AFA: Auto-tuning Filters for Ads Joobin Gharibshah, Mahmuda Rahman and Abraham Bagherjeiran San Jose, CA, USA Abstract Tuning filters to refine Ads eligibility to surface in search results emerges as a pivotal problem. It often necessitates a nuanced approach to cater to diverse requirements from the customers. Adjusting these filters must judiciously balance the preferences of both advertisers and users in the online marketplace. Hence, it requires a multi-objective optimization which often turns out to be hard due to the conflicting nature of the objectives from these customers. In this paper we present AFA: Auto-tuning Filters for Ads - a novel application of Bayesian Optimization for auto-tuning these filters. We specifically develop AFA to employ a probabilistic model to navigate the intricate trade-offs between multiple objectives. It iterates over a feasible solution space and quickly converges to an operating point which ensures showing well performing ads while increasing their scale. This offers a substantial advancement in the automation for digital advertising campaigns. Our approach significantly reduces the reliance on manual adjustments and expensive A/B testing, as demonstrated by empirical results from a large-scale e-commerce platform. Keywords Bayesian Optimization, Surrogate Model, Acquisition Function, Automated Pipeline 1. Introduction Sponsored search is an advertising model used by search engines to display paid advertisements alongside organic search results. When a user enters a query, the search en- gine runs an auction among advertisers who have bid on keywords relevant to the user’s search terms. The win- ners of this auction have their ads displayed in prominent positions on the search results page, typically marked as Figure 1: A visual overview of AFA pipeline to auto tune the ”Sponsored” or ”Ad” to distinguish them from other content. quality filter: (1) Observed data points consist of thresholds and While these advertisements (a.k.a. ads) benefit advertisers metrics lift (set of values for each objective function on that given by increasing the visibility of their products to potential threshold) denoted as < 𝑇 , 𝑃𝑒𝑟𝑓 𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝐿𝑖𝑓 𝑡(𝑇 ), 𝑆𝑐𝑎𝑙𝑒𝐿𝑖𝑓 𝑡(𝑇 ) > users, maintaining user satisfaction is also important for the respectively, initial data points collected based on random thresh- olds from a feasible solution space. (2) Optimizer fits surrogate e-commerce platform to monetize the search from showing function on the observed data points and produce 𝑇 ∗ as the next ads. Thereby, it is crucial to ensure that the participating threshold (T*) to be evaluated (3) Evaluator measures the im- ads are of high quality to comply with user’s desirability. pact of the new threshold 𝑇 ∗ on the objective value and produce Generally, ad platform employs a variety of quality filters another triple < 𝑇 ∗ , 𝑃𝑒𝑟𝑓 𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝐿𝑖𝑓 𝑡(𝑇 ∗ ), 𝑆𝑐𝑎𝑙𝑒𝐿𝑖𝑓 𝑡(𝑇 ∗ ) > to be controlled by hyperparameters 1 . Its core purpose is to appended to the stored observed data points so that we can ex- efficiently navigate through extensive datasets, find content plore next point based on that that aligns closely with the user’s search intent and personal preferences by providing more relevant outcomes. The paper presents a innovative approach for automat- for paying more to the platform. As they often prioritize vis- ically tuning quality filters in the search monetization do- ibility over the quality of their product, the platform needs main using Bayesian Optimization. The approach aims to to maintain a quality filter that strikes a balance between optimize the quality threshold to balance scale improvement their need for exposure for the product (measured by scale) and ad performance. An automated pipeline was developed and its relevance to the user’s preferences. to minimize human effort, time, and errors in this process. 