(Don't worry about the road ahead, no one in the world knows you, I wish you all a happy Labor Day!) ）
First, the business background
The ad engine is at the heart of performance ad monetization. Its function is to find the highest value ad (or ads) from the millions of ad libraries for each ad retrieval, and give an accurate estimate of eCPM for deduction. It is not feasible to directly calculate and sort the advertising library of more than one million in terms of computing power and latency performance. To balance performance and performance, ad engines often use a cascade architecture to form funnels. The content is filtered layer by layer through the funnel, the magnitude gradually decreases, the algorithm becomes more complex, and the accuracy becomes more and more accurate.
Second, the introduction of each layer of the funnel
The funnel structure of Hima advertising engine is divided into five levels, from top to bottom, retrieval, recall, coarse, fine and rearranged.
The first level of the funnel is retrieval. It uses the entire advertising database as the candidate set to screen out the set that satisfies the targeting and has relatively high value. Because it is the first layer of primary screening facing millions of candidates, only matching algorithms with low computational complexity and high parallelization can be used. Algorithms are usually Boolean operations, such as matching traffic profiles to ad targeting, and hot and cold indexing of the plan lifecycle.
The second level of funnel is recall. The candidate set is the output of the retrieval funnel and is on the order of about 100,000. In practice, the goal of recall is to improve the relevance of the output ad set to traffic, and algorithms such as I2I and U2I are often used. Ads with high relevance can have better clicks and conversions in the future, resulting in higher advertising revenue and better user experience. Due to the abundance of relevance goals, such as contextual relevance, historical content consumption and advertising relevance, multiple recall channels are usually used in the funnel to carry different relevance goals. A lot of prior knowledge and the user's non-advertising behavior can be directly introduced into the advertising system to optimize the effect by adding recall channels. Explainable strategies are usually done at this layer.
The third level of the funnel is the coarse row. Its goal is to uniformly screen recalls, outputting a fixed-cap set of ads to reduce the amount of computation for the next fine funnel. Ideally, coarse sorting should provide consistent sorting with fine sorting and retain a relatively rich set of candidates. From the perspective of computing power, the amount of advertising entering the coarse row can already support the model to estimate and sort at the 10,000 level. In practice, rough models often use simple linear logistic regression or double-tower cache depth models.
The fourth level of funnel is fine-rowed. Given traffic requests, the goal of fine-tuning is to screen out the highest-value ads and give an estimated eCPM. Because of the dependence on deductions, it is necessary not only to sort eCPMs, but also to ensure the accuracy of the absolute value as much as possible. The highest-value ads are the result of balancing the interests of both the media and the advertiser, and balancing the short-term and long-term interests. These are reflected in eCPM estimation and bidding strategies. Fine Layout first uses deep learning to estimate the click-through rate and conversion rate. This is where algorithm optimization is the most complex and powerful. Secondly, the bidding strategy is designed to give the final bidding ranking score based on multi-faceted information. The bidding ranking score is not only affected by eCPM, but also takes into account recall relevance, crowd prior information, commercial value access, negative feedback access, and cold start long-term value.
The fifth level of the funnel is rearrangement, which breaks up similar content in multi-ad slot scenarios such as information flow to ensure diversity.
Like looking for a pearl in a vast ocean, the five-level funnel is screened through layers to obtain the most valuable Chosen One.
3. Funnel optimization practice of horizontal and vertical combination
For the practice of funnel optimization, firstly, the optimization done independently by each layer of the funnel is introduced from a horizontal perspective, including multi-objective unified coarse arrangement, system optimization, hot and cold indexing, etc. Then, from a vertical perspective, cold start is used as an example to describe how to optimize business effects through multiple layers of the funnel.
Search Funnel Optimization Practices: Simple Strategies Also Pay
Retrieval funnels are often difficult to support slightly more complex performance strategies for the full database of candidates. However, as long as you have a thorough understanding of the business, even a simple strategy can achieve good returns. Hot and cold indexes are one such example.
