April 13, 2024
Pinterest Engineering
Pinterest Engineering Blog

Erika Solar ML Engineer | Advertiser Development Modeling Group; Ogheneovo Dibie Engineering Supervisor | Advertiser Development Modeling Group

Old, rustic boat sinking in ocean — Photo by Jason Blackeye on Unsplash
Photograph by Jason Blackeye on Unsplash

On this weblog submit, we describe a Machine Studying (ML) powered proactive churn prevention resolution that was prototyped with our small & medium enterprise (SMB) advertisers. Outcomes from our preliminary experiment counsel that we will detect future churn with a excessive diploma of predictive energy and consequently empower our gross sales companions in mitigating churn. ML-powered proactive churn prevention can obtain higher outcomes than conventional reactive handbook effort.

Like many ads-based companies, at Pinterest, we’re intently targeted on minimizing advertiser churn on our platform. Historically, advertiser churn is addressed reactively. Particularly, a gross sales individual reaches out to an advertiser solely after they’ve churned. This strategy is difficult as a result of it’s extremely troublesome to “resurrect” a buyer as soon as they go away the platform. To handle the challenges with addressing churn reactively, we current a ML-powered proactive strategy to advertiser churn discount. Particularly, we developed a mannequin that may predict the probability of advertiser churn within the close to future and empowered our gross sales workforce with insights from this mannequin to stop in danger accounts from churning.

On this weblog, we cowl the:

  • Churn prediction mannequin’s design and implementation
  • Experimentation within the managed North America SMB phase

Our workforce constructed a ML mannequin to foretell advertiser’s churn probability within the subsequent 14 days. We use the Shapely Additive Rationalization (SHAP) bundle to estimate the mannequin’s options’ contribution to the churn prediction. We offer the mannequin churn prediction together with prime contributing options to gross sales. Gross sales makes use of this info to prioritize their effort to mitigate churn for advertisers in danger. We’ll speak about every part in additional element within the following subsections.

Mannequin Structure

The preliminary model of our mannequin relies on a snapshot Gradient Boosting Choice Tree (GBDT) structure. We selected GBDT for the next causes:

  1. GBDT is a extensively used mannequin with good efficiency on small to medium sized tabular knowledge* (our knowledge matches on this description).
  2. SHAP works nicely with GBDT to estimate options’ contributions.
  3. Mannequin function significance is straightforward to generate with GBDT.
  4. It will possibly additionally function a very good baseline mannequin for future mannequin enhancements, e.g. a sequential mannequin.

*Snapshot means we use all the data obtainable as much as a given timestamp to foretell the churn likelihood within the subsequent 14 days with respect to that timestamp.

Goal Variable

After thorough evaluation and session on the enterprise wants, we determined to make use of the next goal variable definition (see Determine 1).

7/01 to 07/07 is 7 day spend >0. 07/07 to 07/21 is 14 days. 07/21 to 07/27 is 7 day spend >0 ? If yes, then Label 0: active. If no, then Label 1: churn.
Determine 1: Goal Variable Definition

For our use case, we distinguish between an energetic and churned advertiser as follows:

  • Lively advertiser: spent within the final 7 days
  • Churned advertiser: no spend within the final 7 days

We solely predict the churn probability for energetic advertisers. Particularly, we predict if they’ll churn within the subsequent 14 days.


There are over 200 options used within the mannequin. These options are aggregated throughout totally different statistical measures–e.g. min, avg, max and many others — over a spread of time home windows such because the previous week / month previous to the inference dates. We additionally embody week over week and month over month change options to replicate current developments. These options will be grouped within the following classes:

  • Efficiency: impressions**, clicks, conversions, conversion values, spend, price per 1000 impressions, price per click on, clickthrough charge
  • Purpose: purpose attainment ratio, distance to purpose
  • Funds: price range and utilization
  • Advertisements supervisor actions: creates, edits, archives, customized stories
  • Property: gross sales channel, nation, business, tenure, measurement, spend historical past
  • Marketing campaign configuration: focusing on, bid technique, goal sort, marketing campaign finish date

**View greater than 1 second.

Function Contribution

We use the SHAP library to estimate the function contribution to mannequin likelihood output. Sigmoid of the sum of the options’ SHAP contribution is the same as mannequin likelihood. From SHAP function contribution, we will know what the important thing drivers are of excessive churn likelihood. We then spotlight them for the Gross sales workforce to stop churn.

We use an offline educated mannequin to deduce energetic advertisers’ churn likelihood every day.