1.1.2. User’s preferences 1.1. Challenges The e-commerce platform is also committed to meeting the E-commerce platforms face the challenge of setting the right essential needs of user who uses this platform to find high quality threshold for ads while ensuring sufficient ad expo- quality product based on their search queries as input. It is sure to increase scale. This issue arises from the differing the platform’s responsibility to provide them with relevant needs of two main stakeholders: advertiser and user and high quality search results for their queries. To achieve this, an appropriate quality filter is required to surface prod- 1.1.1. Advertiser’s preferences ucts in the search results that are most relevant to the user’s Advertisers enlist their items on e-commerce marketplace to query. User’s preference is calculated as the ratio of the sell and expect greater visibility of their product in exchange click over number of impressions they have (to measure filter’s performance). This dilemma between the preferences of users and ad- AdKDD ’24, August 26, 2024, Barcelona, Spain vertisers presents a challenging problem in e-commerce Envelope-Open jgharibshah@ebay.com (J. Gharibshah); mahrahman@ebay.com (M. Rahman); abagherjeiran@ebay.com (A. Bagherjeiran) platforms. On one hand, they need to increase the quality Orcid 0000-0003-2264-8097 (J. Gharibshah) threshold to meet users’ expectations of obtaining more © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). relevant items. On the other hand, they must keep the 1 https://support.google.com/google-ads/answer/6167130?hl=en quality filter to a degree where the platform can accommo- accessed on 07/11/2024 CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings date advertisers’ appetite for increased visibility. Therefore 1.3.1. Initializer they aim to enhance ad scales without compromising the is a process to collect and store data points observed over performance. the course of running our system. We store a triplet (𝑇 , 𝑉𝑂𝑏𝑗1 (𝑇 ), 𝑉𝑂𝑏𝑗2 (𝑇 )) for each filter threshold 𝑇 that we ex- 1.2. Motivation plore. This triplet has a threshold along with two corre- The dynamic nature of the marketplace, characterized by sponding objective values, i.e., scale lift and performance continuous growth, frequent updates, and new feature roll lift. The pipeline starts with some initial data points which outs, further complicates the task by rendering the static are collected via running A/B tests over various thresholds thresholds obsolete in short order. Manual tuning of filters within the exploration range to give the process a warm is not only time-intensive but also lacks efficiency, under- start. scoring the need for an automated, scalable solution. Main inspirations for this work are: 1.3.2. Optimizer utilizes a Bayesian optimizer to solve multiple objectives 1.2.1. Need for a Formal Definition consisting of scale and performance. This optimizer will Capturing and quantifying the intricate correlations be- read initial data points i.e (𝑇 , 𝑉𝑂𝑏𝑗1 (𝑇 ), 𝑉𝑂𝑏𝑗2 (𝑇 )) and build a tween two different objectives (Scale and Performance), surrogate model based on those data points. Then it utilizes which often exhibit complex interactions, is a hard task. an acquisition function [2] to suggest a new threshold 𝑇 ∗ The absence of a formal function to encompass both the which is the next optimal point to be evaluated by Evaluator. objectives, exacerbates this issue and hinders the system- atic exploration of the trade-offs inherent to the objectives. 1.3.3. Evaluator Consequently, without a clear mathematical framework to is the component which runs an A/B test to evaluate the navigate the multi-dimensional objective space, manual op- objective function with respect to the suggested threshold. timization becomes not only cumbersome but also prone to Then, we compute a new triple as (𝑇 ∗ , 𝑉𝑂𝑏𝑗1 (𝑇 ∗ ), 𝑉𝑂𝑏𝑗2 (𝑇 ∗ )) sub-optimal decision-making, as it relies on intuition rather which consists of the suggested threshold 𝑇 ∗ by the opti- than analytical precision. This emphasize the necessity for mizer and corresponding objectives values evaluated by the a structured approach that can effectively balance these evaluator. This newly explored data point is then feedback competing goals and facilitate the discovery of an optimal to the initilizer. solution that results in a desired outcomes. Thus, the pipeline constantly updates the quality thresh- old to respond to new changes, ensuring that search results 1.2.2. Opportunity Cost remain relevant. As a common practice, multiple experiments covering var- ious operating points and parameters via grid or random 1.4. Contributions search [1] are utilized to find a proper thresholds for fil- This paper presents several key contributions: ters. This manual process is expensive and time-consuming which often needs to be repeated whenever there is a change • We formalize the process of general filter tuning in the environment. On an e-commerce platform, the vast problem by leveraging a Bayesian optimization