Search-based Interest Model

• SIM is a framework that models user interests by retrieving a relevant subsequence from massive behavior histories and aggregating them for enhanced personalization.
• The model addresses scalability and noise issues with a two-step process: a coarse search for high-relevance behaviors followed by fine-grained attention-based aggregation.
• SIM demonstrates superior performance in industrial applications, achieving notable CTR uplifts and latency reductions in display advertising and search ranking deployments.

A Search-based Interest Model (SIM) operationalizes user interest modeling for large-scale sequential recommendation and search ranking tasks via a two-stage “search-and-aggregate” mechanism. SIMs address the scalability and noise issues intrinsic to lifelong user behavior sequences (often tens of thousands of events), by first searching for behaviors relevant to the current query or candidate item, then modeling user interests over the resulting compact, high-signal subsequence. This cascading design allows SIMs to achieve both real-time efficiency and superior personalization on industrial systems, as demonstrated in display advertising and search ranking at Alibaba and Kuaishou (Qi et al., 2020, Guo et al., 2023). Modern instantiations of SIM incorporate multi-modal, multi-behavior, and disentangled representations, further enhancing accuracy and interpretability (Shen et al., 15 Jul 2024, Si et al., 2023).

1. Foundational Principles and Motivation

SIMs emerged to solve two core challenges in extracting actionable user interests from lifelong behavioral data:

1. Scalability: Naive attention or sequence models are computationally infeasible for sequences of length $T \gg 10^3$ due to $O(Td)$ or $O(T^2 d)$ complexity.
2. Relevance Focusing: Most user behaviors are irrelevant to the current query or candidate, so modeling all equally introduces noise and dilutes signal.

The SIM paradigm formalizes user interest modeling as a search-then-aggregate pipeline:

• Search: Select a small, high-relevance subsequence from the user's behavior history, using item-aware or query-aware similarity scores.
• Aggregate: Model the interaction between the target (candidate item, query) and the selected behavioral subsequence using deep attention or fusion mechanisms.

This explicit separation enables sub-linear serving complexity, handles sequences $T \approx 10^4$–$10^5$, and improves recommendation/search accuracy by suppressing behavioral noise (Qi et al., 2020, Guo et al., 2023).

2. Core Architectural Components

SIM implementations share a common two-stage structure, though instantiations vary:

General Search Unit (GSU) / Relevance Search Unit (RSU)

• Purpose: Coarse search to select top-$K$ most relevant historical behaviors with respect to the current candidate (and, in search, the user query).
• Mechanisms:
  • Hard-search: Category or ID-matching, e.g. $r_i = \mathbf{1}_{C_i = C_a}$.
  • Soft-search: Learned similarity (e.g., inner product of projected embeddings).
  • Two-stage search: First coarse relevance to the query, then refinement to the candidate item (Guo et al., 2023).
  • Multi-modal or cross-behavioral expansion: Incorporate queries, item content, images, attributes (Shen et al., 15 Jul 2024, Si et al., 2023).

Exact Search Unit (ESU) / Fused Attention Unit (FAU)

• Purpose: Model precise relationships between candidate and selected behaviors.
• Mechanisms:
  • Multi-head attention using candidate embedding as query and filtered behaviors as keys/values.
  • Fusion of multiple modalities or behavior types, with decoupled attention over item IDs and attributes, often modulated by engagement-based gating (Guo et al., 2023).
  • Aggregation into a fixed-size user-interest vector for prediction.

Component	Function	Example Paper
GSU / RSU	Coarse search & filtering	(Qi et al., 2020, Guo et al., 2023)
ESU / FAU	Fine-grained interest modeling (attention)	(Qi et al., 2020, Guo et al., 2023)
PQ/ANN Search	Efficient top-$K$ retrieval (optional)	(Shen et al., 15 Jul 2024)

3. Mathematical Formulations and Objective Functions

SIMs employ a suite of mathematical tools to realize the search-and-aggregate paradigm:

Search Scoring:

• Soft-search similarities: $$r_i = \langle W_b \mathrm{Embed}(b_i), W_a \mathrm{Embed}(a) \rangle$$.
• Two-level relevance:
  • Query relevance: $r_{b_t}^q = (e_q W^Q)(e_{b_t} W^K)^\top / \sqrt d$.
  • Candidate item relevance on the subset: $$r_{b}^i = (e_i W^{Q'})(e_{b} W^{K'})^\top / \sqrt d$$ (Guo et al., 2023).

Attention-based Aggregation:

• Multi-head item-aware attention: For each head $h$,

$$q_h = W_{ah} e_a,\;\; k_h = Z W_{bh}^T,\;\; A_h=\mathrm{softmax}(k_h q_h^T),\;\; u_h = A_h^T Z$$

with outputs concatenated and fed to an MLP (Qi et al., 2020).

Multi-modal Fusion and Alignment:

• Modality projections and fusion: $\tilde Q = W_q Q,\,\tilde T = W_t T,\,\ldots$
• Multi-modal attention score:

$$q_t = \sum_{m} \gamma_m x_t^{(m)},\quad k_\ell = \sum_{m} \gamma_m x_\ell^{(m)}$$

• Multi-modal contrastive alignment and InfoNCE losses (Shen et al., 15 Jul 2024, Si et al., 2023).

