Dual-Stable Interest Exploration

• Dual-Stable Interest Exploration (DSIE) is a recommendation framework that pairs long-term group identity with short-term behavioral signals to mitigate feedback-loop issues.
• It supports multi-interest modeling and novelty control by aligning complementary preference views, ensuring both relevance and exploration in recommendations.
• DSIE is implemented through dual-phase, dual-level, or dual-scale mechanisms that optimize user representation reliability and promote effective discovery.

Dual-Stable Interest Exploration (DSIE) denotes a family of recommendation designs in which user preference is represented through two comparatively stable signals and then used to support exploration, multi-interest modeling, or novelty control. In "When Relevance Meets Novelty: Dual-Stable Periodic Optimization for Exploratory Recommendation" (Lin et al., 1 Aug 2025), DSIE is an explicit module that jointly models long-term group identity and short-term individual interests. In adjacent work, the same phrase is used as an interpretive description for dual-level LLM-guided multi-interest modeling in LDMI (Wang et al., 15 Jul 2025) and for dual-phase playtime-guided recommendation in DP$^2$Rec (Zhang et al., 9 Aug 2025). A separate sequential-recommendation paper uses the acronym DSIE for "Dual-Scale Interest Extraction," centered on global-scale inherent preference and local-scale multi-interest extraction rather than the later dual-stability framing (Chen et al., 2023). The recent literature therefore suggests that DSIE is best understood as a technical motif in recommender systems: stabilize interest discovery by pairing two complementary preference views rather than relying on a single user representation.

1. Terminology, scope, and problem setting

The immediate motivation for DSIE-style modeling is the well-known feedback-loop pathology of recommendation systems. Traditional recommenders learn from historical user-item interactions and then reinforce those same preferences, which limits exposure to new interests, induces content fatigue, and restricts discovery of latent preferences (Wang et al., 2024). In exploratory recommendation, this pathology appears as an imbalance between relevance and novelty; in multi-interest recommendation, it appears as collapsed or redundant interest vectors; in sparse domains, it appears as unstable grouping of behaviors into interests.

The technical literature identifies several recurring failure modes. Existing multi-interest models often rely on heuristic co-occurrence assumptions, predefine a fixed number of interests, and infer interest assignments from sparse interaction patterns, which can yield unstable or noisy partitions (Wang et al., 15 Jul 2025). Exploratory LLM frameworks can over-rely on short-term sequences and omit long-term group identity, which produces biased user modeling (Lin et al., 1 Aug 2025). In game recommendation, binary interaction views underutilize playtime and multimodal semantics, even though playtime carries fine-grained interest intensity and games exhibit cross-category semantic relations (Zhang et al., 9 Aug 2025). Sequential recommendation work on "Dual-Scale Interest Extraction" further argues that purely local multi-interest extraction ignores global correlation across the entire sequence and therefore misses a stable inherent preference (Chen et al., 2023).

A common source of ambiguity is nomenclature. "Dual-Stable Interest Exploration" refers explicitly to a user-modeling module in CoEA (Lin et al., 1 Aug 2025), but the same phrase is also used conceptually to describe other architectures with two stabilizing preference views (Wang et al., 15 Jul 2025, Zhang et al., 9 Aug 2025). By contrast, the acronym DSIE in (Chen et al., 2023) expands to "Dual-Scale Interest Extraction." The overlap is conceptual rather than terminological: both families oppose one-vector user modeling and both seek robustness through two coordinated preference representations.

2. Dual stability as a design principle

Across the cited literature, DSIE-like systems differ in implementation but share a common structural idea: the recommendation model becomes more reliable when interest exploration is anchored simultaneously by two complementary sources of stability.

Source	Duality	Stabilization mechanism
CoEA (Lin et al., 1 Aug 2025)	Long-term group identity + short-term individual interests	Group CSID and group profile text with recent categories
LDMI (Wang et al., 15 Jul 2025)	User-individual level + user-crowd level	Alignment with collaborative interests and synthesized-user contrastive learning
DP$^2$Rec (Zhang et al., 9 Aug 2025)	Strong interest phase + weak interest phase	Dual-beta mixture, dual graphs, and multimodal random walks
Dual-Scale Interest Extraction (Chen et al., 2023)	Global-scale inherent preference + local-scale multi-interests	Contrastive global encoder and preference-guided aggregation
DESMIL (Liu et al., 2022)	Representation stability + OOD stability	HSIC-based decorrelation with sample reweighting

In CoEA, the two dimensions of stability are explicit: group-level stability and short-term behavioral coherence. Users are clustered into collaborative semantic groups through long-term behavior, while the short-term sequence is mapped to a small set of recent categories; the model therefore treats a user as "their group identity plus their recent trajectory" (Lin et al., 1 Aug 2025). In LDMI, the duality is architectural rather than semantic: user-individual LLM clustering is stabilized by collaborative interests, and user-crowd modeling is stabilized by synthesized users selected for compactness and representativeness (Wang et al., 15 Jul 2025). In DP$^2$Rec, the two stabilities are probabilistic phases of preference intensity, estimated from playtime as strong versus weak interest (Zhang et al., 9 Aug 2025). In the earlier Dual-Scale Interest Extraction framework, the duality is scale-based: a global inherent preference regularizes local interests and supports novel recommendation (Chen et al., 2023).

