Meta & Children's Online SafetyFeed Ranking HPM
Raw text
Case 4:22-md-03047-YGR Document 3012-37 Filed 05/02/26 Page 1 of 5 Exhibit 168 Case 4:22-md-03047-YGR Document 3012-37 Filed 05/02/26 Page 2 of 5 Feed Ranking HPM (bi-weekly) [HYPERLINK \I "vzvky0jbvg3i" \h] Process • • • • Every other Thursday - Driver solicits input (team EM/PM/DS) about what to include in biweekly HPM Thursday - PMs write their team's HPM, select one highlight and one shoutout for this HPM Thursday/ Friday Morning - DS write their team's data updates, select one insight and one shoutout for this HPM Friday by 2pm ET - PMs and DS add fill out the doc below using [HYPERLINK VI "vzvky0jbvg3i" \h] Friday by 5pm ET - (alternated) write top of mind and publish HPM in [ HYPERLINK "https://fb.workplace.com/groups/1840194289545280" \h ] [ HYPE [ HYPE RLINK RLINK V Poster V 9/29 10/13 10/27 11/10 11/27 12/8 12/22 "ggib0 "xm5ix yy7zq xa" ih zjy6u5 2" \h] ] X X X (maya ) x X X X X X Feed Ranking HPM - [12/11/2023] [HYPERLINK "mailto Top of Mind @meta.com" \h] Entire Feed Ranking has a great half. Connected Ranking side, we exceeded our global cap-15 session goal and VPV goal. Blending/IFR joint-holdout also exceeds both global and teen cap- 15 session goals. This is nearly mission impossible, when we set up goals in early H2. I still HIGHLY CONFIDENTIAL (COMPETITOR) ΜΕΤΑ3047MDL-065-00400107 Case 4:22-md-03047-YGR Document 3012-37 Filed 05/02/26 Page 3 of 5 • • [Interactions] IFR quality metrics (interaction rates) are fairly stable month over month, with IFR worth your time score up 1.2%. IFR continues to outpace connected content in terms of like rate. [VM Utilization] Increased to 57% (up from 12% at start of half) ([ HYPERLINK "https://fburl.com/unidash/syy4u2ow" \h ]). Can't Fails Priorities 1 Improve explainability of Blending and IFR distribution (milestones) 2 Increase personalization in Blending with more decision made by VM (milestones) 3 Migrate Connected Feed VM to a Semilinear Structure (milestones) 4 Increase Connected Feed model size and complexity (milestones) Status is Current Target Expected At Risk 3 4 4 On Track 3 4 4 Done 3 3 3 On Track 1 4 4 Feed Core Ranking We launched Connected Feed MB3 model this week, which improved prediction task accuracy and topline metrics, and officially completed our model can't fail milestones in H2. Compared to the beginning of H2, this model update increased our training data by 10X for all production models (100% teens and 60% for genpop), model size by 200X, and incorporated a more comprehensive set of features including cross-surface, teen-specific, and embedding features. Overall, this update improved offline prediction accuracy (~5% penter_comment_surface, ~2% pexternal_share, ~1.2% p_comment, ~0.75% preshare) and drove ecosystem value (+0.33% genpop direct reshares, +0.21% genpop likes, +0.22% teen direct reshares, +0.08% high ME imps). We expect the model to mature in the next 6 weeks and continue to see growing topline impact in the backtest. Increasing connected feed model training data has shown +0.06% cap15 sessions for both teen and gen pop, validating our hypothesis that increasing model training data drives value for teens and marginal users ([ HYPERLINK "https://fb.workplace.com/groups/233247545227463/permalink/789919946226884/" \h ]). At the beginning of H1, we hypothesized that low training data volume limits our ability to deliver model value for Teens. In H2, we took a big bet to significantly increase training data sampling rate on connected feed for both teens (5% -> 100%) and gen pop (5% -> 60%). We have now launched increased training data for 4 out of 5 production models, and are on track to complete the full migration by EOH. HIGHLY CONFIDENTIAL (COMPETITOR) ΜΕΤΑ3047MDL-065-00400119 Case 4:22-md-03047-YGR Document 3012-37 Filed 05/02/26 Feed Ecosystem Ranking We launched a LOE Tuning on p(enter comment surface) and p(reels iv entry) predictors in Top Rank VM that drives +.033% Global Capped Sessions, and multiple tier 1 wins. (HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3676262732605084/" th]) We have been leveraging the[ HYPERLINK "https://fb.workplace.com/groups/233247545227463/permalink/743324394219773/" \n ][ HYPERLINK "https://fb.workplace.com/groups/233247545227463/permalink/743324394219773/" \h ] framework this half to find better predictor weights in a post Linear VM migration world, and had some successful launches recently ([ HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3655368638027827/" \h ],[ HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3655368638027827/" \h ][ HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3655368638027827/" h]). This launch