GAIP: Grouped Ad Image Preference Dataset

• GAIP is a large-scale public dataset comprising over 40 million users and 9.5 million ad images to capture group-wise advertising preferences.
• It employs advanced embedding techniques, cross-attention mechanisms, and product-aware k-means clustering to construct robust group features.
• The dataset enables precise evaluation with metrics like CTR, NDCG@5, and AUROC, informing targeted marketing and group-conditioned CTR optimization.

The Grouped Advertising Image Preference Dataset (GAIP) is the first large-scale, public resource for benchmarking and modeling group-wise preferences in advertising image generation. Constructed from expansive e-commerce advertising log data, GAIP provides a systematic framework for understanding and exploiting the diversity of user response to visual ads at the group level, enabling researchers to address the limitations of traditional “one-size-fits-all” optimization in click-through rate (CTR) maximization (Lu et al., 2 Feb 2026). GAIP supports robust research in preference-conditioned generation, group-aware evaluation, and the design of targeted marketing algorithms with its rich, multi-modal schema and rigorously curated groupings.

1. Dataset Construction and Grouping Methodology

GAIP is built on three weeks of advertising impressions from a major e-commerce platform, comprising 40,027,535 unique users, 2,085,969 products, and 9,565,154 distinct ad images. Each user–ad exposure is recorded with full context: user profile, product title, ad image, and binary click label.

User and Product Embedding:

• Images $I$ and product titles $T$ are encoded into $d$-dimensional vectors:

$$\mathbf e_v = E_I(I) \in \mathbb R^d,\quad \mathbf e_t = E_T(T) \in \mathbb R^d.$$

• User attributes $A = \{a_1, \ldots, a_N\}$ are embedded (via sum and MLP) into $\mathbf e_u \in \mathbb R^d$.
• Cross-attention is applied to condition $\mathbf e_u$ on product text and then on image:

$$\mathbf e_{u|t} = \mathrm{CA}(Q=\mathbf e_u, K=\mathbf e_t, V=\mathbf e_t),\quad \mathbf e_{u|c} = \mathrm{CA}(Q=\mathbf e_{u|t}, K=\mathbf e_v, V=\mathbf e_v).$$

Product-aware Clustering:

• For each product $s$, the set of user embeddings $\mathcal E_s$ is collected.
• K-Means is run for $K \in [K_{\min}, K_{\max}]$, with the final $K_s^*$ chosen to maximize the Silhouette score:

$$K_s^* = \arg\max_K \mathrm{Silhouette}(\mathcal E_s, K).$$

• Clusters $C_1,\ldots,C^*_{K_s}$ minimize the within-cluster sum of squared deviations, centering each group on its centroid $\boldsymbol\mu_k$.

Group-feature Construction:

• For each cluster $C_k$, both centroid and “peripheral” (percentile-selected) embeddings are aggregated:

$$G_{s,k} = \{\boldsymbol\mu_k\} \cup \{\mathbf p_k^{(j)} \mid j\in \mathcal J\} \subset \mathbb R^d.$$

• The final data quadruple for each (image, product, group) is $(I_s^i, T_s, G_{s,k}, \mathrm{CTR}(I_s^i, k))$, with low-exposure groups filtered.

2. Dataset Statistics and Structural Overview

GAIP comprises 610,172 group clusters, with an average size near $65.6$ users (median $\sim 50$; 60% $\le 100$; 10% $> 200$). Every group is associated with aggregated user-attribute statistics (histogram or mean/std by field). CTR labels for images within groups have a mean $\approx 0.015$, standard deviation $\approx 0.0045$, 5th percentile $\approx 0.007$, and 95th percentile $\approx 0.025$.

Core Table Schemas

Feature Type	Group-level	Image-level
Identifiers	group_id (int32), product_id (int32)	image_id (int32), product_id (int32)
Embeddings	centroid_emb (float32[d]), percentiles	visual_emb (float32[d])
User Statistics	categorical histograms, numeric means	—
Performance	—	impressions_k, clicks_k, CTR_k (per group)

All group and image features are cross-referenced via product/group IDs and organized for efficient access in Parquet or HDF5 formats.

