Beyond Co-purchase Relation: Evolution of Complementary Recommendations at Allegro
Aleksandra Osowska-Kurczab, Klaudia Nazarko, Eliška Kosturová, Lidia Wojciechowska, Michał Bień
cs.IR · cs.AI · cs.LG
Quick verdict
Allegro's production Two-Tower+ComCat framework lifts GMV up to 21% at scale; solid industry report but only internal baselines and proprietary data.
Summary
This RecSys 2026 industry paper presents AlleCompanion, a production-scale complementary-product retrieval framework deployed at Allegro, serving over 20 million monthly active users. The system is a Two-Tower model with content features (title, price, taxonomy category, seller ID, and product attributes) trained on 90 days of co-purchase data with sampled softmax, mixed negative sampling, and temperature scaling. A Category Adapter concatenates the query embedding with a target complementary-category embedding, and an auxiliary category-reconstruction loss steers retrieval toward logically complementary categories. Co-purchase noise is reduced via heavy-buyer exclusion, a minimum pair-count threshold, and a category-alignment heuristic that rebalances an annotated 400-pair sample from 36% to 61% complementary pairs. ComCat, a decoupled multi-source category mapping combining automated co-purchase mining, human-in-the-loop annotations, expert rules, and a same-category relation, supplies target categories at inference; it covers 99.8% of live traffic despite 42% source-category coverage. Offline, AlleCompanion reaches Recall@20 of 0.4567 versus 0.0447 for a vanilla Two-Tower. Online A/B tests (two weeks, full traffic) show +8-9% GMV on organic product-page carousels, +15-21% GMV in cart, and roughly 50% higher sponsored ad revenue, though pre-cart placements were neutral or slightly negative. Limitations include internal-only baselines, proprietary data, and relative-only offline CTR/CVR reporting.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A category-constrained Two-Tower (Category Adapter plus auxiliary category-reconstruction loss) outperforms vanilla and hard-filtered Two-Tower variants on Recall@20 (0.4567 vs 0.0447 / 0.3782) while respecting target-category constraints (Table 1).
- The ComCat multi-source mapping covers 99.8% of live traffic despite only ~42% source-category coverage and lifts offline CTR by +450% and CVR by +398% relative to the heuristics-only mapping (Table 4).
- Training on the full filtered co-purchase distribution beats expert-restricted training data for relevance, so expert and LLM logic is best pushed into the mapping layer rather than the training set (Section 4.2, Table 3).
- Online A/B tests show +8-9% GMV on organic product-page carousels, +15-21% GMV in-cart, and roughly 50% higher sponsored ad revenue, leading to deployment on three placements serving 20M+ monthly users (Table 5).
Strengths
Full-lifecycle industrial evidence with honest negative results
Five online A/B experiments over two weeks at 100% traffic (Section 4.4, Table 5), including deployed wins (in-cart GMV +21.25%* Web / +15.73%* App) and candidly reported failures (pre-cart Filtered-Transactions variant: v-CVR -0.28%* on Web; neither pre-cart strategy deployed).
Data-noise analysis grounded in human annotation
Section 3.2.2 quantifies filter effects on a 400-pair labelled sample: raising pair-count thresholds mostly isolates substitutes (28%->43%) rather than complements (29%->36%), and the category-alignment heuristic rebalances the pool from 36/16/48% to 61/4/35% complementary/substitute/unrelated.
Decoupled mapping layer is a genuinely useful engineering contribution
Careful metric attribution and self-aware discussion
Weaknesses
No external baselines
Offline traffic gains reported only in relative terms
Thin statistical reporting around the A/B battery
Small annotation sample without agreement statistics
Coverage and consistency trade-offs underexplored
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This nine-page industrial paper (RecSys '26) describes AlleCompanion, a production complementary-product retrieval framework at Allegro.com, and ComCat, its decoupled category-mapping layer. Four contributions are claimed: a category-conditioned Two-Tower architecture, the multi-source ComCat mapping, an empirical study of dataset definitions, and lessons from offline ablations plus online A/B tests. The headline numbers are substantial: offline Recall@20 of 0.4567 vs 0.0447 for a vanilla Two-Tower (Table 1), and deployed A/B wins of +8-9% GMV on organic product-page carousels, +15-21% GMV in-cart, and roughly 50% sponsored ad-revenue growth (Table 5), serving 20M+ monthly active users.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is AlleCompanion?
AlleCompanion is Allegro's production-scale complementary-product recommendation system. It is a Two-Tower neural retrieval model whose Category Adapter steers candidate generation toward logically complementary categories (e.g., camera -> lens, not another camera), trained on 90 days of filtered co-purchase data and deployed across product-page and cart placements serving over 20 million monthly active buyers.
What is ComCat in this paper?
ComCat is a multi-source, directed category mapping that decides which complementary categories to retrieve from for a given query category. It merges automated co-purchase mining, human-in-the-loop annotations, rule-based expert logic, and an optional same-category (substitute) relation. It is decoupled from the model, so category policies can change without retraining, and it covers 99.8% of Allegro's live recommendation traffic.
How much business impact did Allegro measure?
In two-week, full-traffic A/B tests: GMV rose 8-9% on organic product-page carousels, 15-21% in the in-cart placement (statistically significant, p<0.005), and sponsored-placement ad revenue grew roughly 50%. Pre-cart placements were neutral or slightly negative and were not deployed. Offline retrieval improved from Recall@20 of 0.0447 (vanilla Two-Tower) to 0.4567.
Is the code or dataset for this paper publicly available?
No. The paper reports results on Allegro's proprietary transaction and catalogue data and does not release code, data, or complete hyperparameter settings, so the exact results cannot be reproduced externally.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure content (the Fig. 1 architecture diagram and Fig. 2 qualitative example screenshots) is image-only and absent from the text extraction; assessment of them relies on the prose descriptions, which are complete.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.