0x00 한

Research / Daily reports

  • Sources read: 1
  • Canon files updated: 0
  • Radar entries: 1
In one lineHuawei published a CIKM 2026 paper that replaced the ranking model of an app distribution platform with 100M+ DAU, from MMoE to PTDG. PTDG builds a separate task dependency graph for each item.

1Replaced MMoE with a ranker that changes inter-task connection strength per item — Huawei

Source Personalized Task Dependency Graphs for Mitigating Signal Erosion in Multi-Task Recommendation — arXiv 2609.04862 v1, submitted 2026-09-04, paper (CIKM 2026 accepted). All 11 authors are affiliated with Huawei Technologies (Shanghai / Nanjing / Dongguan / Beijing). We confirmed this from the affiliation listing in the HTML version and the @huawei.com domain.

Background Ranking models in industrial-scale recommendation systems use a multi-task learning (MTL) structure that predicts several objectives together. A single model learns hierarchically a conversion funnel that runs from shallow interactions to deep conversions, such as Click → Add-to-Cart → Purchase. The commonly used MMoE (Multi-gate Mixture-of-Experts) and its successors fix the inter-task dependency structure of this funnel in the model structure. Experts share information across tasks through gating, and a head for each task predicts each objective.

The problem this paper points out is a rigid inductive bias. That is, it assumes the inter-task dependency strength is the same across all traffic. Inter-task correlation varies greatly with item characteristics. Ad items have a strong immediate-click objective, while game items have a long conversion path from download to active use. A static funnel cannot reflect this heterogeneity. As a result, signal erosion is observed toward the lower part of the conversion funnel (sparse, deep-funnel objectives). It is a phenomenon where the signal attenuates as hierarchical message passing accumulates. This paper proposes PTDG, an MTL ranker that rewires the conversion funnel connections per item, as a replacement for MMoE.

It helps to note the relationship with several threads that already appear in the canon.

  • Meta HSTU (a recommendation-specific sequence Transformer), Wukong, Netflix Foundation Model, and Alibaba TransRetrieval in the scaling law thread improved ranking quality by expanding the parameter and sequence axes. PTDG differs: it makes the dependency structure within the MTL architecture itself the object of learning.
  • Ads domain foundation models such as Meta GEM cover several surfaces by sharing a sequence encoder. PTDG personalizes the signal routing between tasks within a single surface.

What we learnedA. PTDG builds a separate task dependency graph for each item, and keeps causal order with a mask

  • PTDG (Personalized Task Dependency Graphs) generates a different dependency graph for each item through low-rank approximation. This graph rewires the strength of the dependency pathways verified.
    • It predicts from item features a T×T dependency matrix I ∈ ℝ^(T×T), with each task as a node and each dependency as an edge.
    • A low-rank decomposition compresses the parameters here.
  • Physical causal constraints are kept with hard causal masking. A hard mask is applied to the dependency graph so that constraints like Click → Pay (there can be no payment without a purchase) are never reversed verified.
    • GCN (a neural network that gathers and propagates information from neighboring nodes in a graph)-based propagation creates item-adaptive shortcuts under this mask.

What we learnedB. Sparse conversion tasks get their own parameter path so they are not buried under dense-task gradients

  • Adaptive Progressive Masking (APM) stabilizes optimization by decoupling the shared parameters of sparse and dense tasks according to task sparsity. It splits the parameter path so that a sparse conversion task such as payment is not overwhelmed by the gradient of a dense task such as click verified.

What we learnedC. Offline, AUC on sparse conversion tasks rose by up to +1.45% over the baseline

  • On the public KuaiRand1K dataset and the authors' own industrial dataset, AUC on sparse conversion tasks rose by up to +1.45% over the baseline, while dense tasks stayed comparable verified.

What we learnedD. In a two-week A/B on an app distribution platform, CVR rose +1.2% and eCPM +1.9%

  • PTDG was deployed on a mainstream app distribution platform. The paper described this surface as: "a mainstream app distribution platform, serving over 100 million daily active users".

    • We infer this surface to be Huawei AppGallery inferred. The paper names the company but anonymizes the specific surface.

  • The A/B ran for two weeks. It had a 10% base bucket and a 10% experimental bucket, and kept the original production model on the remaining 80%.

    • Significance was assessed with a two-sample t-test (p<0.05) on daily aggregate metrics.

  • The paper states that CVR rose +1.2% and eCPM rose +1.9%, and that these improvements are statistically significant.

    • CVR is the conversion rate, and eCPM is a metric of ad revenue per impression.
    • The baseline is the currently deployed MMoE. It is the same architecture used in the offline comparison in §4.2 Table 2.

    "Results show that PTDG achieved statistically significant improvements on core business metrics, with a +1.2% lift in CVR and +1.9% lift in eCPM." — arXiv 2609.04862 v1 §4.5, 2026-09-04

What we learnedOther points

  • The emails of all authors are on a Huawei domain.