The life cycle phenomenon of advertising can be observed in Hima advertising. The conversion rate is low when learning from cold starts, but traffic exploration is required. As advertising matures and conversion rates increase, it continues to bring leads to advertisers and revenue to media. Then the number of ad exposures increases, and the audience content fatigue makes the ads gradually appear empty consumption with exposure but no conversion. Advertising is starting to go into a recession. To address this business phenomenon, we designed a hot and cold indexing strategy. In terms of strategy, a single index is upgraded to four new, hot, cold, and zero indexes, and the recall probability of different indexes is different, and the configuration of computing resources is different. The learning and growth plan prioritizes the occupation of computing resources and the over-funnel, and the recession plan reduces the computing resources and reduces the over-funnel. After the strategy was launched, the overall eCPM for self-operated ads was +3.0%. Retrieval performance end-to-end 99-line latency reduction of -7.5%.
In hot and cold indexes, we first extract the ad status characteristics. Secondly, the lifecycle of the ad and the index that should be placed are calculated according to the characteristics and configured rules. From business practice, we can conclude that the key characteristics of the program lifecycle include: delivery type (new acquisition/activation/lead) and conversion number. Because the target is shallow and the production of professional materials is fast, the life cycle rotation is fast. The final goal of lead advertising is often payment, and it needs to go through a long back-link such as lead acquisition and private domain operation, so the life cycle rotation will be slower.
Recall funnel optimization: multiple recalls with strong explanatory and high scalability
Enter the second layer of recall in the funnel and start focusing on the relevance of the ad to the traffic. Typically, relevance is directly proportional to performance behaviors such as user clicks. Therefore, the higher the relevance, the stronger the monetization power of the ad. At the same time, high relevance can also increase the user experience, so that the advertising content is not too abrupt with the user's interests or context.
The recall funnel optimizes relevance using a multi-way recall framework. Multiplexing refers to designing different channels for different recall targets, with each channel focusing on only one target. In this decentralized recall method, the channel and channel form a gap filling effect, which can screen out the best advertisements to the greatest extent. In addition, multi-channel recalls are also interpretable, which is convenient for business analysis and iterative updates. Finally, multi-channel recall has good scalability, and recall pathways can be added in a timely manner based on business observation and prior knowledge.
As shown in the figure below, the current multi-channel recall includes: the exploration value of cold-start recall attention ads, the relevance value of recall attention scenarios to the current playing album, and the relevance value of historical user behaviors and advertisements based on user playback behavior ItemCF recall.
Coarse Row Funnel Optimization: Unified collection sorting as a shadow of fine rowing.
The coarse funnel has the function of connecting the upper and lower levels. After the first few layers of funnel ad candidates, the candidate set has been screened to the 10,000 level, but the number is still too much. It is still not practical to calculate directly with a finely arranged complex model. Computing power cannot cope with this level of computing requests, and even if more computing power is used for parallel computing, the computing power cost will be much greater than the traffic benefit. Therefore, it is necessary to further reduce the 10,000-level material to the 100-level through the coarse funnel. From a positioning point of view, the best case scenario for coarse row is to be a shadow of fine row, to maintain consistent order with it, and to provide the richest possible set of candidates.
All-in-one, multi-business targeted advertising coarse strategy
Hima Performance Advertising also faces the unique unified ordering of advertisements with different objectives. In a funnel, there will be three advertising businesses with different goals, namely external loop effect advertising, internal circulation native advertising, and time-constrained anchor advertising. The goal of the external loop effect campaign is to simply maximize eCPM. Internal loop native ads, because their ad content is an internal streamer's album, need to take into account the experience relevance such as eCPM and context. Time-constrained streamer ads need to be based on delivery completion goals. In addition, the ecological value objective of cold start advertising needs to be considered.