This suggests that DSIE is less a single algorithm than a recurrent design principle. The first branch typically contributes persistence, collaborative regularity, or global semantics; the second branch contributes recency, personalization, exploration capacity, or local specificity. The central modeling problem is then no longer merely extracting multiple interests, but calibrating the interaction between these two branches so that the model neither collapses into a single dominant preference nor fragments into unstable micro-interests.

3. The explicit DSIE module in CoEA

The most literal formulation of DSIE appears in CoEA, where DSIE is the front-end user modeling module for exploratory recommendation (Lin et al., 1 Aug 2025). It processes long-term and short-term interaction histories in parallel and produces two outputs: a long-term Group Collaborative Semantic ID (CSID) with a textual group profile, and a short-term set of recent categories.

The long-term branch begins by splitting a user sequence into long-term and short-term parts. The long-term sequence is filtered to retain only items satisfying a cumulative-click threshold,

$$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$

with $\tau=5$ in experiments (Lin et al., 1 Aug 2025). Each retained item is embedded, user sparse features are embedded as $\mathbf{e}_{\text{sparse}}$, and the resulting sequence is encoded by an $L$-layer Causal Self-Attention stack. The long-term user representation is the last-position output,

$\mathbf{u}_{\text{long}}=\mathbf{H}^{(L)}_T.$

This vector summarizes stable preferences over filtered long-term interactions together with sparse features (Lin et al., 1 Aug 2025).

To convert continuous long-term preference into collaborative group identity, CoEA applies Residual-Quantized VAE. The encoder maps $\mathbf{u}_{\text{long}}$ into a latent $\mathbf{z}$, residual quantization assigns codebook vectors across $^2$0 levels, and the resulting discrete tuple

$^2$1

defines the user’s Group CSID (Lin et al., 1 Aug 2025). Users sharing the same CSID form a collaborative semantic group $^2$2. For interpretability, representative users closest to the group centroid are selected and their histories are given to a Profile LLM, which generates a textual group description such as “Tech enthusiasts: Prefer flagship mobile phones and foldable screen devices.” (Lin et al., 1 Aug 2025).

The short-term branch is deliberately simpler. Each short-term item is mapped by a deterministic category function $^2$3, then deduplicated into

$^2$4

The paper emphasizes that this representation focuses on the diversity of recently engaged categories rather than frequency, avoids heavy clustering at industrial scale, and provides up-to-date signals for novelty generation (Lin et al., 1 Aug 2025).

These dual outputs condition both LLMs in CoEA. The Novelty LLM receives prompts of the form

$^2$5

where $^2$6 is the group profile text, and generates candidate novel categories $^2$7. The Relevance LLM receives

$^2$8

scores each generated category, and filters by $^2$9 to obtain aligned exploratory categories (Lin et al., 1 Aug 2025). DSIE is therefore not an item encoder or a standalone retrieval model; it is a user-side conditioning mechanism that stabilizes exploratory recommendation by jointly exposing LLMs to long-term collaborative identity and short-term behavioral context.

4. Dual-level and dual-phase realizations

LDMI provides a dual-level realization of DSIE for multi-interest recommendation (Wang et al., 15 Jul 2025). At the user-individual level, an LLM groups a user’s engaged item titles into semantic clusters,

$^2$0

with GPT‑4o run at temperature $^2$1 for deterministic outputs (Wang et al., 15 Jul 2025). Independently, a capsule network produces $^2$2 collaborative interests $^2$3, where $^2$4 and $^2$5 is best in the experiments. LLM clusters are embedded by attention over item embeddings to form $^2$6, and an alignment module computes

$^2$7

The stated interpretation is that over-coarse clusters can be effectively split across interests, while over-fine clusters can be merged through shared alignment weights, producing stable semantically aligned interests (Wang et al., 15 Jul 2025).

The second LDMI branch operates at the user-crowd level. Similar users are aggregated into cliques, yielding synthesized users with dense behavior sequences. Because applying the LLM to all synthesized users would be costly and redundant, LDMI formulates a Max Covering Problem to select a subset $^2$8 of size $^2$9 that covers as many valuable items as possible. LLM-derived clusters on these synthesized users supervise a contrastive loss,

$$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$0

which enforces intra-cluster compactness and inter-cluster separation in the shared item-embedding space (Wang et al., 15 Jul 2025). In this reading, DSIE consists of individual-level granularity control and crowd-level representation disentanglement.