    • The co-first author is Fuyuan Liu, and the corresponding author is Tiandeng Wu. The other authors are Yaqun Fang, Wei Zhou, Zehao Zhou, Wenping Chen, Qishun Mei, Jiaxin Zhou, Heng Chang, Yi Cao, and Jiandong Ding.
    • All of their emails are @huawei.com.

  • The end of the paper includes a generative AI usage disclosure (GenAI Usage Disclosure). This kind of notice follows CIKM/ACM rules and has become common in recent industry papers.

    "Generative AI tools assisted with language editing and code debugging. All experimental results were produced by author-written scripts and verified against saved artifacts." — GenAI Usage Disclosure at the end of the paper

Canon update Added a PTDG entry at the top of the new-this-month section of radar.md. Huawei is not a service the canon (this repository's notes by service and topic) tracks, so no services canon is created. If a second company makes a similar personalized task dependency statement (a combination of GCN, causal masking, and APM), it becomes a candidate for a new section in topics/multi-task-ranking.md.

Why it matters An approach that makes the dependency structure itself, the architectural axis of MTL, the object of learning occupies a different role from the scaling axes (parameters, sequence, data) and the loss axis that the canon has often covered. Here the loss axis means pxtr (the per-action prediction probabilities a multi-task model produces) noise calibration and multi-objective decoding. In particular, if Xiaohongshu GateDiffInt redesigned the input sequence of the CVR head with diffusion and LLM intent distillation, then PTDG in effect redesigned the dependency structure between CVR heads. Statements that redefine the CVR head itself appearing from two companies recently is a pattern observed here.

However, the paper has only the CVR/eCPM lift, and did not disclose individual metrics for dense tasks (for example, click), the specific version of the baseline MMoE, or the sample count of the industrial dataset. We will watch whether these details come out in the CIKM 2026 presentation slides (2026-11-07–11, Rome).

Candidates also reviewed from arXiv cs.IR

Parsed 30 papers, and 1 passed: PTDG. The three below were reviewed because they have industrial production statements, but their domains do not match the canon's interest (feed ranking/retrieval), so they were not attached to threads.

  • AlleCompanion (arXiv 2609.05063, RecSys 2026 OARS workshop) is a complementary product recommendation system that Allegro, Poland's largest e-commerce company, deployed to production.
    • It filters out co-purchase noise with a two-tower structure (two encoders that embed the two inputs separately) and a Category Adapter, and ComCat creates category-level constraints.
    • ComCat combines expert rules, human-in-the-loop, LLM reasoning, and statistical mining.
    • We could not confirm the effect metrics because the abstract is cut off at the end.
    • Complementary recommendation is a category retrieval problem for basket-building surfaces (such as the carousel below the cart), so its domain differs from the organic feed and home surface ranking threads.
    • If a second company (such as a production statement on Amazon "Frequently bought together") makes a similar category-constrained two-tower statement, we will consider creating a domain thread then.
  • IGPO (arXiv 2609.04813, EMNLP 2026 Industry Track) is a training-free approach from the AI search of an anonymous commercial smart assistant.
    • In a production-deployed AI search pipeline, it learns inventory-grounded policy guidelines instead of fine-tuning or RL, and puts these guidelines into the prompt at serving time.
    • It has been deployed since May 2026, and a 14-day A/B gave CTR +3.17% and audited bad cases −38.9% verified.
    • It focuses on AI search (conversational search), and its domain differs from recsys feed ranking.
    • However, its problem statement that "fine-tuning/RL works poorly in the early deployment stage when the catalog is updated often" partly overlaps with recsys cold start.
    • If a similar policy-guideline statement appears from a second company, we will put it back on the review list.
  • SAM-D2Q (arXiv 2609.04961, CIKM 2026 Oral Full Paper) is a multimodal Doc2Query that Alibaba deployed in AliExpress production search.
    • Doc2Query is a document expansion technique that generates and attaches to each document pseudo-queries that would find it.
    • SAM-D2Q attaches visual signals to the product title to generate pseudo-queries, and aligns them to the search business objective (conversion) with RL-based preference alignment.
    • It is a document expansion problem in e-commerce search, so its domain differs from the feed ranking threads.
    • We record it only as the latest statement on Alibaba's search line.

These 4 are all the candidates with production deployment or A/B statements. The topics of the remaining 26 are as follows:

  • Domain-limited: medical, legal, baseball, music finance QA, essay grading, programming tutor, multi-omic visualization, web curation, phonetic encoding
  • Agentic memory benchmark
  • RAG improvements: CAGE, tree-based framework
  • Academic SID (semantic ID, a token sequence that represents an item as a hierarchical code) and multimodal recsys: EPIC diffusion follow-up, MURAL uncertainty edge learning, Latent-Aligned VLM, and others
  • embedding surgery
  • Programming tutor audit
  • Brand recommendation LLM audit

What to check next