acceptodds
Under review as a conference paper at ICLR 2027

SDBC-RO: STRONG-DEPENDENCY BEHAVIOR CHAIN GUIDED MULTI-VIEW BEHAVIOR SEMANTICS-AWARE REPRESENTATION OPTIMIZATION

Abstract

With the continuous expansion of data scale and increasingly diverse interaction scenarios on e-commerce platforms, accurately capturing users’ true preferences and decision intentions from massive heterogeneous behavioral data has become a central challenge in Multi-Behavior Recommendation (MBR). Current mainstream MBR methods have an inherent flaw: their modeling relies heavily on predefined behavioral partial order relations, which largely deviate from users’ real behavior patterns. This easily causes over-strengthening of auxiliary behaviors and marginalization of target behaviors, ultimately leading to representation distortion. In addition, existing methods typically fail to fully leverage both users’ behavioral habits and category preferences, resulting in incomplete semantic representations. To address the above issues, we propose Strong-Dependency Behavior Chain Guided Multi-View Behavior Semantics-Aware Representation Optimization (SDBC-RO). We first define a strong-dependency behavior chain (SDBC) in the form of ”auxiliary behavior set set > target behavior”. This design avoids the cumulative interference from multi-level auxiliary behaviors while preserving the dominant role of the target behavior. Meanwhile, we jointly leverage large language models and knowledge graphs to model behavior semantics and behavioral partial order relations, so as to generate high-quality behavior representations. On this basis, we develop a two stage multi-view representation optimization framework guided by SDBC. In the former stage, a two-level heterogeneous hypergraph is constructed from user–item–category ternary interactions to learn the basic representations of users, items, and categories. In the latter, behavior semantics-aware knowledge graphs are constructed from user–user, item–item, and category–category relational views to deeply integrate behavior semantics with the three types of representations. Finally, adaptive cross chain fusion and category-aware representation enhancement are employed to improve recommendation performance. Extensive experiments on two real world benchmark datasets, Tmall and Taobao, demonstrate that SDBC-RO substantially outperforms state-of-the-art baselines, achieving relative improvements of 17.78% and 10.09% in HR@10 and NDCG@10 on Tmall, and 23.21% and 3.68% on Taobao, respectively

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
multi-behavior recommendation, strong-dependency behavior chain, two-level heterogeneous hypergraph, multi-view behavior semantics-aware knowledge graph

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FgxUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SDBC-RO for multi-behavior recommendation. It groups interactions by auxiliary-behavior combinations and learns user, item, and category representations through two-level heterogeneous hypergraphs. Behavior text embeddings and multiple knowledge-graph views further refine these representations, which are combined through attention and category-aware fusion. Experiments on Tmall and Taobao report improvements over fourteen baselines, supported by component ablations and parameter analyses.

Critical strengths
  1. The method addresses heterogeneous paths to purchasing by distinguishing different auxiliary-behavior combinations and direct purchases.

  2. Table 2 reports the prevalence and conversion rates of behavior groups, providing useful motivation for modeling them separately.

  3. The experiments include two datasets, fourteen baselines, and seven ablation variants, offering reasonably broad coverage of the proposed components.

Critical weaknesses
  1. The behavior graph in Section 3.3.1 includes both directions between every pair of distinct behaviors under one relation. As written, it does not represent a partial order and leaves no non-self negative triples under head/tail corruption. If this is abbreviated notation, the actual edge construction and negative-sampling rules are missing.

  2. The paper does not specify the data split, ranking candidates, or temporal boundaries for graph construction. Table 3 also needs clarification: with one test positive per user and identical ranked lists and users, the Taobao NDCG increase from K=10 to K=20 cannot exceed 0.0048/log2(12), approximately 0.00134, whereas the reported increase is 0.0198. The unspecified protocol prevents determining whether this bound applies.

  3. The meaning of "strong dependency" is not operationally defined. Moreover, w/o Adaptive-Chain changes interaction grouping and propagation structure together by replacing the grouped architecture with one chain. It therefore does not isolate the proposed mechanism or substantiate the claim that two-level propagation eliminates preference superposition.

Questions for authors
  1. Which directed behavior edges are actually used, how are they determined, and what triples are eligible negative samples?

  2. What data-splitting and ranking protocol is used, and how does it explain the Taobao HR and NDCG changes between K=10 and K=20?

  3. How is "strong dependency" determined during chain construction, and must auxiliary actions occur before the target interaction?