number of advertisements and transactions necessitates that method in AFA, considering the need of multiple modifications are made with accuracy and speed to stay at- competing stakeholders. tuned in the market. As the customer tastes and industry • AFA provides a fast-converging approach that sig- tendencies shift quickly, it leaves only a brief time window nificantly reduces human effort and time for tuning to detect and react to these developments successfully. De- quality threshold. layed tuning can lead to outdated search results, diminishing • AFA has been successfully deployed within an e- user experience and potentially leading to a loss in sales, commerce platform, demonstrating its effectiveness customer trust and platform’s reputation. in refining quality filters for the search engine and attesting to its scalability in large-scale industry set- 1.3. System Overview tings. Our proposed system employs a probabilistic model that cap- The structure of the paper is as follows: Section 2 de- tures the complex relationship between the quality thresh- scribes implementation details in the optimization process old and the multi-objective function, which includes both and evaluation metrics. Section 3 analyzes the results of performance and scale lift. By utilizing Bayesian optimiza- AFA. Section 4 reviews related literature. Section 5 con- tion, we iteratively update the quality filter settings, effi- cludes with a summary and future research directions. ciently navigating the search space to identify the optimal threshold that satisfies the dual objectives as demonstrated in Figure 1. This approach not only accounts for the inher- 2. Implementation Details ent uncertainty in user behavior but also reduces the need for extensive manual tuning and A/B testing. In the pipeline presented in Figure 1, there are three com- To this end, we introduce AFA as an auto filter tuning ponents as we explained in section 1.3. In this section, we pipeline for ads to tune the quality filter. AFA consists of focus on the Optimizer and provide more details regarding three major steps that we illustrated in the figure 1 and we its implementation and evaluation. introduce them here. 2.1. Optimization Process Bayes’ Theorem is formulated as follows: In this section, we describe our objective function and its 𝑃(data|𝜃)𝑃(𝜃) formulation. 𝑃(𝜃|data) = (6) 𝑃(data) Objective function In AFA, we aim to find a threshold value for quality filter that satisfies two objectives related to where: 𝑃(𝜃) is the prior probability of the hypothesis be- advertisers and users needs: maintaining performance lift fore seeing the data. 𝑃(data|𝜃) is the likelihood of the data and a positive scale lift in ads. The lift amounts calculated under the hypothesis. 𝑃(data) is the marginal likelihood for this purpose are all relative differences between the or evidence, the probability of the data under all possible suggested value and the current value in the system as AFA hypotheses. 𝑃(𝜃|data) is the posterior probability of the hy- updates the threshold in each iteration. (more detail in pothesis after seeing the data. In our case, 𝜃 is representing section 3 and equation 14). the threshold We formulate these objectives as follows: Based on this Bayes’ Theorem, we build a surrogate model Increasing scale: AFA looks for quality thresholds which will rely on the observed data points (thresholds) and which increase the ad scale in search result by: make a prediction for unobserved data points. The surrogate model is as follows: 𝑡 ̄ ∈ 𝑎𝑟𝑔𝑚𝑎𝑥 𝛿𝑠(𝑡) (1) a) Surrogate model is a probabilistic model used to ap- 𝑡∈Thresholds proximate the unknown objective function 𝑓 (𝑥) that we where 𝛿𝑠(𝑡) represents ad scale changes. wish to optimize. The surrogate model, denoted as 𝑓 (𝑥), ̂ is Maintaining performance: AFA also looks for that used to predict the output of 𝑓 (𝑥) given new inputs 𝑥, and quality thresholds to minimize performance lift via: to estimate the uncertainty of that prediction. The surrogate model we used is a Gaussian Process (GP), 𝑡 ̄ ∈ 𝑎𝑟𝑔𝑚𝑎𝑥 − |𝛿𝑝(𝑡) | (2) which is defined by a mean function 𝜇(𝑥) and a covariance 𝑡∈Thresholds function (kernel) 𝑘(𝑥, 𝑥 ′ ). The GP surrogate model for any subject to the constraint |𝛿𝑝(𝑡) | < 𝐶. where 𝛿𝑝(𝑡) represents input point 𝑥 is: performance changes. ̂ 𝑓 (𝑥) ∼ 𝒢 𝒫 (𝜇(𝑥), 𝑘(𝑥, 𝑥 ′ )) (7) We combine these two objective in one as follows: This formulation allows the Bayesian optimization algo- 𝑡 ̄ = argmax (𝛿𝑠(𝑡) − |𝛿𝑝(𝑡) |) (3) rithm to not only predict the function