# Brain Feature Value report

{% admonition type="danger" name="This API is deprecated and will be removed in July 2026" %}
  The new API docs can be found here: https://apidocs.mediamath.com/apis/reporting-api
{% /admonition %}

## Brain Feature Value Report

Sample usage: GET https://api.mediamath.com/reporting/v1/std/brain_feature_value/meta?v1

### Background
The Brain Feature Value report gives you transparency into the top 100 and bottom 100 feature values (e.g. site A, creative 123, exchange 1) that impact predictions for each optimization Goal Type (for example, ROI, CPA, VCR).  This shows how each Brain Feature Value directly influences and informs impression prediction and pricing.  Descriptions of Current Brain Features themselves are available here. When T1 generates a Brain, there can be 200K+ predictions for Feature Values. To make these predictions more accessible, T1 selects the top 100 and bottom 100 Feature Values (which can vary daily) generated for that day for your viewing per optimization goal.

The information in this report is provided when a new Brain model is generated, but some days in the reporting period may be missing (which is normal). Depending on the training data, we may not always generate a new model so you may see gaps (not all days will generate the report). The report time rollups are updated daily, typically before 18:00 UTC. The data contained for the date in the report will be for the latest model picked up on that day and loaded into reporting.  Retention is on a rolling 30 days, with a date range therefore of up to 30 days.  T1 continues to train Brain models just in case the campaign is set live again, maximizing your opportunity to take advantage of the Brain Optimization on spend. Once there is no more training data (after 23 days of no activity) T1 will stop generating new models until spend recommences.

The information included in the Brain Feature Value report is not relevant in the following circumstances:
* Campaigns powered by a Custom Brain (for more on Custom Brain see here)
* CPA/ROI strategies within a campaign using 3rd-party attribution
* CPA strategies within a campaign with post-view attribution discounted below 100%

### How to run a Brain Feature Value report via T1 instead of by API:
1. Navigate to the Reports module
2. Click on the Data Export tab
3. Type the name of the report in the File Name field
4. Select the Brain Feature Value report from the Report Type drop-down list
5. Select the date range you want your report to cover
6. Select Agency, Advertiser and Campaign from the relevant drop-down lists
7. Select your preferred Dimensions

### The Brain Feature Value report itself
This report exports multiple campaigns and Goal Types at once.

### Sample CSV view of a Brain Feature Value report
| | A | B | C | D | E | F | G | H | I |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | start_date | end_date | organization_id | advertiser_id | agency_id | campaign_id | model_goal | feature_report_type | feature |
| 2 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | top_features |  conn_speed |
| 3 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | top_features |  exchange_id |
| 4 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | top_features |  dma_id |
| 5 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | top_features |  fold_position |
| 6 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | top_features |  size |
| 7 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | top_features |  os |
| 8 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | top_features |  site_id |
| 9 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | bottom_features |  id_vintage |
| 10 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | bottom_features |  region_id |
| 11 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | bottom_features |  day_of_week |
| 12 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | bottom_features |  size |
| 13 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | bottom_features |  device_id |
| 14 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | bottom_features |  channel_type |
| 15 | 10/8/2019 | 10/8/2019 | 100001 | 10002 | 10003 | 111111 | cpa | bottom_features |  country_id |

### What's in the Brain Feature Value report

| Column | Example* | Description |
|---|---|---|
| Start Date | 8/31/2019 | Date when these values were being used in bidding. |
| End Date | 8/31/2019\t| Date when these values were being used in bidding. |
| Organization ID | 100000 | |
| Agency ID | 10001 | User-specified (Step 6, below). |
| Advertiser ID | 20000 | User-specified (Step 6, below). |
| Campaign ID | 30000 | MediaMath unique ID for campaign. User-specified (Step 6, below). |
| Model Goal | CPA, ROI, CPC, etc | The Goal Type that the model was trained for. Each Goal Type will have a different Brain model. |
| Feature Report Type | bottom_features or 
top_feature | bottom_features: a feature value that was in the bottom 100 (lowest \"Mean\").  top_features: a feature value that as in the top 100 (highest \"Mean\") |
| Position | 0, 1, 2, 3, 4, ..., 9 | For bottom_features, a Position of 0 means it had the lowest \"Mean\".  For top_features, a Position of 0 means it had the highest \"Mean\". |
| Feature | isp_id, day_of_week, size | The Feature that the next column’s Feature Value belongs to, used as input into the Brain Machine Learning model. For a full Feature list, see here. |
| Feature Value| Verizon, Tuesday, 728x90 | Readable name of the Feature Value. Some Feature Values display in the report as hashed or otherwise manipulated in a way that does not accurately map back to a human-readable name, fields including Site ID, Category ID, and Region ID. In Q1 2020 or later, we hope to change the hashing to return these back to human readable names but until then it is in the backlog. Please provide feedback if you would find this an important feature to support sooner. The pipe character 'I' indicating that hashing has been applied may occur when there is a large number of Feature Values in a given Feature.  For example; site_id: \"539409862I130821751I231186\" or isp_id: \"WindstreamITelefonica\" |
| Is Numeric | Y, N | A few of the Features (data types used in the prediction) are numerical in nature and will have a different impact on the prediction depending on the actual number (Y = Yes, it is numeric; N = No, it isn't numeric). One example is \"id_vintage\" which represents how long we’ve had an ID for the particular user we’re trying to predict the Response Rate for. Longer time means better quality (as they don't clear cookies/cache etc.) and improves the odds of conversions. The \"id_vintage\" field simplifies the large time spectrum into 4 specific ranges. |
| Index | 123.855316, -60.73726 | This is based on the Mean. The index column gives you a comparative indication of the relative importance of features based on normalized attribution. While the mean values might not make much sense for a global understanding of the model, the index gives an indication of what the model sees as important features when its trained. The higher the Index number, the higher the importance. |
| Mean | 0.00000648, -0.00000308 | When we train a model, we run a sample of the available data through our proprietary machine learning algorithm. The mean shown in this report is the Mean Attribution of the feature across samples of attribution. While technical, it may interest Data Science users. For the layman, it describes the \"average\" contribution of that Feature Value to the Predicted Response Rate. |
| Bid Impact | 1.27748963, 0.93687577 | The marginal multiplicative contribution (> 1.0 increases, > 1.0 decreases) of a feature value (e.g. a particular exchange or site or creative) to a Predicted Response Rate (norm_rr).  The bid_impact is helpful for understanding pricing impact on that feature value, and to calculate CPM. An oversimplification: the baseline score is also impacted by a calibration coefficient. * T1 logs the predicted Action Rate of specific impressions in the Log Level Data Service under a field called “norm_rr”. Please speak to your MediaMath Account representative for more information about accessing this data. |

*Note: Only 1 value will appear, not a comma-delimited list, although feature_value may be a pipe-delimited list)


Endpoint: GET /brain_feature_value/meta