1 comment

1:We will supplement the exact positive-edge set and negative sampling filtering rules in the method description of Section 3.3.1, to ensure the optimization objective is fully specified and consistent with the actual implementation. In our actual implementation, the only directed positive edges are Click > Cart and Cart > Buy, encoded as the triplets (Click, r, Cart) and (Cart, r, Buy). They represent the direct progressive relation between behaviors, reflecting the transition from initial interest to stronger purchase intention, and no reciprocal edges are included. For TransE negative sampling, we corrupt each positive triplet by replacing either its head or tail entity with a candidate drawn from {Click, Cart, Buy}. Candidate triplets that coincide with ground-truth positive triplets or form invalid self-loops are filtered out, and the remaining corrupted triplets are used as negative samples. We have explicitly specified this sampling and filtering rule in Section 3.3.1 of the revised manuscript. 2:We thank the reviewer for raising this question. Our data splitting and ranking protocol follow the widely adopted conventional paradigm in multi-behavior recommendation: we split the dataset using the standard leave-one-out strategy, where one target behavior interaction of each user is held out for testing and the remaining interactions are used for training. For evaluation, we generate a unified ranking list for each user and compute HR@K and NDCG@K on this list, with only the truncation threshold K set to 10 and 20 separately. The previously observed inconsistency in metrics arose because the early statistics aggregated the maximum values across different experimental runs. We have now standardized our reporting to present averaged results over multiple independent runs under identical settings. The revised Taobao results in Section 4.3 of the revised manuscript are HR@10=0.2415, HR@20=0.2463, NDCG@10=0.1215, NDCG@20=0.1227. The NDCG increment from K=10 to K=20 is 0.0012, which is below the theoretical upper bound of approximately 0.00134 under this evaluation scheme and thus reasonable. Critical Weakness 3:We agree with the two insightful comments raised by the reviewer. First, in Section 3.2.1 of the revised manuscript, we will explicitly operationalize the concept of "strong dependency" as structural relations centered on target behaviors. For each user-item interaction pair, we only incorporate auxiliary behaviors that occur before the first Buy action, based on which four types of behavior chains are constructed: Click > Buy, Cart > Buy, Click & Cart > Buy, and Only Buy. For the Click & Cart > Buy chain, we do not impose a fixed order between Click and Cart as long as both precede the Buy behavior. Second, the current w/o Adaptive-Chain variant in Section 4.4 of the revised manuscript essentially serves as a composite comparison. When replacing the SDBC-based grouping architecture with a cascade strategy similar to DA-GCN, both the interaction grouping logic and the message-passing mechanism change simultaneously. Accordingly, this variant should not be described as a strict single-variable ablation for a single propagation path. Instead, this set of comparisons verifies that under the same multi-behavior propagation paths, the strong-dependency partitioning mechanism of our model can effectively alleviate the interest shift caused by preference accumulation in existing cascade paradigms. Question 3:We agree with the two insightful comments raised by the reviewer. First, in Section 3.2.1 of the revised manuscript, we will explicitly operationalize the concept of "strong dependency" as structural relations centered on target behaviors. For each user-item interaction pair, we only incorporate auxiliary behaviors that occur before the first Buy action, based on which four types of behavior chains are constructed: Click > Buy, Cart > Buy, Click & Cart > Buy, and Only Buy. For the Click & Cart > Buy chain, we do not impose a fixed order between Click and Cart as long as both precede the Buy behavior. Second, the current w/o Adaptive-Chain variant in Section 4.4 of the revised manuscript essentially serves as a composite comparison. When replacing the SDBC-based grouping architecture with a cascade strategy similar to DA-GCN, both the interaction grouping logic and the message-passing mechanism change simultaneously. Accordingly, this variant should not be described as a strict single-variable ablation for a single propagation path. Instead, this set of comparisons verifies that under the same multi-behavior propagation paths, the strong-dependency partitioning mechanism of our model can effectively alleviate the interest shift caused by preference accumulation in existing cascade paradigms.

Reviewer_ayxVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SDBC-RO, a multi-behavior recommendation framework designed to improve user preference modeling by capturing dependencies between auxiliary and target behaviors. The paper argues that existing methods rely heavily on predefined behavior orders and fail to fully capture users' behavioral patterns and category preferences. To address these limitations, SDBC-RO introduces strong-dependency behavior chains and integrates multi-view behavioral and semantic information to enhance user and item representations. Experiments on two real-world recommendation datasets, Tmall and Taobao, demonstrate improvements over existing baselines in recommendation performance.

Critical strengths
  1. The paper considers diverse user interaction patterns rather than assuming that all users follow a fixed sequence of behaviors, which is a reasonable direction for improving preference modeling in multi-behavior recommendation.

  2. The integration of behavioral information and category-level preferences provides a more comprehensive view of user interests and helps capture relationships beyond direct user-item interactions.