value at unobserved 𝑡∈Thresholds points but also quantify the prediction uncertainty, which subject to the constraint |𝛿𝑝(𝑡) | < 𝐶. In this function, we is crucial for balancing exploration and exploitation during apply the same weight to both objectives, although these the optimization process. In our GP, we used a White Noise weights may vary in different scenarios. as a kernel function in Bayesian optimization[3], This kernel To impose the constraint 𝐶 on the objective function we function is defined to represent the idea that observations employ a penalty and reward mechanism. We penalize and have some amount of uncorrelated noise. reward the objective function when the explored thresholds The White Noise kernel is defined as: fails and succeeds to maintain the constraint respectively. Thereby, we formulate the optimization problem as follows: 𝑘(𝑥, 𝑥 ′ ) = 𝜎𝑛2 Δ(𝑥, 𝑥 ′ ) (8) where 𝜎𝑛2 is the noise variance, a hyper-parameter that repre- 𝑓 (𝑡) = 𝑡 ̄ = 𝑎𝑟𝑔𝑚𝑎𝑥 (𝛿𝑠(𝑡) − 𝛿𝑝 ∗ (𝑡) ) (4) 𝑡∈Thresholds sents the variance of the noise in the observations. Δ(𝑥, 𝑥 ′ ) is the Kronecker delta function, which equals 1 if ( x = x’ where the modified performance will be defined as: ) (i.e., the points are identical) and 0 otherwise [3]. The White Noise kernel is just one possible choice among many kernels for GPs 𝑅(𝑡) ∗ 𝛿𝑝(𝑡) −𝐶 ≤ 𝛿𝑝(𝑡) < 𝐶 𝛿𝑝 ∗ (𝑡) = { (5) b) Acquisition function is a function that guides the 𝑃(𝑡) ∗ 𝛿𝑝(𝑡) 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 optimization process by determining where to sample next. The acquisition function balances exploration of the search Here, 𝑃(𝑡) and 𝑅(𝑡) are the amount of penalty and reward space (sampling where the model is uncertain) with exploita- we apply on the 𝑃𝑒𝑟𝑓 𝑜𝑟𝑚𝑎𝑛𝑐𝑒𝐿𝑖𝑓 𝑡. tion (sampling where the model predicts high performance). This generic formulation can accommodate additional ob- Two common acquisition functions which we used in jectives and constraints, as well as custom-defined penalty AFA are: and reward values, according to business requirements. Al- 1) Expected Improvement (EI): This function measures though we have defined the constraint in closed-form, we the expected amount of improvement over the current best have not devised a closed-form definition for the objec- observation 𝑓 (𝑥 + ) at a new point 𝑥. tive components that would allow us to use the Lagrange multiplier method. However, we discuss related work and 𝐸𝐼 (𝑥) = 𝔼 [max(𝑓 (𝑥) − 𝑓 (𝑥 + ), 0)] (9) alternative approaches in Section 4. In our pipeline, at each iteration, the Optimizer solves where 𝑓 (𝑥) is the objective function and 𝑓 (𝑥 + ) is the best for the objective function defined in Equation 4 by using a observed value so far. Bayesian optimization approach, as we describe below. 2) Probability of Improvement (PI): This function mea- Bayesian Optimization employs a probabilistic model sures the probability that sampling at a new point 𝑥 will to represent the uncertainty about the objective function’s lead to an improvement over the current best observation behavior, and it updates this model iteratively using Bayes’ 𝑓 (𝑥 + ). Theorem as new data points are observed. 𝑃𝐼 (𝑥) = 𝑃(𝑓 (𝑥) > 𝑓 (𝑥 + )) (10) The selected acquisition function will generate a new One of the standout results of using AFA for automatic threshold, which will be passed to the evaluator for testing filter tuning is its marked efficiency in both time and human in the next step. To ensure fast convergence we adopted effort. In a comparative analysis between manual tuning PI . We explain the evaluation process in Section 2.2. and AFA-assisted tuning, we observed a significant reduc- tion in the number of iterations required to optimize the 2.2. Evaluation Metrics quality filter. Manual efforts necessitated 12 iterations to achieve satisfactory results, whereas AFA achieves compa- We need to evaluate the optimizer’s efficiency using the rable improvements in the ad scale metric and maintained Evaluator through an A/B test. We execute this phase of performance within just 3 iterations. the pipeline at the conclusion of each iteration, which, in our Moreover, the total number of data points needed for case, spans a period of six days. We set the recommended AFA was halved, with only 6 data points (including 3 initial threshold by AFA as a treatment in an A/B test and get the data points) compared to the 12 required for manual tuning. real-time impact on both the scale and performance since This reduction in data points translates directly into savings