Training Losses:

• Cross-entropy for CTR or ranking: $$L_{\mathrm{CTR}} = \sum_i \mathrm{CE}(\hat y_i, y_i)$$.
• Multi-task setups: combinations of cross-entropy, alignment, relevance, and triplet losses (Si et al., 2023).

4. Industrial-Scale Deployment and System Insights

SIMs are specifically architected for low-latency, high-throughput serving in large-scale production environments:

• Alibaba deployment: SIM supports behavior lengths up to $T=54{,}000$, outperforming precedent models restricted to $T \leq 1000$. System improvements include a two-level User Behavior Tree (UBT) structure for category-based hashing (22 TB), yielding sub-millisecond GSU lookups by category and <5 ms additional serving latency over previous memory-network models (Qi et al., 2020).
• Kuaishou production search: QIN’s (a SIM variant) RSU executes two cascaded approximate nearest neighbor (ANN) searches (query then item) over large user histories ($N \approx 10^4$), with FAU and MLP for final ranking, all under 30 ms SLA (Guo et al., 2023).
• Latency reduction via approximate retrieval: Multi-modal product quantization reduces attention/retrieval latency by up to 10×, with >95% recall at online scale (Shen et al., 15 Jul 2024).

5. Extensions: Multi-Modal, Multi-Behavior, and Disentangled Interests

SIM frameworks are extensible to multi-modal and multi-behavioral contexts:

• SEMINAR (Shen et al., 15 Jul 2024) incorporates search queries, item text, images, and structured attributes, with a multi-modal pretraining search unit (PSU) optimizing across alignment, next-pair, and query–item relevance tasks. Product quantization allows efficient multi-modal attention at scale.
• SESRec (Si et al., 2023) disentangles user interest into “similar” (reinforcing S&R behaviors) and “dissimilar” (novelty-seeking) parts, extracting both via transformer co-attention and triplet losses, and fusing these components for next-item prediction.
• QIN (Guo et al., 2023) uses engagement-based gating and decoupled attention over multiple modalities, further enhancing representational capacity and model interpretability.

These architectures support new industrial requirements, such as:

• Efficient utilization of lifelong sequences (billions of interactions).
• Alignment and de-duplication of user intent signals across diverse modalities.
• Rich explanation of user interest profiles in complex ecosystems.

6. Empirical Performance and Comparative Evaluation

SIM-based models have demonstrated consistent state-of-the-art results across diverse datasets and industrial benchmarks:

• Alibaba display advertising: SIM achieved +7.1% CTR and +4.4% RPM vs. MIMN, with offline AUC improvement from 0.6541 (MIMN) to 0.6604 (SIM hard-search) and 0.6625 (SIM soft-search) (Qi et al., 2020).
• Kuaishou search: QIN delivered 7.6% CTR uplift in online A/B tests, and 24.1% increase in “efficient view” rate (Guo et al., 2023).
• SESRec outperformed both single-stream sequential and prior search-aware recommendation methods, raising NDCG@10 from 0.3787 (best sequential) and 0.3762 (best search-aware) to 0.4054 (Si et al., 2023).
• SEMINAR’s multi-modal product quantization achieved a 10× latency reduction (from ∼20 ms to 1.5 ms per user) while maintaining recall@64 ≈ 0.98 (Shen et al., 15 Jul 2024).

Model	Domain	Online CTR Gain	Offline Metric
SIM	Display Ad (Alibaba)	+7.1%	AUC +0.0084
QIN	Search (Kuaishou)	+7.6%	NDCG@4 +18–29%
SESRec	E-comm/video rec	–	NDCG@10 +7%
SEMINAR	Short-video rec	–	Recall@64 0.98

7. Broader Implications and Open Design Questions

SIMs redefine user modeling by enabling fine-grained interest extraction from lifelong, multi-modal, and multi-intent histories. Key implications include:

• Two-stage search with query and item context is crucial for high-precision, scalable personalization.
• Decoupling heterogeneous signals (ID vs. content, S&R behaviors, image vs. text) and multi-modal alignment are vital for avoiding representational bottlenecks.
• Engagement and fine-grained gating mechanisms dynamically weight the importance of past behaviors, facilitating nuanced recommendation and ranking.

Notable limitations and open avenues:

• ANN index maintenance introduces engineering complexity; generalizing to real-time interest and negative feedback remains underexplored (Guo et al., 2023).
• Multi-modal product quantization and alignment require careful hyperparameterization and infrastructure for extreme scale (Shen et al., 15 Jul 2024).
• The integration of session-level intent shifts and non-click behaviors (likes, comments) offers future research potential.

The SIM paradigm thus underpins a new class of industrial recommender and search systems, balancing accuracy, interpretability, and system efficiency across unprecedented behavioral sequence lengths and modal diversity (Qi et al., 2020, Guo et al., 2023, Shen et al., 15 Jul 2024, Si et al., 2023).