  3. The combination of heterogeneous graph structures and semantic information offers a reasonable approach to capturing complex relationships among users, items, and their interactions.

Critical weaknesses
  1. The proposed framework mainly combines existing techniques, including IHGNN, TransE, TransD, attention mechanisms, and BPR. Its main novelty lies in the organization of behavior chains, but the differences from existing methods that support flexible behavioral dependency modeling are not sufficiently clarified.

  2. The proposed method directly associates combinations of auxiliary behaviors with the target behavior without explicitly estimating dependency strengths. It remains unclear whether this design effectively alleviates the accumulated influence of auxiliary behaviors. Moreover, compared with existing methods such as DA-GCN, it is difficult to distinguish the benefits of the proposed behavior modeling strategy from those introduced by additional graph representation learning components.

  3. The proposed semantic enhancement relies on encoding only three predefined behavior descriptions into text embeddings, which appears relatively straightforward. The performance gains from this component are also limited, and further comparisons with conventional learnable behavior embeddings would help justify the necessity of introducing pretrained language representations.

Questions for authors
  1. Could the authors further clarify the key methodological differences between SDBC-RO and existing multi-behavior recommendation methods, particularly those that already support flexible behavioral dependency modeling, such as DA-GCN? What are the unique contributions beyond the combination of existing graph learning techniques?

  2. Why is the proposed behavior chain design characterized as strong-dependency modeling rather than simply behavior grouping? What specifically distinguishes the proposed design from conventional behavior grouping strategies? Could the authors provide a quantitative analysis of the dependency strengths between auxiliary behaviors (Click and Cart) and the target behavior (Buy) to support this claim?

  3. Given that the semantic enhancement relies on only three predefined behavior descriptions, what additional benefits do pretrained text embeddings provide over conventional learnable behavior embeddings? Could the authors provide a more direct comparison to justify the necessity of this semantic encoding component?

1 comment

1: We greatly appreciate the reviewer’s constructive feedback, which has helped us recognize that the original manuscript did not sufficiently clarify the core contributions and innovations of this work. The employed techniques, including IHGNN, TransE, TransD, the attention mechanism, and BPR, serve as standard foundational algorithms for our representation learning pipeline. The key contribution of this work is twofold: first, we propose the strong dependency behavior chain that faithfully captures the real conversion patterns of user behaviors; second, based on this novel construct, we develop a ternary hypergraph-integrated multi-behavior representation semantic enhancement mechanism, as well as a knowledge graph-enabled multi-view representation semantic optimization method. The proposed strong dependency behavior chain, rooted in real user auxiliary-to-target behavior conversion patterns. In preprocessing, for each user-item interaction pair, we sort out user behaviors by timestamp, take the first Buy as the termination point of the interaction sequence, and remove all subsequent interactions. We then categorize the records into four disjoint chains: Click > Buy, Cart > Buy, Click&Cart > Buy, and Only Buy. For chains with combined auxiliary behaviors (e.g., Click&Cart > Buy), we only require both actions to occur before Buy, without restricting their internal order. All chains are mutually independent with non-overlapping records, and target behaviors in each chain are fully converted from its preceding auxiliary behaviors — this is exactly the meaning of strong dependency. The conversion rates in Table 2 further quantify this dependency. Therefore, compared with existing methods that support flexible behavioral dependency modeling, the proposed strong dependency behavior chain can reduce the interference of irrelevant auxiliary behaviors on the target behavior and effectively eliminate the undesired superposition effect across multi-level auxiliary behaviors. We have supplemented and refined it in Section 3.2.1 of the revised manuscript. 2: We sincerely thank the reviewer for the valuable comments on strong dependency behavior chain. The original explanation of this definition was indeed insufficient, and we have supplemented and refined it in Section 3.2.1 of the revised manuscript. In preprocessing, for each user-item interaction pair, we sort out user behaviors by timestamp, take the first Buy as the termination point of the interaction sequence, and remove all subsequent interactions. We then categorize the records into four disjoint chains: Click > Buy, Cart > Buy, Click&Cart > Buy, and Only Buy. For chains with combined auxiliary behaviors (e.g., Click&Cart > Buy), we only require both actions to occur before Buy, without restricting their internal order. All chains are mutually independent with non-overlapping records, and target behaviors in each chain are fully converted from its preceding auxiliary behaviors — this is exactly the meaning of strong dependency. The conversion rates in Table 2 further quantify this dependency. Through the statistical analysis of two datasets in Table 2, it is not difficult to observe that different strong dependency behavior chains exhibit significant differences in the number of interaction records and Buy conversion rates. The click behavior has the largest number of interaction records and the widest coverage, which can reflect users' relatively broad interest preferences, but its Buy conversion rate is relatively low. In contrast, the cart behavior, as well as the co-occurrence interaction records of click and cart behaviors, have a smaller quantity of interaction records and relatively limited coverage, yet they deliver a much higher Buy conversion rate. It can be concluded that although high-frequency behaviors cover a wider range of user interest preferences, their contribution to target behavior prediction is generally lower than that of low-frequency behaviors. Therefore, it is necessary to subdivide the types and scopes of auxiliary behaviors to predict their actual contribution to the target behavior and their real impact on user preferences. 3: We acknowledge that this semantic enhancement module is not the main source of performance gain for the model, yet it still contributes positively. It provides external textual semantic priors for the three types of behaviors prior to relation learning. An ablation experiment w/o LLM has been added in Section 4.4 of the revised manuscript: while keeping the downstream architecture identical, the pretrained semantic representations are replaced with randomly initialized learnable behavior ID embeddings. The performance drop of this variant on both datasets offers the direct comparison requested by the reviewer. This variant underperforms the full model on both datasets, verifying that pretrained textual semantic priors deliver consistent gains over arbitrary learnable ID representations.

