PaperlayerView exact revision on arXiv

Source-linked AI summary

PRISM: Personalized Item Representations in User History

Renzhi Wu, Zikun Cui, Junjie Yang, Li Sheng, Yongjun Xie, Jerry Fu, Nipun Mathur, Sri Reddy, Hong Li

arXiv:2610.39175v1cs.IRcs.AI

TL;DR

Shared item embeddings do not express how an item relates to a particular user and encounter, motivating personalized history inputs. PRISM reuses upstream ranking computation to create compact event embeddings, and on Surface 1 reports an average downstream NE change of −0.15% versus −0.07% for VISTA at similar history lengths.

  • Problem

    Shared item embeddings give the same item the same initial representation across users, despite differing user–item relationships.

    Source: §1 · p. 1

  • Method

    PRISM reuses upstream ranking computation and jointly compresses states from its user, user–item, and same-request candidate components into personalized event embeddings for downstream sequence modeling.

    Source: §1 · p. 2

  • Results

    −0.15% average downstream normalized entropy (NE) change for PRISM, compared with −0.07% for VISTA at similar history lengths on Surface 1.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Personalized history length provides an empirical scaling dimension for improving recommendation quality.

    Source: §1 · p. 2

Abstract

from arXiv · show

Sequential recommenders typically construct history inputs from item identifiers or content features using an embedding map shared across users. The same item therefore yields the same initial embedding representation, even when users have different relationships to the item. We propose personalized item representations in user history, where each item's embedding expresses its relationship to the user. Personalization becomes a property of the history inputs themselves, supplying richer evidence for later sequence modeling. We instantiate this idea with PRISM, which reuses upstream ranking computation to generate personalized item embeddings. An autoencoder jointly compresses all layer states from a user tower, a user--item tower, and same-request candidate mixing into one embedding per scored pair for downstream replay. Offline evaluation on Surface 1, a major recommendation surface at billion scale, reports an average downstream normalized entropy (NE) change of $-0.15\%$, compared with $-0.07\%$ for VISTA at similar history lengths. We further identify an empirical sequence-length scaling law: over history lengths of 26--240 events, NE gain follows a fitted power law with exponent 0.468, increasing from 0.06\% to 0.17\%. Therefore, personalized item representations in user history provide a new scaling dimension for improving recommendation quality.

1 Introduction

Shared item embeddings do not express how an item relates to a user or encounter. PRISM creates personalized history inputs by reusing upstream ranking computation, and its Surface 1 evaluation reports a downstream NE change of −0.15% versus −0.07% for VISTA at similar history lengths.

1 Introduction

  • A shared initial embedding cannot express different relationships to an item, so PRISM personalizes history entries for each user and encounter.Even repeated events involving the same user and item can receive different representations as interests and context evolve.Source: §1 · pp. 1–2
  • PRISM reuses upstream ranker computation by retaining a compact embedding per scored pair and replaying embeddings chronologically as history.This amortizes generation computation across later reads.Source: §1 · pp. 1–2
  • Its user tower, user–item tower, and same-request candidate mixing capture user state, item relation, and immediate session context.A detached autoencoder jointly compresses hidden states from every layer, including candidate-specific and pooled session states, into one personalized embedding.Source: §1 · p. 2
  • −0.15% average downstream normalized entropy (NE) change for PRISM, compared with −0.07% for VISTA at similar history lengths on Surface 1.The evaluation covers a major recommendation surface at billion scale.Source: §1 · p. 2
  • At similar history lengths, PRISM achieves an average normalized entropy (NE; lower is better) change of −0.15%, versus −0.07% for VISTA; NE gain follows fitted power law g(L) ∝ L^0.468.Personalized item representations in user history provide a scaling dimension for improving recommendation quality.Source: §1 · p. 2

2 Related Work

PRISM supplies item representations specific to each historical encounter, whereas related approaches include models that learn user-dependent states from behavioral inputs and systems that reuse other intermediate representations.

2 Related Work

  • PRISM differs from sequence encoders and interest models by supplying an item representation specific to each historical encounter with the user.The passage contrasts this with models that learn user-dependent states from behavioral inputs and CMSL’s multiple latent history contexts.Source: §2 · p. 3
  • VISTA caches user-history summary tokens, while OneTrans reuses intermediate states through cross-request KV caching.PRISM instead stores a compact code for each scored user–item pair, retaining user and context evidence at each encounter.Source: §2 · p. 3
  • PRISM jointly compresses states from every layer of its user, user–item, and same-request components into personalized history tokens for later supervised reading.The passage relates this design to intermediate-representation methods and autoencoder uses including dimensionality reduction and collaborative filtering.Source: §2 · p. 3

3 Method

PRISM generates fixed-width codes from event-time user, candidate, and request inputs, then compresses multilevel ranking states into representations for downstream history modeling. Its method combines user and pair towers, same-request session mixing, joint compression, and replay through a candidate-conditioned reader.

3.1 User and User–Item Towers

  • The user tower encodes history and user features; its final state enters pair-state learning, while multitask prediction losses train both towers.Source: §3.1 · p. 4

3.2 Session Mixing Within a Request

  • The session mixer jointly processes detached pair states, yielding candidate-specific context and pooled request summaries used for item- and aggregate-level predictions.Source: §3.2 · p. 4

3.3 Multi-Level Hidden-State Compression

  • Joint compression combines user, pair, and session states across layers into one event-specific code; its target width is independent of request candidate count.Source: §3.3 · p. 4

3.4 Downstream Sequence Modeling

  • Downstream replay retains eligible prior-event codes under time, availability, retention, and length constraints, adds time and event encodings, and lets a candidate-conditioned reader summarize them.Source: §3.4 · p. 5

4 Experiments

On Surface 1, PRISM history improves downstream normalized entropy across engagement groups, and longer retained histories show increasing gains with diminishing returns.

4.1 Setup and Metrics

  • On Surface 1 production data, PRISM history was added alongside ordinary features and evaluated using NE, where lower is better.PRISM exports embeddings during 14 training days; downstream models train on the following three production-data days and evaluate on the next day's held-out data.Source: §4.1 · p. 5

4.2 Downstream Ranking Results

  • −0.15% combined average NE change versus the Surface 1 ranking baseline; positive-engagement tasks change −0.17% and negative-engagement tasks −0.11%.Lower NE is better.Source: §4.2 · p. 5, Table 1 · p. 6

4.3 Scaling Law with Sequence Length

  • NE improvement rises from 0.06% at 26 events to 0.17% at 240, fitting g(L) ≈ 0.0131L^0.468 with diminishing gains per added event.Within the tested 26–240-event range, doubling history length corresponds to approximately 1.38× the NE improvement.Source: §4.3 · p. 6, Figure 3 · p. 6

5 Conclusion

The conclusion presents personalized history as a scaling dimension for recommendation: PRISM encodes encounter-specific user–item information, and Surface 1 evaluation and a history-length sweep support this direction.

5 Conclusion

  • Personalized history inputs express how an item relates to a particular user at a particular encounter.This makes the history representation encounter-specific.Source: §5 · p. 7
  • PRISM reuses upstream ranking computation, jointly compressing every layer's user-tower, user–item-tower, and candidate-mixer states into compact event embeddings.The embeddings are used for downstream sequence modeling.Source: §5 · p. 7
  • On Surface 1, PRISM's average NE change is −0.15% versus −0.07% for VISTA at similar history lengths, while a 26–240-event sweep shows a power-law trend.The observed trend has diminishing gains per additional event.Source: §5 · p. 7

Related papers

Loading 2610.39175v1…