RARe: Raising Ad Revenue Framework with Context-Aware Reranking

TL;DR

RARe tackles revenue optimization in e-commerce recommender systems by balancing revenue and relevance via context-aware reranking. It introduces a two-model framework consisting of context-aware clickers (GBDT-C and SAINT-Q) and a lightweight MLP reranker trained with a revenue-aware loss, with revenue computed from predicted clicks and item revenues and a KL-like difference metric controlling perturbations to the original ranking. The paper provides formal definitions for revenue $\mathcal{R}(\theta)$ and the relevance difference $\mathcal{D}(\pi,\pi')$, and introduces the RAReD dataset for benchmarking. Experiments show revenue gains in the 4–12% range with modest drops in NDCG, demonstrating that context-aware clicks can significantly improve monetization while preserving user engagement.

Abstract

Modern recommender systems excel at optimizing search result relevance for e-commerce platforms. While maintaining this relevance, platforms seek opportunities to maximize revenue through search result adjustments. To address the trade-off between relevance and revenue, we propose the $\mathsf{RARe}$ ($\textbf{R}$aising $\textbf{A}$dvertisement $\textbf{Re}$venue) framework. $\mathsf{RARe}$ stacks a click model and a reranking model. We train the $\mathsf{RARe}$ framework with a loss function to find revenue and relevance trade-offs. According to our experience, the click model is crucial in the $\mathsf{RARe}$ framework. We propose and compare two different click models that take into account the context of items in a search result. The first click model is a Gradient-Boosting Decision Tree with Concatenation (GBDT-C), which includes a context in the traditional GBDT model for click prediction. The second model, SAINT-Q, adapts the Sequential Attention model to capture influences between search results. Our experiments indicate that the proposed click models outperform baselines and improve the overall quality of our framework. Experiments on the industrial dataset, which will be released publicly, show $\mathsf{RARe}$'s significant revenue improvements while preserving a high relevance.

Figures (5)

• Figure 1: Comparison of the classical GBDT model (left) and the proposed GBDT-C model (right). The GBDT-C model requires expanding the original dataset's feature dimension by stacking the neighbor items' features. The colored circles represent the features of the items, the black arrows correspond to the model click model processing of the input features, and $1/0$ on the right indicate the model predictions.
• Figure 2: The illustration for the proposed Chunked Intersample Attention transformation. Tuples $(Q, K, V)$ are generated separately for different search results but with the same linear layers. Therefore, they are labeled with the same colors. Then, these tuples are transformed via the attention mechanism and stacked for next layers processing. The output of the model is shown on the right, where $1/0$ corresponds to the predicted click.
• Figure 3: The reranking model training mode. Input data (the initial search query results) is passed to the reranking model, which generates scores for reranking. Based on these scores, a new re-ranking is generated. This ranking and all possible transpositions are fed to the clicker model, which predicts clicks on the received results. These clicks, the costs per click, and the scores from the reranking model are used in the loss for training the reranking model.
• Figure 4: The scheme of the reranking model in the inference mode. The reranking model takes the items' features and revenues from the initial search results and reorders them to maximize the expected revenue while preserving relevance.
• Figure 5: The Delta Revenue metric $\bar{\mathcal{R}}$ and two relevance metrics for a range of hyperparameters $\alpha$ from (\ref{['eq::r_reg']}). The larger the $\alpha$ is, the more solid the color is. (a) The Difference metric is used to estimate the relevance drop: the smaller the Difference, the less the order of the items is changed. (b) The NDCG metric estimates the relevance change compared to the original search results. The vertical lines show the NDCG metric at the original ranking, as well as with losses of 0.1%, 0.5%, and 1%.