3. Annotation Protocol and Label Quality Control

The CTR for an image $I$ in group $k$ is defined as: $$\mathrm{CTR}(I, k) = \frac{\text{clicks}(I, k)}{\text{impressions}(I, k)}.$$ Only group–image pairs with at least $100$ impressions are retained to ensure statistical stability in CTR estimation. Each record provides both raw counts and the precomputed CTR.

The annotation protocol exclusively relies on real user interaction logs, distinguishing GAIP from human-preference datasets that depend on subjective pairwise votes. However, methods such as Thurstone/MAP aggregation and pairwise preference protocols, as in AIGI-VC (Tian et al., 2024), could be applied for future human-centric extensions.

4. Dataset Splits and Evaluation Strategies

GAIP is partitioned at the product level to prevent data leakage:

• Train: 70% of products (~1.46M products, ~430K groups)
• Validation: 10% of products (~208K products, ~60K groups)
• Test: 20% of products (~417K products, ~120K groups)

All images and group clusters referring to a product appear exclusively within the assigned split.

Key Evaluation Metrics:

• NDCG@5 between group pairs for a single product, quantifying group-specific response diversity:

$$\mathrm{NDCG@5}(k_1 \to k_2) = \frac{1}{\mathrm{IDCG}_5} \sum_{j=1}^5 \frac{2^{\mathrm{rel}_j} - 1}{\log_2(j+1)},$$

with true group $k_2$ CTRs as relevance.

• AUROC for click prediction within each group.
• Pair-Accuracy for offline reward models:

$$\text{PairAcc} = \frac{1}{N} \sum_{n=1}^N \mathbf{1}(\hat r_n(I_w) > \hat r_n(I_\ell)),$$

where $I_w$ exceeds $I_\ell$ in held-out group CTR.

Comparisons to baseline “one-size-fits-all” optimizers (CAIG, CG4CTR) are grounded in the relative CTR gain observed in novel or minority groups.

5. Usage Guidelines and Best Practices

Loading and Preprocessing:

• Use scalable frameworks (e.g., Spark, Pandas) to load tables.
• Normalize group and image embeddings to unit norm if desired.
• Encode categorical and numeric statistics as recommended (one-hot, histogram, standardization).
• Dimensionality reduction (e.g., PCA to $d=128$) is suggested for large-scale training.

Model Input Construction:

• For each sample $(s, k, i)$, combine group features $G_{s,k}$ with product text $T_s$, and use $(I_s^i, \mathrm{CTR}(I_s^i, k))$ as the supervised target.

Training and Evaluation Recommendations:

• Enforce a minimum exposure threshold of $100$ impressions per group–image for reliability.
• When fine-tuning vision-LLMs, it is advised to freeze the image encoder and update only group-fusion modules (LoRA is recommended for efficiency).
• Proactively monitor group-level CTR to detect and mitigate overfitting to dominant groups or leakage across splits.

Baseline and Comparative Analysis:

• CTR-optimizing models that ignore group structure are expected to show diminished generalization to minority or atypical clusters (Lu et al., 2 Feb 2026).

6. Extensions, Cross-Dataset Context, and Broader Significance

GAIP’s schema aligns with emerging methodologies in visual advertising evaluation and human-centric preference collection, as seen in AIGI-VC (Tian et al., 2024). While GAIP builds its labels from behavioral signals (CTR), AIGI-VC leverages structured pairwise preference annotation and Thurstone/MAP score aggregation, offering a blueprint for introducing interpretable explanations and fine-grained rationales in future GAIP-like datasets.

Key insights from AIGI-VC applicable to GAIP include:

• Effective group formation can be shaped by campaign, demographic, or creative axes, with MAP-based aggregation scalable to any grouping.
• Hierarchical annotation (pairwise preferences, textual rationales) grounded in specific cues (clarity, emotion, composition) is recommended for future human-in-the-loop GAIP extensions.
• Benchmarking should consider accuracy, correlation, and consistency in both coarse and fine preference dimensions.

A plausible implication is that combining GAIP’s behavioral diversity with interpretable, multi-tier annotation protocols—as pioneered in AIGI-VC—will accelerate progress toward both highly granular targeting in advertising and more generalizable models for group-conditioned visual communication (Lu et al., 2 Feb 2026, Tian et al., 2024).