We then tested this idea with a set of experiments on simple ML models, comparing a centralised baseline (pooling everyone\u2019s data in the same place and training a model on the full puzzle) with a federated counterpart that keeps the puzzle distributed across several data nodes. Alongside this, we examined how fragmentation to an increasing number of participating nodes affects overall performance.<\/p>\n\n\n\n
Those initial experiments rested on two assumptions:<\/p>\n\n\n\n
- \n
- All the data puzzle pieces look alike. Conceptually, this assumes that every data partner observes and records similar patterns and behaviors. This is a simplification of the real-world, where brands, agencies, and platforms address a diverse spectrum of audiences, products, and experiences.<\/li>\n\n\n\n
- The federated ML model follows a traditional \u201ctabular\u201d architecture. Every row in its input is an observation (e.g. a marketing campaign), every column is a an aspect of that observation (e.g. the targeted audience, the advertised product). The last column is the outcome we are trying to model (e.g. the campaign\u2019s performance). The model then looks for statistical associations among the columns from scratch, without having its own prior knowledge of the world. Again, this is a simplification of today\u2019s AI\/ML model landscape, where LLMs with vast prior world knowledge dominate.<\/li>\n<\/ul>\n\n\n\n
In this next phase of our work, we relax both of these assumptions by exploring:<\/p>\n\n\n\n
- \n
- Federated Learning in the presence of nodes with heterogeneous datasets<\/li>\n\n\n\n
- Federated Fine-tuning of LLM-based predictive models<\/li>\n<\/ul>\n\n\n\n
\n\n\n\nContextualising the data: Marketing campaigns<\/strong><\/h2>\n\n\n\n
Before turning to the experiments, it helps to revisit what the marketing dataset represents and where it comes from. We use synthetic datasets generated by a purpose-built pipeline<\/a> designed to mirror the dynamics of real marketing campaigns. This approach gives us fine-grained control over each dataset\u2019s composition, allowing us to construct targeted scenarios that stress-test the model under controlled conditions while retaining full visibility into the factors that drive campaign outcomes.<\/p>\n\n\n\n
At the heart of this pipeline is a signed marketing knowledge graph, which encodes how marketing attributes interact. Each node is an attribute value\u2014the brand<\/strong> being promoted, the audience<\/strong> it targets, the content\u2019s creative<\/strong> tone and style, the platform<\/strong> it runs on, or the geography<\/strong> it reaches\u2014and each edge captures whether two attributes are compatible (positive signal), incompatible (negative signal), or unrelated (missing signal).<\/p>\n\n\n\n
This graph is the ground truth from which we sample subgraphs to build various synthetic datasets. Each dataset contains campaigns defined as combinations of these attributes: one brand, one audience, one creative (image or video), one platform, and one geographical location, drawn from the graph. Labels follow from the signals linking them: campaigns built from compatible pairings tend to over-perform (Positive<\/strong>), those from incompatible ones tend to under-perform (Negative<\/strong>), and mixed cases fall in between (Average<\/strong>). For full details on how the graph is built and sampled, see the synthetic data generator pod<\/a>.<\/p>\n\n\n\n
\n\n\n\n![\"\"](\"https:\/\/cms.research.wpp.com\/wp-content\/uploads\/2026\/06\/graph-4-1024x708.png\")
Figure 1:<\/strong> Visualisation of two synthetic marketing subgraphs for the Base and 10% Flip datasets. \u201cFlipping\u201d means changing the edge type of every edge incident to a randomly selected 10% of nodes (highlighted by the dashed rectangle).<\/figcaption><\/figure>\n\n\n\n Generating heterogeneous data<\/h3>\n\n\n\n
For our experiments, we utilised the ground-truth graph to extract four subgraphs for four respective datasets. To meet the requirement of using heterogeneous datasets, we deliberately add structural noise by randomly \u201cflipping\u201d the connected edges of a subset of subgraph nodes from positive to negative and vice versa. Figure 1 illustrates this on a sample subgraph: the left-hand panel (\u201cBase Subgraph\u201d) shows it in its original state. The right-hand panel (\u201c10% Flipped Subgraph\u201d) then shows a flip in action: we randomly select a fraction of nodes (here Brand A<\/em>, boxed by the dashed rectangle) and reverse the signal on every edge connected to them, turning positive (green) edges negative (red) and vice versa. The result is a graph that tells a partly contradictory story, simulating the noisy, conflicting evidence of real marketing data.<\/p>\n\n\n\n
The dataset names follow directly the fraction of nodes flipped, represented as percentage in each name. More specifically:<\/p>\n\n\n\n
- \n
- Base<\/strong> (0%) leaves the graph untouched as a clean reference<\/li>\n\n\n\n
- 10% Flip<\/strong> reverses a random 10% of nodes (the case in Figure 1)<\/li>\n\n\n\n
- 20% Flip<\/strong> and 50% Flip<\/strong> apply the same step to larger fractions, adding steadily more noise<\/li>\n<\/ul>\n\n\n\n
Since every dataset shares one origin and differs only in flip rate, this provides a controlled way to observe how the model copes as conflicting evidence accumulates\u2014each dataset large enough to reach stable performance even without training to full equilibrium.<\/p>\n\n\n\n
\n\n\n\nFrom tables to text: Federated LLM fine\u2011tuning with LoRA<\/strong><\/h2>\n\n\n\n
The synthetic datasets start out as tables, where each row is a single campaign described by its features (brand, geography, audience, platform, and creative), and the last column records how that campaign performed\u2014Positive, Negative, or Average. To prepare the data, we first replace these three labels with simple number codes (0, 1, or 2). We then treat the problem as a text classification task, where the model’s job is to read a campaign and produce a single answer at the end indicating its predicted performance.<\/p>\n\n\n\n
Before training, every dataset goes through the same preparation steps, so all the inputs look consistent. Each campaign row is rewritten as a short, plain-language description that the LLM model can read, with each feature clearly labelled. For example, one campaign might look like this:<\/p>\n\n\n\n
Platform: Platform A\nBrand: Premium outdoor apparel brand\nAudience: Young professionals interested in fitness and travel\nCreative: Humorous\nGeo: [1023, 2045, 9876] # Zip codes expression of the place\n<\/code><\/pre>\n\n\n\nThis description is then placed inside a fixed set of instructions that presents the task in exactly the same way for every experiment. During training, we include the correct performance label so the model can learn the patterns. During testing, we leave the label out, and the model has to work it out on its own\u2014deciding whether the campaign is Positive, Negative, or Average.<\/p>\n\n\n\n
- 10% Flip<\/strong> reverses a random 10% of nodes (the case in Figure 1)<\/li>\n\n\n\n
- Base<\/strong> (0%) leaves the graph untouched as a clean reference<\/li>\n\n\n\n
![\"\"](\"https:\/\/cms.research.wpp.com\/wp-content\/uploads\/2026\/07\/new_graph.png\")
![\"\"](\"https:\/\/cms.research.wpp.com\/wp-content\/uploads\/2026\/06\/overall_f1_comparison-1024x586.png\")
![\"\"](\"https:\/\/cms.research.wpp.com\/wp-content\/uploads\/2026\/06\/new_all_metrics_radar-1024x381.jpeg\")