DP$$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$1Rec instantiates DSIE differently, as a dual-phase model of interest intensity inferred from playtime (Zhang et al., 9 Aug 2025). Playtime is normalized per game via percentile rank,

$$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$2

and each user’s normalized playtimes are modeled by a two-component Beta mixture,

$$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$3

The posterior $$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$4 is the probability that an interaction belongs to the strong-interest component, which the paper interprets as stable, core hobbies, while the weak component captures ephemeral or exploratory interests (Zhang et al., 9 Aug 2025). These two phases define a full graph $$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$5 and a strong-interest graph $$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$6, and the model applies LightGCN propagation plus cross-view contrastive learning to fuse the two signals.

DP$$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$7Rec then uses these phases to guide exploration through playtime-guided multimodal random walks. Walks start from strong-interest games sampled according to normalized playtime, construct category-aware candidate sets using multimodal similarity, and transition with probabilities proportional to an aggregate of multimodal similarity, interest similarity, and a category-balance coefficient (Zhang et al., 9 Aug 2025). The paper characterizes the resulting category-coverage process as a submartingale, because the expected increase in covered categories is non-negative. This is a different DSIE semantics from CoEA or LDMI, but it preserves the same core intuition: exploration should be anchored by a stable signal and then expanded along a controlled secondary signal.

An important precursor is the hierarchical LLM framework for user interest exploration in large-scale recommendation (Wang et al., 2024). It does not use the DSIE name, but it fixes a stable interest space through a 4-level tree of traffic-balanced clusters, fine-tunes an LLM to generate novel cluster descriptions from recent cluster history, and constrains a transformer-based item policy to return items only from the selected cluster. Novelty is defined at the cluster level, not the item level, and offline precomputation yields a deterministic mapping $$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$8 from recent cluster pairs to novel clusters (Wang et al., 2024). This suggests a broader DSIE-compatible view in which stability can reside in the interest space itself, rather than only in user representations.

5. Objectives, optimization strategies, and representation learning

The optimization machinery of DSIE-style systems is heterogeneous, but nearly all variants combine a task loss with an auxiliary stability-inducing loss.

In LDMI, recommendation uses hard routing over aligned interests,

$$I'_{\text{long}}=\left\{I_t^l \in I_{\text{long}} \mid \text{click}_u(I_t^l)\ge \tau\right\},$$9

with an InfoNCE-style recommendation objective $\tau=5$0. The total loss is

$\tau=5$1

where $\tau=5$2 is the synthesized-user contrastive loss and $\tau=5$3; gradients from $\tau=5$4 are updated every $\tau=5$5 iterations to control computation (Wang et al., 15 Jul 2025). Stability is therefore enforced by periodic global clustering pressure on shared item embeddings.

In DP$\tau=5$6Rec, the Interest Intensity Exploration module fits the dual-Beta mixture per user via EM, producing $\tau=5$7 and dual graph views. Cross-view alignment uses user-side and item-side contrastive losses between embeddings learned on $\tau=5$8 and $\tau=5$9, while the full model fuses IIE and MRW embeddings as

$\mathbf{e}_{\text{sparse}}$0

and trains with a balance-aware ranking loss $\mathbf{e}_{\text{sparse}}$1 plus self-supervision $\mathbf{e}_{\text{sparse}}$2 (Zhang et al., 9 Aug 2025). Here the stable branch is probabilistic and graph-based, while the exploratory branch is stochastic and multimodal.

CoEA couples DSIE to Periodic Collaborative Optimization rather than to a single end-to-end retrieval loss (Lin et al., 1 Aug 2025). The Novelty LLM is supervised on category prediction with

$\mathbf{e}_{\text{sparse}}$3

the Relevance LLM is trained with a pairwise ranking loss

$\mathbf{e}_{\text{sparse}}$4

and periodic refinement of the Novelty LLM uses Direct Preference Optimization,

$\mathbf{e}_{\text{sparse}}$5

The closed loop is explicit: the Relevance LLM verifies preferences on incremental data, guides fine-tuning of the Novelty LLM, and then re-evaluates the updated generator (Lin et al., 1 Aug 2025). DSIE is therefore embedded in an evolving optimization cycle rather than treated as a frozen front-end.