For this background advertising algorithm, a unified coarse strategy with multiple objectives was designed. The strategy uses the idea of "exclusive + preferred", which not only ensures that each channel has ads into the fine arrangement, but also maximizes the commercial value of the entire queue. After the strategy is launched, from the AB data, the overall ecpm of self-operated advertising is +5%
Specifically, the unified coarse queue consists of 4 business target queues and 1 public queue, which are 1) cold start, 2) external circulation effect, 3) internal circulation native, 4) time-constrained anchor advertising, and 5) shared queue. In the business target queue, the score is calculated and sorted according to the independent coarse order.
Business-led, growing coarse model
The basis of the coarse strategy is model estimation. For the coarse arrangement model, driven by business scenarios, the advertising algorithm has also done three rounds of iterations. The model ranges from simple linear logistic regression to LTR single-objective two-tower model, and then to multi-objective multi-twin towers with stronger explanatory and more goal-focused approach.
Linear logistic regression modeling and inference are simple and suitable for quick implementation in the early stage of business. With the development of advertising business, the number of advertising spaces is increasing, and the traffic is complex, the algorithm designs a double-tower model of single-target LTR. At this time, the characteristics of deep learning with strong fitting ability, double-tower material vector caching to save computing power, and single-target good maintenance can solve business problems well. With the emergence of new forms of services such as native internal loop and time-constrained anchor advertising in business scenarios, single-target LTR gradually cannot match business needs. At this time, a more detailed CTR and CVR independent target twin tower model is needed.
Fine funnel optimization: the most direct effect and the most complex algorithm
Fine-funnels are closest to ad exposure, so optimizations have the most impact on performance. These include not only common CTR/CVR estimation models, but also system optimization with Hima characteristics. In general, fine-tuning is optimized around two aspects: more accurate eCPM estimation and a more balanced bidding ranking strategy. The two synergize with each other to improve eCPM and ecological health for long-term revenue growth from a media perspective, and ensure cost and maximize volume from an advertiser's perspective.
Model optimization has been covered in previous articles. This section describes the optimization of bidding sorting strategies, including admission, sorting, and optimal budget allocation. First of all, in terms of access, due to the importance of user experience, there is independent negative feedback access to intercept potential high-risk advertisements in real time. In addition, there is also commercial value access screening to improve the value of market traffic and reduce the proportion of advertisers without conversion waste. Secondly, for the price reduction ranking, on the basis of the core CTR and CVR estimates, the bidding ranking is carried out through strategies such as double bidding and crowd premium. Finally, there is the system optimization with the characteristics of Hima, which helps advertisers dynamically allocate and select the budget of advertising space. Let's explain the system optimization a little.
System Optimization is a feature for B-end advertisers. Since Hima has dozens of advertising slots, there are large differences between individuals (click-through rate, people reached, etc.), and it is difficult and costly for advertisers to manually select advertising slots. Under the condition of cost protection, the system will automatically select the ad space instead of the advertiser. After the feature went live, it helped advertisers increase their spend by +33%. At the same time, the media bidding filling was improved, which led to the growth of eCPM, and the number of accounts in a single auction was +133.3%.
The strategy is shown in the diagram above, and the maximum set of deliveries is determined by the creative. Motivate advertisers to make up for it through operational strategies on a daily basis. ROI is the core feedback signal and is defined as: ROI = (Planned Conversion Bid * Number of Placements Conversions) / (Actual Charge). The higher the ROI, the lower the risk of over-cost. First, split the ad slot into two subsets according to ROI: main investment S, ROI greater than 1, and exploration A, ROI less than 1. And calculate the upper limit of the corresponding period, so that S and A can achieve the overall planned ROI of >1 at a uniform cost with each other.
Collaborative optimization through multiple layers of funnel longitudinally - cold start
In the practice of funnel optimization, in addition to the horizontal iteration within the upper layer, we also carried out multi-layer funnel collaborative optimization in the cold start. Cold start refers to the advertising engine to help new advertising content to explore traffic through certain support means. For publishers, it brings new ad revenue increments, enriches ad diversity and increases eCPM. For advertisers, the volume of new advertising is accelerated, reducing the cost of capital and time for exploration.
