pinterest·Feb 13, 2026·3 min readCurated summary

GPU-Serving Two-Tower Models for Lightweight Ads Engagement Prediction

Pinterest replaced its CPU-served two-tower model for ads lightweight ranking with a GPU-serving architecture based on MMOE and DCN. The more expressive model maintained latency comparable to the CPU baseline while reducing offline CTR loss by 5–10%. Separating standard and shopping ad models produced another 5–10% loss reduction and doubled offline iteration speed, with online improvements in CPC and CTR. ## Role of Lightweight Ranking - Lightweight ranking serves as an intermediate stage in Pinterest’s ads recommendation pipeline. - It filters a large pool of candidate ads before more complex downstream ranking models process them. - The two-tower design balances quality and latency: - The Pin tower generates ad embeddings offline through batch updates. - The query tower generates real-time user embeddings. - The prediction score is the sigmoid of the embeddings’ dot product. ## MMOE-DCN Model Architecture - The new system replaces the previous Multi-Task Multi-Domain (MTMD) model. - It combines: - Multi-gate Mixture-of-Experts (MMOE) with MLP-based gating. - Deep & Cross Network (DCN) layers for modeling feature interactions. - Each expert uses both full-rank and low-rank DCN layers. - Unlike MTMD, MMOE handles multi-task and multi-domain learning without relying on separate domain-specific modules. - GPU serving makes it practical to deploy this larger and more computationally demanding model while preserving CPU-baseline latency. ## Scenario-Specific Modeling - Standard and shopping ad scenarios are served as separate models. - Each model is trained only on data relevant to its scenario. - This specialization delivered an additional 5–10% reduction in offline loss. - Separating the models also doubled the speed of offline model iteration. ## Training Efficiency Improvements - **Dataloader optimization** - GPU prefetching prepares the next batch while the current batch is processed. - Additional worker threads take advantage of the 1 TB of CPU memory available on p4d instances. - **Model code optimization** - Operations that previously allocated zero-filled tensors on the CPU were moved to the GPU. - Fused kernels replaced multiple individual kernels to reduce execution overhead. - **Training configuration** - BF16 precision improved processing speed. - Larger batch sizes increased GPU memory utilization. ## Evaluation Results - The model uses downstream ranking scores as labels and optimizes KL divergence between those labels and its predictions. - Evaluation covers both: - Auction winners—ads ultimately inserted and shown to users. - Auction candidates—ads passed to downstream ranking. - Offline loss decreased significantly across all evaluated slices. - Online experiments showed: - Lower cost per click (CPC), which is favorable. - Higher click-through rate (CTR). GPU-serving a more complex MMOE-DCN two-tower model allowed Pinterest to improve ad engagement prediction without sacrificing serving latency. The results support using GPU infrastructure, scenario-specific models, and targeted training optimizations to scale lightweight ranking systems.

Read original(opens in new tab)

Ctr PredictionRecommendation SystemsGpu ServingTwo Tower Models+4

toss·Dec 1, 2025Original article

Toss Next ML Challenge (opens in new tab)

Toss recently hosted the "Toss Next ML Challenge," a large-scale competition focused on predicting advertisement Click-Through Rates (CTR) using real-world, anonymized data from the Toss app. By tasking over 2,600 participants with developing high-performance models under real-time serving constraints, the event successfully identified innovative technical approaches to feature engineering and model ensembling. ### Designing a Real-World CTR Prediction Task * The competition required participants to predict the probability of a user clicking a display ad based on a dataset of 10.7 million training samples. * Data included anonymized features such as age, gender, ad inventory IDs, and historical user behavior. * A primary technical requirement was "real-time navigability," meaning models had to be optimized for fast inference to function within a live service environment. ### Overcoming Anonymization with Sequence Engineering * To maintain data privacy while allowing external access, Toss provided anonymized features in a single flattened table, which limited the ability of participants to perform traditional data joins. * A complex, raw "Sequence" feature was intentionally left unprocessed to serve as a differentiator for high-performing teams. * Top-tier participants demonstrated extreme persistence by deriving up to 37 unique variables from this single sequence, including transition probabilities, unique token counts, and sequence lengths. ### Winning Strategies and Technical Trends * All of the top 30 teams utilized Boosting Tree-based models (such as XGBoost or LightGBM), while Deep Learning was used only by a subset of participants. * One standout solution utilized a massive ensemble of 260 different models, providing a fresh perspective on the limits of ensemble learning for predictive accuracy. * Performance was largely driven by the ability to extract meaningful signals from anonymized data through rigorous cross-validation and creative feature interactions. The results of the Toss Next ML Challenge suggest that even in the absence of domain-specific context due to anonymization, meticulous feature engineering and robust tree-based architectures remains the gold standard for tabular data. For ML engineers, the competition underscores that the key to production-ready models lies in balancing complex ensembling with the strict latency requirements of real-time serving.

Ctr PredictionAIMachine LearningFeature Engineering+4