Churn Danger Class

To assist the Gross sales workforce higher perceive the which means of the mannequin output, we classify accounts into three classes primarily based on their churn likelihood: excessive, medium, and low churn threat. Excessive churn threat captures the accounts which might be principally more likely to churn with excessive precision. Medium churn threat captures the accounts which have a decrease probability of churn. Low churn threat comprises the ‘wholesome’ accounts which might be unlikely to churn within the subsequent 14 days. We choose the thresholds to outline totally different churn threat classes based on the Gross sales workforce’s request of desired precision and recall. Extra particulars will be present in Experiment Outcome.

Our first experiment was targeted on SMB accounts in North America which might be managed by Gross sales Account Managers (AMs). We break up the advertisers randomly into therapy and management teams throughout the experiment inhabitants. For the management group, we don’t make any adjustments to the prevailing Gross sales workforce procedures. For the therapy group, we supported the Gross sales workforce to stop churn with the next info:

  1. Churn Danger Class: Excessive / medium / low churn threat
  2. Churn Purpose Class. We categorized the detailed churn causes into coarse churn classes to ease understanding. The Gross sales workforce carried out investigations utilizing churn classes as instructions.
14 Day Churn Prediction Model — Overall Churn Risk High. Churn Category is Performance and Campaign Setup / Best Practices. Absolute Change in 14d Churn Risk % D/D is -11% down.
Determine 2: Churn Data Widget

Experiment Success Metrics

Our experiment was evaluated primarily based on the next standards:

  1. Mannequin predictive energy, i.e. how nicely our mannequin is ready to determine advertisers which might be more likely to churn
  2. Efficacy of churn prediction in churn discount

Mannequin Predictive Energy

With the intention to decide the mannequin’s predictive energy, we in contrast its on-line efficiency on the management group (i.e. AMs who didn’t have entry to the churn predictions) to what we had noticed offline throughout improvement (i.e. our out-of-sample analysis). Particularly, we measured mannequin efficiency primarily based on:

  1. Mannequin high quality: We in contrast the AUC-ROC and AUC-PR noticed on-line to offline.
  2. Churn threat segmentation: In session with gross sales, we decided thresholds for prime, medium, and low churn threat classes in order that:
  3. Recall in excessive and medium churn threat needs to be above 70%.
  4. Precision in excessive churn threat needs to be round 70%.

This permits gross sales to seize most accounts prone to churning whereas additionally prioritizing methods to work by means of them, i.e. excessive churn threat first (highest precision).

With respect to mannequin high quality, our outcomes point out that the AUC-ROC noticed on-line is inside 1% of the offline AUC-ROC and the net AUC-PR is inside 3% of the offline AUC-PR. This means that the mannequin’s predictive energy in figuring out at-risk accounts is akin to what we noticed offline.

When it comes to churn threat segmentation, our mannequin’s precision, recall, and proportion of the inhabitants captured throughout the excessive and medium threat churn classes had been persistently inside 2–3% of our offline analysis. This means that the segmentation of account threat primarily based on churn probability had been in step with our offline analysis and gross sales expectations.

Efficacy of Churn Prediction in Advertiser Churn Discount

We noticed a 24% (statistically vital) discount within the churn charge of excessive tier pods*** in our experiment therapy group in comparison with the management. This means that accounts whose churn dangers had been uncovered to AMs had been much less more likely to churn than those who weren’t.

*** In excessive tier pods, AMs handle about 50–70 accounts on common.

On this weblog submit, we illustrated the event and implementation of an ML-based resolution for proactive churn prevention at Pinterest. We’re additionally actively investigating sequential mannequin architectures equivalent to Lengthy short-term reminiscence (LSTM) and Transformers, which can higher seize the utilization behaviors of advertisers and reduce the necessity for handbook function engineering equivalent to week-over-week or month-over-month function aggregation utilized in our present mannequin.

Advertiser Development Modeling Group

  • Engineering: Erika Solar, Ogheneovo Dibie, Keshava Subramanya, Mao Ye
  • Product: Shailini Pandya
  • Product Analytics/Knowledge Science: Alex Simons

Gross sales Group

  • Product: Wesley Kwiecien, Grace Yun
  • Gross sales Managers: Abby (Fromm) Lubarsky

Salesforce Group

  • Engineering: Gayathri Varadarangan (She Her), Murthy Tumuluri, Phani Chimata, Gabriela Mihaila, Richard Wu

Optimization Workbench Group

  • Engineering: Phil Worth, Jordan Boaz, Lucilla Chalmer
  • Product: Dan Marantz

[1] When and Why Tree-Based Models (Often) Outperform Neural Networks | by Andre Ye | Towards Data Science

To be taught extra about engineering at Pinterest, try the remainder of our Engineering Weblog and go to our Pinterest Labs web site. To discover life at Pinterest, go to our Careers web page.