increases the weights of p(enter comment surface) and p(reels IV entry), and therefore increases the likelihood a user will enter reels IV or the comment surface based on the first post in feed. Close Friends Feed posts launched globally this week - Feed Ranking collaborated w/ the Entertainment Sharing (ES) Team over the last 2 halves to support [HYPERLINK "https://fb.workplace.com/groups/192691577812157/permalink/1573053673109267/" \h], which was[ HYPERLINK "https://fb.workplace.com/groups/3246924332080190/permalink/6265066933599233/" \h ] by IG leads. Feed Ranking will continue providing ranking support for the[ HYPERLINK "https://fb.workplace.com/groups/398847337271923/permalink/1652347355255242/" \h] this week and continued adoption phase in Q4/H1 2024. ([ HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3682560051975352/" \h ]) Feed Blending [Launched] VM Bundle that adds 5 new predictors and drive +0.024% Cap 15 sessions ([ HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3687927994771891/" \h ]). As part of[ HYPERLINK "https://fb.workplace.com/groups/331709089022333/permalink/562877795905460/" \h ][ HYPERLINK "https://fb.workplace.com/groups/331709089022333/permalink/562877795905460/" \h] we're strengthening the Blending VM with 5 more predictors and moving closer to parity with the IFR VM. This week we launched a bundle of 5 new VM terms to the Blending system - click, follow, save, skip and see less - which collectively drives +0.024% Cap15 Sessions and creates a much stronger VM for us to add personalization on top of. This launch increases IFR load by HIGHLY CONFIDENTIAL (COMPETITOR) Page 4 of 5 ΜΕΤΑ3047MDL-065-00400120 Case 4:22-md-03047-YGR Document 3012-37 Filed 05/02/26 Page 5 of 5 +1.6% topline which on a standalone basis is below the Sessions efficiency threshold. We escalated to Home leads and secured approval to launch as this system improvement unlocks future work (VM tuning, constraint relaxation and PVM) which we expect can all be session drivers in the near future. [Launched] Responsive Blending: enabling source/rank/blend flow on every page ([ HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3688580801373277/" \h ]). Current Feed Ranking flow only runs "source and rank step" (the main delivery step that orchestrates sourcing/ranking/blending) on the first or first two pages, making it difficult to leverage in-sessions signals to source more fresh connected and recommendations posts and improve ranking as users scroll further in feed. Last week we launched a change which runs sourcing and ranking more consistently on each page, as well as making top of feed ranking and blending more efficient. This change brings stronger guarantees on fresh IFR inventory as users scrolls deeper in feed and we are seeing +0.10% revenue, -0.33% TTNC (9ms), 80KW overall capacity regression (100KW CPU regression and 20KW GPU saving), +0.087% VPVD w/ IV for teens, +0.58% scroll depth, with regressions and IFR load increase (+0.24% topline) within threshold. This is a foundational change that will unblock the several subsequent workstreams including [HYPERLINK "https://fb.workplace.com/groups/1011416399422153/permalink/1331559214074535/" \h] and [ HYPERLINK "https://fb.workplace.com/groups/1011416399422153/permalink/1316547502242373/" \h ]. We found an Organic HP1 candidate that drives +0.44% IG revenue with minimal engagement tradeoffs & highly efficient ad imps increase. ([ HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3680129298885094/" \h ]) This candidate leverages (1) time since last session, and (2) # of sessions in the part hour - to determine the likelihood a user sees an ad in position 1. These signals drive highly efficient feed revenue per feed ad impression (4.5: 1), and pass the engagement tradeoff efficiency for all[ HYPERLINK "https://fburl.com/workplace/jxvu2rm6" \h ][ HYPERLINK "https://fburl.com/workplace/jxvu2rm6" \h ]. We are in progress on outlining a path to launch, which is targeted for early Q1 2024. This candidate furthers confidence in our collab w/ the Ads org that we can find better tradeoffs of user value and revenue by leveraging organic signals in ads insertion. Recent Launches [Notifs / Stories / Feed] [HYPERLINK "https://fb.workplace.com/groups/1840194289545280/permalink/3678315429066481/" \h ] (10/30). This is a collaborative launch with IG Growth adds the signal p(time_spent & click), top feed / stories engagement features, and more training data to the UAR (Unified Actor Ranker) value model for push notifications, resulting in +0.067% absolute projected Global DAU, and SS wins to DM Teen MAU, cap15 sessions, and notif CTR. HIGHLY CONFIDENTIAL (COMPETITOR) ΜΕΤΑ3047MDL-065-00400121