Related work broadens the space of stabilization mechanisms. Dual-Scale Interest Extraction learns a global inherent preference $\mathbf{e}_{\text{sparse}}$6 from the entire sequence via self-attention and a BPR-style contrastive loss between original and shuffled sequences, then uses $\mathbf{e}_{\text{sparse}}$7 to guide local multi-interest extraction and soft aggregation (Chen et al., 2023). DESMIL instead focuses on stability under distribution shift: it measures dependencies among interest vectors with the Hilbert-Schmidt Independence Criterion and performs HSIC-based sample reweighting so that the model relies less on subtle, non-causal dependencies among interests (Liu et al., 2022). A plausible implication is that DSIE-like systems can be stabilized at several levels—sequence encoding, interest alignment, graph regularization, or sample selection—without changing the basic dual-view premise.

6. Empirical performance, limitations, and interpretive issues

The available empirical evidence is consistently favorable, though it comes from different tasks and metrics. LDMI reports superiority over single-interest baselines, strong multi-interest baselines, and the LLM-based baseline EIMF on Beauty, Books, and Video Games; on Beauty, the best baseline EIMF reaches $\mathbf{e}_{\text{sparse}}$8, while LDMI reaches $\mathbf{e}_{\text{sparse}}$9, and removing semantic LLM clusters drops $L$0 to $L$1 (Wang et al., 15 Jul 2025). DP$L$2Rec reports that its accuracy-focused setting reaches $L$3, above SMORE at $L$4, while its diversity-focused setting reaches $L$5, above DGRec at $L$6; ablations removing dual-interest recognition or category-based candidate selection reduce both accuracy and diversity (Zhang et al., 9 Aug 2025). CoEA reports on Movielens-1M that full performance reaches $L$7, $L$8, $L$9, and $\mathbf{u}_{\text{long}}=\mathbf{H}^{(L)}_T.$0, and that removing the long-term branch causes a large drop in quality, while a 20-day A/B test on Meituan yields $\mathbf{u}_{\text{long}}=\mathbf{H}^{(L)}_T.$1 and $\mathbf{u}_{\text{long}}=\mathbf{H}^{(L)}_T.$2 (Lin et al., 1 Aug 2025). Dual-Scale Interest Extraction, despite using a different expansion of DSIE, similarly reports consistent gains over MIND, ComiRec, and SINE on Amazon benchmarks, including Video Games Recall@50 improving from $\mathbf{u}_{\text{long}}=\mathbf{H}^{(L)}_T.$3 to $\mathbf{u}_{\text{long}}=\mathbf{H}^{(L)}_T.$4 (Chen et al., 2023). The industrial LLM interest-exploration framework also reports live gains in novel-interest exploration and overall user enjoyment, though the provided summary does not enumerate exact numbers (Wang et al., 2024).

The limitations are similarly recurrent. LLM-driven variants depend on LLM clustering or prompt quality and incur offline preprocessing cost; LDMI uses GPT‑4o, precomputes and caches cluster assignments, and still notes costs for individual calls, crowd-level MCP, and contrastive learning on dense synthetic sequences (Wang et al., 15 Jul 2025). CoEA depends on group definitions induced by RQ‑VAE and raises issues of complexity, update lag, bias, and fairness in behavior-based grouping (Lin et al., 1 Aug 2025). DP$\mathbf{u}_{\text{long}}=\mathbf{H}^{(L)}_T.$5Rec assumes exactly two Beta components, does not model temporal order directly, and fits EM per user, which may be heavy at platform scale (Zhang et al., 9 Aug 2025). DESMIL shows that multi-interest models can overfit to inter-interest dependence and become unstable under distribution shift, motivating explicit control of dependence rather than assuming multiple interests are automatically disentangled (Liu et al., 2022).

One misconception is to treat DSIE as a single standardized architecture. The literature does not support that reading. In CoEA, DSIE is a named module with long-term group identity and short-term categories (Lin et al., 1 Aug 2025). In LDMI and DP$\mathbf{u}_{\text{long}}=\mathbf{H}^{(L)}_T.$6Rec, it functions as a conceptual lens for dual-level or dual-phase stabilization (Wang et al., 15 Jul 2025, Zhang et al., 9 Aug 2025). In (Chen et al., 2023), DSIE is an acronym for a different phrase, "Dual-Scale Interest Extraction." Another misconception is to equate stability with rigidity. In all of these systems, stability is introduced to make exploration more trustworthy: aligned LLM clusters, strong-interest seeds, group identities, or decorrelated interests are used precisely so that novelty, serendipity, or cross-category discovery can be pursued without collapsing relevance.

Taken together, the literature presents DSIE as a response to the inadequacy of single-view user modeling in modern recommendation. Whether implemented through Group CSIDs and recent categories, LLM semantic clusters and synthesized users, dual-Beta intensity phases, global and local sequence encoders, or HSIC-controlled multi-interest learning, the shared technical claim is that exploration becomes more effective when the model anchors user preference in two coordinated stable structures rather than in a monolithic embedding.