e-commerce platforms should be their own assets). For example, the YOYI tag system is divided into audiences with two different demand levels: interest tags and purchase intentions. Correspondingly, we will refer to the user's behavior characteristics on the official website to determine the different demand levels of users, and use this as one of the modeling verification criteria. At the same time, YOYI has access to the traffic of almost all trading platforms and mainstream media in the domestic market, covering a large area of user Internet behavior; on the other hand, we integrate the user's natural behavior and advertising behavior. From the user's perspective, as long as actual browsing and clicking behaviors occur, it is a manifestation of the user's interest in information, and there is no strict distinction between whether the information is natural information or advertising information. The above two points enable YOYI to fully restore the user's Internet behavior trajectory. The complete user behavior trajectory ensures that the model has complete available features and accurate feature values; the rich first-party Panel data provides a real and effective training and verification set for model training. These two points together ensure the effectiveness of the user judgment model. In actual operation, feature classification and feature extraction are performed according to all behaviors such as the scene, media, time, frequency, browsing time, search, clicking on ads, browsing ads, etc. of user behavior, and the user demand stage behavior model is trained to perform label modeling. In this way, we can make more accurate judgments on user needs and demand levels. At the same time, multi-channel unified identification of target users is another problem that technology needs to solve. At present, only giant companies with large mobile and PC users have the opportunity to complete large-scale unified ID identification of Internet PC and mobile terminals, allowing DSPs to easily overcome this problem. However, there is no large-scale unified cross-screen ID standardization service in China, so DSPs need to build their own multi-screen unified ID system. YOYI's approach is to build its own cross-screen user identification algorithm, and then use the more standard third-party cross-screen ID in the market as a training set for optimization and verification. For example, the same person on PC and mobile will have some similarities in wifi access, geographic location, behavior trajectory, and catalyst habits, which are the basis for building a unified ID algorithm. The combination of machine learning algorithms for large-scale data and artificial rules can deconstruct and understand data from many dimensions, thereby solving the problem of unclear rules and inconsistent behavior. Feature extraction and modeling of users in PC, mobile, and cross-dimensional dimensions of PC and mobile are performed to train a unified ID model, and the standard third-party cross-screen ID is used as the evaluation standard, and the accuracy rate reaches a usable level. Secondly, we need to analyze the value of the target users. Not every absolutely accurate "mother and baby" group is "just" eager to buy milk powder when encountered in every scenario. The user-level conversion rate estimation and click-through rate estimation problem is the key problem of performance advertising. Estimating the click-through rate and conversion rate of a specific advertisement for a specific user in a specific location is a typical machine learning problem of large-scale data. We have built a user feature system, an advertising feedback feature system, a traffic feature system, and a cross-feature system of various dimensions. We use the classic LR as the estimation model and GBDT as a high-dimensional feature extraction model to estimate the click-through rate and conversion rate. Both offline evaluation and online effects have good performance. After estimating the click-through rate and conversion rate, we can perform KPI-oriented CPM calculations based on marketing goals. Taking automobile Leads as an example, Ecpm = CPA * conversion rate. CPM bids are made according to different users and different traffic to achieve the best effect within a limited budget. In addition, another often overlooked but crucial strategy is the ability to prevent cheating. How to identify the validity of the data input into the model and exclude false traffic, clicks, and even conversions is a core topic that needs to be developed in a separate article. Here is a record to remind you of its importance. The conversion rate after data purification is crucial. On the one hand, it is an important factor in the screening and sorting of advertisements during delivery, which determines which advertisement is the most relevant to the user in the current scenario. On the other hand, with the authorization of the advertiser, the algorithm can automatically bid for the target user to ensure that the CPA bid limit is appropriately exceeded to track the core user group and avoid being "snatched" by competitors. Finally, after finding the person and knowing his/her focus and corresponding demand stage in the scenario, we should consider using the appropriate communication method to impress him/her. At present, due to the richness of the current advertiser's creativity, production capabilities and media review cycle, our algorithm application in intelligent creativity is still in the early stages of development, and there is still a lot of room for the explosion of the effectiveness of creative combinations. However, we have seen that some third-party creative companies in the market have made efforts to advance in this direction, and a creative revolution in Internet advertising is just around the corner. Industry-customized application of algorithms and data Taking the application of algorithms and data in the automotive industry as an example, we will explain that general algorithms and data frameworks need to be optimized through industry customization to achieve the best effect. In the actual advertising of automobiles, we found an interesting phenomenon: the time dimension of user car-related behavior is a very useful feature for the final sales leads. We analyzed that some car enthusiasts, although they have been paying attention to car forums or browsing car knowledge for a long time, have no need to buy a car in the short term, so they do not have much effect on sales leads. However, for those who have a need to buy a car in the short term, the timeliness of their behavior will be very obvious. Therefore, we differentiated the modeling factors of key industries based on the general model, so that specific explicit features in some industries can play a unique role in identifying user needs and become an effective weapon to meet the different marketing goals of advertisers. In summary, there are three aspects of the core application of algorithms and data in DSP: one is the identification of users at different stages of various interests, one is the determination of the value of users in different media scenarios, and the other is the selection of effective information displayed to users. Around these issues, the user interest and effect value algorithm system constitutes the core algorithm system of YOYI, which is continuously improved in practice to help advertisers achieve their marketing goals.