Client Data Client data will never be shared, nor even used in its raw form to ensure data protection and competitive IP. Only highly abstracted statistics will enter the data graph: data that describes data, not the underlying data itself.

PCG Overview

To understand the Performance Index we first give an overview of how the PCG is structured and that the DCI Dashboard is a component of the Data Composition Index (DCI) Core Element.

DCI Breakdown

The DCI allows the de-composition of campaign performances to understand how well a campaign performed and what the major drivers of this performance along the defined recipe-dimensions are.Therefore the DCI should act as a simple, easy to understand campaign KPI across all forms of media that accurately reflects performance relative to well-defined benchmark criteria.DCI should enable easier decisions e.g. between data providers or DSP settings as well as enable other graph features such as recipe benchmarks or optimisation via the Performance Index.

The DCI consists of 4 components that are calculated sequentially

Performance Index Calculation

To calculate the Index, we look first at the campaign goals for each campaign across our network which are matched and then standardised across a logical set of benchmark groups

After the standardisation process, a model is incorporated to predict performance using recipe dimensions as predictors.

Finally, the model balances are scaled from 0-100 and used as the final index score.

This can be seen below in more detail below:

Standardised Outcomes

First we need to make CpM / CpV / etc. campaigns comparable

Open Decisions:

Multi-dimensional / weighted target vs. singular mapped target
Relevant target map

Determining the target KPI

Insertion Order Goal	Campaign KPI
CPM	CPM
CtR	CtR
CpV	CpV
Other	CPM

Scaling KPIs

The assigned target KPI of each campaign is first standardised using a robust scalar to remove outlier bias and scale all KPIs to the same dimension.

In a second step they are then rescaled from 0-100 for ease of interpretation. All values that are outside of 1 SD of the median values are automatically scaled to the extremes to reflect extreme under- / over performance.

Open Decisions:

Best scaling approach
Re-scaling before modelling necessary?
Standard Deviation range of cutting of re-scaling or alternative outlier treatment

Example:

Campaign KPI	Scaled Raw Score
5€ CPM	50
25% CTR	60
0,02€ CPV	55
7€ CPM	30

Mapping Campaign Composition

We map all relevant performance drivers to a simplified campaign composition Segment data as well as DSP Settings are projection mapped to reduce complexity and bring categorical and semantic values to the same mathematical level.

UMAP projection is used to reduce campaign recipes to a two-dimensional vector trained on 2022 data, producing spatial similarity between campaigns that share a lot of settings.

This allows computationally efficient searches and benchmarks as well as more intuitive visualisations.

Open Decisions:

Embedding and projection vs. different model approaches
DB-based Graphing vs. KNN Networks
Pre-Selection of relevant dimensions and feature engineering vs. open, raw data approach

Modelling Performance Attribution

Using an attribution approach we model the performance and score residualsUsing the UMAP benchmark criteria we build a predictive model for the raw performance scores.
The residuals (delta prediction vs. truth) are then taken and scaled to a 0-100 final performance score.
Therefore, the performance score reflects the performance relative to the benchmark set and not an absolute performance.

Open Decisions:

Benchmark group vs. benchmark predictors
Tightness of benchmarks determines relativity of performance
Absolute target elements