metrics like performance are dependent to user behaviour extensive AB testing spanning for multiple weeks as well and is not possible to be computed offline precisely. The as post test analysis efforts. AFA autonomously computes objectives calculated by AFA are: the next threshold to test, minimizing the need for manual Scale is calculated as the ratio of the total number of evaluation. times ads are displayed to the number of qualified queries Time efficiency was also significantly improved. Each (i.e. buyer searches) that trigger the ads, expressed by the data point in the AFA corresponded to outcome of one week formula: of experimentation using our A/B testing platform, leading 𝑁 𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡𝑖𝑚𝑒𝑠 𝐴𝑑 𝐷𝑖𝑠𝑝𝑙𝑎𝑦𝑒𝑑 to an optimization timeline of just 4 weeks. This consisted 𝑆𝑐𝑎𝑙𝑒 = . (11) of a week of initial data collection followed by 3 weeks of 𝑁 𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑄𝑢𝑎𝑙𝑖𝑓 𝑖𝑒𝑑 𝑄𝑢𝑒𝑟𝑖𝑒𝑠 iterations. Conversely, the manual approach spanned over Performance is measured as the proportion of clicks an 12 weeks, with each iteration taking one week - yet resulting advertisement receives relative to the number of times it is in a sub-optimal solution for the problem. shown (impressions), represented by the formula: In summary, the deployment of AFA for quality filter 𝑇 𝑜𝑡𝑎𝑙 𝐶𝑙𝑖𝑐𝑘𝑠 tuning within our experimental framework yielded a 4X 𝑃𝑒𝑟𝑓 𝑜𝑟𝑚𝑎𝑛𝑐𝑒 = . (12) increase in efficiency in terms of both time and human 𝑇 𝑜𝑡𝑎𝑙 𝐼 𝑚𝑝𝑟𝑒𝑠𝑠𝑖𝑜𝑛𝑠 effort. This demonstrates AFA’s potential to accelerate the We introduce quality and revenue here as additional out- tuning process, ultimately reducing the laborious and time- put metrics in order to track business impacts. intensive nature of manual filter tuning. Revenue is the income earned from displaying ads on a platform. It can be calculated as: 3.2. Business Targets 𝑅𝑒𝑣𝑒𝑛𝑢𝑒 = Buyers’ action (i.e clicks) × Revenue per action In this section, we present the amount of improvement (13) achieved by employing AFA in our real world production Quality is the indicator of the relevance of the ads with platform, targeting a large online user population over a respect to query. two-week experiment period. The amount of lift in an A/B test is calculated by compar- We compared two thresholds: the recommended thresh- ing the evaluation metric (e.g., scale) between the treatment old from AFA, which converged after three iterations via group (T), 𝑀𝑒𝑡𝑟𝑖𝑐𝑇 , and the control group (C), 𝑀𝑒𝑡𝑟𝑖𝑐𝐶 , using an automated pipeline, and the best threshold suggested the formula by manual tuning, which required human adjustment over (𝑀𝑒𝑡𝑟𝑖𝑐𝑇 − 𝑀𝑒𝑡𝑟𝑖𝑐𝐶 ) twelve iterations. As indicated in Table 1, both methods 𝐿𝑖𝑓 𝑡 = × 100% (14) improved the scale, but AFA showed a higher impact on busi- 𝑀𝑒𝑡𝑟𝑖𝑐𝐶 ness metrics. AFA not only improves the quality score 80% The amount of lifts will be used by the Optimizer and also but also we achieved a 0.72% higher performance compared stored in the table introduced at Section 2 for the ongoing to manual tuning, resulting in a mere 0.18% performance exploration. loss when using AFA, as opposed to a 0.66% performance loss with the manual approach. Considering that perfor- 3. Experiment Results mance was our constraint during the optimization phase, a smaller loss signifies a better operating point. Moreover, we Analyzing the outcomes of filter tuning to meet multiple observed a 44% improvement in ad revenue with AFA; man- objectives simultaneously, is not a trivial task. It requires ual tuning resulted in a 0.22% loss in revenue, while AFA a deep and precise understanding of the interplay between managed a 0.01% gain. We also observed a startling gain different objectives. However, in our algorithm, we have of 80% quality as well. In summary, our approach enabled formalized this relationship in a robust manner, allowing the display of more ads without sacrificing performance. measurable and accurate improvements. In this section, we Furthermore, it improved quality and revenue which are discuss our experimental results. key business metrics. In Figure 3, we provide an empirical example illustrating the search results page for a buyer query ”leather jacket”. 