Under the characteristics of Hima Media Traffic, the functional positioning of cold start makes it face the following challenges. First of all, due to the insufficient exposure of new ads, the models and statistical strategies based on posterior exposure data are not very accurate, which makes it more difficult to get exposure opportunities through layers of funnels. The second is the E&E efficiency problem, which inevitably has the short-term loss of traffic value when exploring in cold starts. Exploration has a cost, but it needs to be manageable. Finally, due to the relatively small volume of Hima advertising, advertisers have higher requirements for traffic quality, and the in-depth effect of advertising (such as payment, credit, etc.) during the cold start period affects the subsequent budget increase of advertisers. It can be seen that the cold start problem is systemic and cannot be solved by a single layer of funnel. In practice, we carried out multi-layer funnel joint optimization according to the sorted out problems.
First, a "new index" is designed in the first level of the retrieval funnel, Icon 1, where cold-start ads that meet the rules are placed. Ads retrieved from this index enter a separate channel that connects multiple layers and does not compete with other ads, Icon 2, thus ensuring that a certain number of cold-start ads enter the fine funnel.
Then, to control the cost of exploration, we pre-allocate advertisers with a two-tier quota of plan and impression, Figure 3. If there are multiple new plans in a single account, the prior quality score of the plan will be calculated and the TopN will be selected as a cold start plan. A single Cold Start plan has an impression quota, and ads will only be placed in the new index if the number of impressions is less than the allocated quota. After the introduction of quota control, the traffic utilization rate has been improved, and the overall ECPM of self-operated advertising is +6%.
Next, in the recall funnel, we use the platform delivery data and the external cooperation data collected to prioritize the recall from the high-quality people mined offline according to the advertising SKU, Figure 4. This increases the concentration of the target group in the exposure of cold-start advertising and optimizes the in-depth effect. The offline crowd is constantly mining and iterating. The high-quality crowd recall results in a cold start program CTCVR +16.7%.
Finally, the fine funnel will be pre-sorted and the top 1 ads will be supported and weighted, icon 5. In pre-sorting, the algorithm performs pCVR admission to ensure the quality of cold-start traffic. At the same time, the pre-sorting and sorting branch introduces the quota usage speed as the traffic shortage factor, so that the plans that are too slow are ranked first, and the cold start plans that are too fast are ranked later, so as to maintain the uniform speed of cold start. Finally, the weighting coefficient is calculated based on the advertising type, estimated CTR, CVR and recall, Figure 6. Ensure that ads that are more compatible with traffic are prioritized to achieve E&E balance.
postscript
Looking back at the above funnel optimization, it can be seen that in the scenario of high QPS and low latency of ad requests, the hierarchical funnel architecture is the optimal solution to balance the use of computing power and business effect in the face of millions of ad libraries. From the perspective of complex system optimization, funnel optimization reasonably splits the ad engine and provides more optimization grippers. This allows the algorithm to have more entry points for optimization.
While optimization in the fine funnel is most direct and effective for improving ad performance, the optimization of advertising algorithms does not stop there. Starting from what you want from the refinement, reasonable business sorting and optimization of the first few levels of funnel can also bring about the improvement of business results.
In the practice of funnel optimization, it is also observed that the local optimal is inconsistent with the global optimum. Each layer of the funnel is optimized to iterate on the local goals it sets. However, due to factors such as the locality of information, computing power limitations, and the error of a single model, if the goal is to be completely local optimal, the overall eCPM effect will be flat or even negative. In general, maintaining a certain diversity is one of the solutions. For example, if you always choose the ad with the highest coarse score in the coarse funnel, the same advertiser will occupy the majority of the fine auction queue. By increasing diversity, this problem was solved and eCPM improved.
作者:at.liu
Source-WeChat public account: Himalaya technical team
Source: https://mp.weixin.qq.com/s/txQZj1crDigg3d9Jlf_wQw