{"id":1760,"date":"2026-09-30T14:56:29","date_gmt":"2026-09-30T14:56:29","guid":{"rendered":"https:\/\/cms.research.wpp.com\/?post_type=research_feed&p=1760"},"modified":"2026-09-30T15:25:55","modified_gmt":"2026-09-30T15:25:55","slug":"performance-aware-campaign-recommendations-from-real-marketing-data","status":"publish","type":"research_feed","link":"https:\/\/cms.research.wpp.com\/?research_feed=performance-aware-campaign-recommendations-from-real-marketing-data","title":{"rendered":"Performance-Aware Campaign Recommendations from Real Marketing Data"},"content":{"rendered":"\n

<\/p>\n\n\n\n

Abstract<\/h2>\n\n\n\n

Marketing teams often need to decide which campaign setup is worth testing before any budget is spent. The hard part is that performance does not depend on a single variable. It depends on the interaction between multiple factors, such as the brand, its audience, the campaign objective, the platform, the geography, and many other contextual signals.<\/p>\n\n\n\n

This proof of concept from the Performance AI team at Open Intelligence (OI PoC) addresses this as a multimodal prediction and recommendation problem. We trained a single model that learns from historical campaign performance, predicts click-through rate (CTR), and uses the same learned embedding space to recommend campaign completions. The result is a system that can move from \u201cwhat happened before?\u201d to \u201cwhat should we try next?\u201d while staying grounded in real performance data<\/strong>.<\/p>\n\n\n\n

This work builds on the earlier Campaign Performance Modelling Pod<\/a>, which explored how AI can support campaign performance prediction using synthetic campaign data (see the Technical Report<\/a> and Blog Post<\/a>, respectively). The OI PoC analysis takes the next step by applying the same broader modelling agenda to observed\/real performance data<\/strong>, trying to also examine both tasks of CTR prediction and recommendation in order to answer: given a partially specified campaign setup, what completion is most likely to perform well?<\/p>\n\n\n\n

The prepared dataset contains 20,067 unique real campaign setups derived from 21,247,221 raw daily rows. The evaluation is designed as a warm-start compositional generalization problem: the model sees known modality values during training, but must generalize to unseen full combinations at test time. This is important because it mirrors the real planning use case more closely than simple row memorization.<\/p>\n\n\n\n

In our evaluation, the model reached a test RMSE of 0.0151, MAE of 0.0052 and R2 of 0.6001 on unseen campaign combinations. For closed-vocabulary brand-fixed recommendations, the system achieved 76.2% positive precision@10 versus 50.2% for random selection from the same candidate pools, while reducing negative recommendations from 18.1% to 2.9%. For open-vocabulary recommendation, the strongest evidence appears in broader query settings where the model can still identify geometrically coherent completions that rank high relative to each brand\u2019s historical CTR range.<\/p>\n\n\n\n

1. Motivation<\/h2>\n\n\n\n

Campaign planning is highly combinatorial. A planner is not making one decision in isolation, but several interacting decisions at once: which audience to target, which campaign objective to optimize for, which platform to use, and which geography to prioritize for a given brand. Each of these choices can materially change the outcome, and the effect of any one decision depends on the others around it. In practice, that means the space of possible campaign setups grows much faster than the amount of historical evidence available to support them.<\/p>\n\n\n\n

This follows naturally from the previous Campaign Performance Modelling<\/a> work. The earlier synthetic-data analysis asked whether AI could help move campaign planning from intuition-led guesswork toward more systematic performance foresight. Here, we ask a complementary question: once real OI performance data is available, can we learn a performance-aware representation that supports both CTR prediction and recommendation?<\/p>\n\n\n\n

This is exactly the gap OI PoC addresses. We want to help answer two related questions before budget is committed:<\/p>\n\n\n\n

    \n
  1. If a team is considering a particular campaign setup, what CTR should they expect?<\/li>\n\n\n\n
  2. If part of the setup is already fixed, what completion is most likely to perform well?<\/li>\n<\/ol>\n\n\n\n

    The challenge is that historical data is both rich and sparse at the same time. At the raw row level there is plenty of volume, but once performance is aggregated into distinct campaign setups, only a small fraction of the theoretically possible combinations has ever been observed. A useful model therefore cannot depend on exact lookup or memorization. It needs to learn reusable structure from known examples and then transfer that structure to unseen combinations.<\/p>\n\n\n\n

    This creates several practical modelling challenges:<\/p>\n\n\n\n

      \n
    1. Prediction models can accidentally memorize historical rows instead of learning reusable structure.<\/li>\n\n\n\n
    2. Recommendation systems can look accurate when the available candidate pool is already easy, even if the model is not adding much ranking value.<\/li>\n\n\n\n
    3. Performance labels are not globally comparable across brands, because some brands naturally operate at different CTR ranges than others.<\/li>\n\n\n\n
    4. Recommendation quality depends not just on whether a setup looks broadly similar to past positives, but on whether its internal combination of brand, audience, objective, platform and geography is coherent.<\/li>\n<\/ol>\n\n\n\n