3.1. Execution Efficiency Using the threshold recommended by AFA, we were able Our experiments demonstrate that AFA could achieve the to display more sponsored ads compared to the threshold defined objectives within three iterations through our tun- determined by manual tuning. Specifically, AFA enabled ing pipeline. Figure 2 illustrates how the model evolved the presentation of four sponsored ads, whereas the manual over multiple iterations. tuning approach yielded only two. This empirical evidence (a) Surrogate model and acquisition function after (b) Surrogate model and acquisition function after (c) Surrogate model and acquisition function after feeding initial data points second iteration third iteration Figure 2: Changes in the surrogate model and acquisition function over three iterations. The dashed green line represents the surrogate model fitted to the red dot observations. The green shade indicates the uncertainty for each threshold based on the surrogate function. The blue line represents the acquisition function, and the blue dot marks the next point suggested for testing, based on the maximum value of the acquisition function that optimizes the defined objective. (a) An example of items retrieved for a query using suggested threshold (b) An example of items retrieved for a query using suggested threshold by AFA by a manual approach Figure 3: Comparing the retrieved items for a query ”leather jacket” in our search engine using thresholds suggested by AFA and the manual tuning approach, we observe notable differences. Figure 3(a) demonstrates that using the threshold recommended by AFA, we retrieved four high-quality sponsored items. In contrast, Figure 3(b) shows that with the manually tuned threshold, only two sponsored items were retrieved. Table 1 4. Related Work Comparing business metrics lift results between manual filter tuning vs AFA tuning approach There is limited research on automatic filter tuning in ads Method Scale Performance Quality Revenue industry, and even fewer studies on pipelines to facilitate Manual Tuning 0.79% -0.66% 0.05% -0.22% the process. Prior works can be categorized in the following AFA 0.88% -0.18% 0.09% 0.01% categories. AFA vs Manual Tuning 11% 72% 80% 44.0% Black box optimization (BBO) methods are essential for optimizing functions without closed-form expressions, which are often encountered in real-world scenarios involv- supports our claim that AFA fine-tunes the threshold more ing complex systems. Evolutionary algorithms (EAs) and ge- precisely than manual methods, thereby improving ad scale. netic algorithms (GAs), such as those described by [4], have Additionally, we observed that all sponsored items were rele- been fundamental in exploring search spaces in a gradient- vant and of high quality, which correlates with the improved free manner. While effective, these methods can require performance and ad revenue as detailed in Table 1. a prohibitively large number of evaluations to converge, which is not always practical [5]. To overcome the limitations of traditional EAs and GAs, surrogate-based optimization (SBO) techniques have been domain and potentially in other fields. developed. Bayesian optimization was introduced by [2] Looking forward, our future plans involve (a) extending as Efficient Global Optimization (EGO), utilizing surrogate the application of the proposed pipeline to other facets of models to approximate objective functions. Bayesian Opti- our e-commerce platform and (b) incorporating additional mization (BO), a subset of SBO, has gained traction for its dimensions into the optimization approach, thereby broad- sample efficiency and effectiveness in noisy evaluations, as ening the scope and impact of AFA. highlighted by [6]. The incorporation of Bayesian optimization into multi- objective optimization (MOO) has been an area of active References research. [7] presented a framework for efficiently optimiz- [1] J. Bergstra, Y. Bengio, Random search for hyper- ing black-box functions with multiple objectives. Bayesian parameter optimization, Journal of Machine Learning optimization and MOO used to improve their recommenda- Research 13 (2012) 281–305. tion models in feeds and notification [8]. [2] D. R. Jones, M. Schonlau, W. J. Welch, Efficient global Multi-objective optimization (MOO) addresses com- optimization of expensive black-box functions, Journal plex problems where multiple, often conflicting, objectives of Global Optimization 13 (1998) 455–492. URL: https: must be optimized simultaneously. Researchers like [9] //api.semanticscholar.org/CorpusID:263864014. have significantly contributed to this field with algorithms [3] C. E. Rasmussen, C. K. I. Williams, Gaussian Processes such as NSGA-II, which efficiently guide the search towards for Machine Learning, The MIT Press, 2006. Pareto-optimal solutions under constraints such as limited [4] D. E. Goldberg, J. H. Holland, Genetic algorithms and evaluations. The extension of BBO principles to MOO has machine learning, Machine Learning 3 (1988) 95–99. enabled the application of these techniques in various do- URL: https://api.semanticscholar.org/CorpusID: mains, including e-commerce, where balancing trade-offs is 2043246. crucial. [5] M. Jaderberg, V. Dalibard, S. Osindero, W. M. Recent work in MOO has focused on improving the ef- Czarnecki, J. Donahue, A. Razavi, O. Vinyals, ficiency and scalability of these algorithms. For instance, T. Green, I. Dunning, K. Simonyan, C. Fernando, [10] offers a robust approach for handling many-objective K. Kavukcuoglu, Population based training of neural problems by introducing NSGA-III. Additionally, the inte- networks, 2017. arXiv:1711.09846 . gration of machine learning models, as seen in work on the [6] E. Brochu, V. M. Cora, N. de Freitas, A tutorial on SPEA2 algorithm, has improved the handling of complex bayesian optimization of expensive cost functions, objective landscapes [11]. However, most of these works with application to active user modeling and hierar- focus on the closed form of objective functions. chical reinforcement learning, ArXiv abs/1012.2599 Hyper-parameter tuning: Hyper-parameter tuning is (2010). URL: https://api.semanticscholar.org/CorpusID: a critical step in machine learning that involves selecting 1640103. the optimal set of hyper-parameter for a learning algorithm [7] S. Daulton, M. Balandat, E. Bakshy, Differentiable ex- to maximize its performance. Hyper-parameter are the con- pected hypervolume improvement for parallel multi- figuration settings used to structure the learning process, as objective bayesian optimization, in: Advances in Neu- opposed to model parameters that are learned from the data. ral Information Processing Systems, 2020. There are several work focusing on search mechanism like [8] LinkedIn Engineering, Using bayesian optimization grid and random search which are the simplest and most for balancing metrics in recommendation systems, commonly used approaches [1]. There are other efforts on LinkedIn Engineering Blog, n.d. URL: https://www.li Gradient-Based Optimization which uses gradient informa- nkedin.com/blog/engineering/recommendations/usin tion to guide the search for optimal hyper-parameters [12]. g-bayesian-optimization-for-balancing-metrics-in-r There are bandit based approaches which dynamically allo- ecommendat, [Online; accessed 20-April-2023]. cates resources to a set of hyper-parameter configurations [9] K. Deb, S. Agrawal, A. Pratap, T. Meyarivan, A fast and and rapidly eliminates poor-performing options [13]. elitist multiobjective genetic algorithm: Nsga-ii, IEEE Our research introduces an automated pipeline that ap- Transactions on Evolutionary Computation 6 (2002) plies new advancements to automatically tune thresholds 182–197. for advertising systems, a problem not previously solved. [10] H. Seada, K. Deb, U-nsga-iii : A unified evolutionary It combines Bayesian Belief Optimization (BBO) and Multi- algorithm for single , multiple , and many-objective Objective Optimization (MOO) in a novel way, providing a optimization, 2014. URL: https://api.semanticscholar. practical solution to a real-world challenge. org/CorpusID:9620515. [11] E. Zitzler, M. Laumanns, L. Thiele, Spea2: Improving 5. Conclusion the strength pareto evolutionary algorithm, 2001. URL: https://api.semanticscholar.org/CorpusID:16584254. Our study conclusively demonstrates that AFA can au- [12] D. Maclaurin, D. Duvenaud, R. Adams, Gradient-based tonomously and effectively tune quality filters in adver- hyperparameter optimization through reversible learn- tising programs, outperforming manual tuning methods in ing, in: International Conference on Machine Learn- terms of both speed and precision. AFA significantly re- ing, 2015, pp. 2113–2122. duces the number of necessary evaluations, enabling faster [13] L. Li, K. Jamieson, G. DeSalvo, A. Rostamizadeh, A. Tal- convergence towards optimal settings. This automated ap- walkar, Hyperband: A novel bandit-based approach proach streamlines the optimization process and mitigates to hyperparameter optimization, in: International the potential for human error and bias. The implications of Conference on Learning Representations, 2017. our findings are suggesting that AFA could be instrumental in various optimization tasks across the digital advertising