      The goal of OI PoC was to build a model that could generalize across known campaign ingredients in new configurations. This is a warm-start compositional generalization problem: the model has usually seen the individual ingredients before, but not the exact recipe. That is why the project combines CTR prediction and recommendation in a shared multimodal framework. The prediction side estimates likely performance for unseen setups, while the recommendation side uses the same learned representation to identify strong or plausibly strong completions for partially specified queries.<\/p>\n\n\n\n

      2. Dataset<\/h2>\n\n\n\n

      The raw source dataset contained 21,247,221 daily campaign rows. After aggregation by campaign setup and filtering low-signal rows, this reduced to 20,067 unique learnable combinations. Furthermore, brand values were anonymized before being used as modelling inputs. This preserves the ability to learn brand-specific performance behaviour while keeping the modelling dataset separated from directly identifiable brands.<\/p>\n\n\n\n

      The five campaign modalities and their coverage are summarized in Table 1:<\/p>\n\n\n\n

      Modality<\/th>Coverage in the prepared dataset<\/th><\/tr><\/thead>
      Brand<\/td>75 anonymized brands<\/td><\/tr>
      Audience<\/td>more than 1,000 audience representations<\/td><\/tr>
      Objective<\/td>7 objectives<\/td><\/tr>
      Platform<\/td>42 platform representations<\/td><\/tr>
      Geography<\/td>167 geography representations<\/td><\/tr><\/tbody><\/table>
      Table 1. Campaign modalities and coverage in the prepared dataset.<\/figcaption><\/figure>\n\n\n\n

      The train, validation and test split used 17,056 training rows, 1,505 validation rows and 1,506 test rows. The test set was designed to avoid exact-combination leakage. <\/p>\n\n\n\n

      The split diagnostics are reported in Table 2:<\/p>\n\n\n\n

      Split diagnostic<\/th>Result<\/th><\/tr><\/thead>
      Exact 5-way test combinations seen in training<\/td>0 \/ 1,506<\/td><\/tr>
      Truly new 5-way test combinations<\/td>1,506 \/ 1,506<\/td><\/tr>
      Test brands seen in training<\/td>71 \/ 71<\/td><\/tr>
      Test objectives seen in training<\/td>7 \/ 7<\/td><\/tr>
      Test platforms seen in training<\/td>39 \/ 39<\/td><\/tr>
      Test geographies seen in training<\/td>117 \/ 117<\/td><\/tr><\/tbody><\/table>
      Table 2. Test-set leakage and warm-start diagnostics.<\/figcaption><\/figure>\n\n\n\n

      This means the evaluation is not a pure cold-start task. It is a test of whether the model can recombine known entities into unseen campaign setups.<\/p>\n\n\n\n

      3. Why CTR Needed Brand-Relative Labels<\/h2>\n\n\n\n

      Click-through rate (CTR) serves as our example performance metric\u2014a practical starting point given the available data. We use it across the dataset, including campaigns with objectives other than clicks, to explore whether the approach can support prediction and recommendation. The positive, neutral and negative labels therefore describe brand-relative CTR, not success against each campaign\u2019s stated objective. Stronger click-through performance does not necessarily indicate stronger awareness, engagement or video-view outcomes. The next step is to extend the approach to available metrics better suited to those objectives.<\/p>\n\n\n\n

      CTR distribution statistics are shown in Table 3, proving that CTR was highly skewed:<\/p>\n\n\n\n

      Statistic<\/th>CTR<\/th><\/tr><\/thead>
      Mean<\/td>0.0106<\/td><\/tr>
      Median<\/td>0.0027<\/td><\/tr>
      Skewness<\/td>7.8710<\/td><\/tr>
      Kurtosis<\/td>108.02<\/td><\/tr><\/tbody><\/table>
      Table 3. Summary statistics of the CTR distribution.<\/figcaption><\/figure>\n\n\n\n

      A global definition of \u201cpositive\u201d campaign performance would over-favour brands with naturally higher CTR and penalise brands with lower typical CTR. Instead, each brand\u2019s own CTR distribution was split into rank-based thirds: negative, neutral and positive.<\/p>\n\n\n\n

      This created labels that mean \u201cgood for this brand\u201d, rather than \u201chigh CTR globally\u201d. The anonymized brand identifiers are still consistent across rows, which means the model can learn within-brand CTR patterns without needing access to the original brand names.<\/p>\n\n\n\n

      The final label distribution is reported in Table 4:<\/p>\n\n\n\n

      Label<\/th>Rows<\/th><\/tr><\/thead>
      Negative<\/td>5,984<\/td><\/tr>
      Neutral<\/td>8,030<\/td><\/tr>
      Positive<\/td>6,053<\/td><\/tr><\/tbody><\/table>
      Table 4. Brand-relative CTR label distribution.<\/figcaption><\/figure>\n\n\n\n

      Rank-based splitting was used because many brands had zero-heavy CTR distributions. Standard quantile thresholds often collapsed to the same value, while rank-based labels preserved useful within-brand ordering.<\/p>\n\n\n\n

      4. Model Approach<\/h2>\n\n\n\n

      OI PoC uses one shared model for both prediction and recommendation.<\/p>\n\n\n\n

      Each modality starts from a 256-dimensional embedding and passes through its own projector:<\/p>\n\n\n\n