Reviewer_w7d2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SDBC-RO for multi-behavior recommendation. It groups auxiliary behaviors into target-oriented behavior chains and combines two-level user-item-category hypergraphs, multi-view knowledge-graph modeling, and adaptive representation fusion. Experiments on Tmall and Taobao report improvements over existing baselines.

Critical strengths
  1. The paper addresses a relevant problem: exploiting auxiliary behaviors without overwhelming the target behavior in multi-behavior recommendation.

  2. It reports substantial gains over the strongest listed baseline on both datasets and includes component-level ablations.

Critical weaknesses
  1. The central notion of strong dependency is insufficiently justified. SDBC mainly partitions observed behavior combinations (e.g., Click>Buy and Click&Cart>Buy), rather than learning or validating actual behavioral dependencies. Temporal ordering within chains is not modeled, as acknowledged in the conclusion.

  2. Section 3.3.1 defines its edges over all ordered pairs of distinct behaviors, which includes reciprocal relations and is inconsistent with a partial order. Applying TransE with one shared relation to such pairs may yield contradictory constraints; the authors should clarify the graph construction and optimization.

  3. The claimed LLM-based modeling amounts to encoding only three manually written behavior descriptions using BAAI/bge-m3. The paper does not convincingly establish an advantage over simpler learned behavior embeddings under comparable capacity.

  4. Presentation quality needs substantial improvement. Figure 1 is overcrowded, Figures 2 and 3 contain rasterized graphics, and the dense notation makes the method difficult to follow. A simpler explanation and clearer figures would improve verifiability.

Questions for authors

Please refer to weakness.

1 comment

1: We sincerely thank the reviewer for the valuable comments on the core concept "strong dependency". The original explanation of this definition was indeed insufficient, and we have supplemented and refined it in Section 3.2.1 of the revised manuscript. In preprocessing, for each user-item interaction pair, we sort out user behaviors by timestamp, take the first Buy as the termination point of the interaction sequence, and remove all subsequent interactions. We then categorize the records into four disjoint chains: Click > Buy, Cart > Buy, Click&Cart > Buy, and Only Buy. For chains with combined auxiliary behaviors (e.g., Click&Cart > Buy), we only require both actions to occur before Buy, without restricting their internal order. All chains are mutually independent with non-overlapping records, and target behaviors in each chain are fully converted from its preceding auxiliary behaviors — this is exactly the meaning of strong dependency. The conversion rates in Table 2 further quantify this dependency. We note that this work only restricts the before-after relationship between auxiliary behaviors and the target behavior, and does not discuss the temporal order among different auxiliary behaviors, which we leave for future research. 2: We thank the reviewer for pointing out this ambiguity in the mathematical formulation. The behavior progressive knowledge graph adopted in our work includes only three behavior entities, namely Click, Cart, and Buy, with merely two directed positive edges: (Click, r, Cart) and (Cart, r, Buy). No bidirectional relations exist between any arbitrary behavior pairs. The shared relation r denotes the direct progressive relation of behavioral intention, transitioning from a lower-intent behavior to a higher-intent behavior. We will redefine this graph with an explicit edge set in Section 3.3.1, and consistently refer to it as the directed behavior progressive graph throughout the manuscript, to eliminate this ambiguity and avoid the notation being misinterpreted as the complete set of all ordered behavior pairs. 3: This semantic module only performs semantic encoding for the three behavioral descriptions: Click, Cart and Buy, and it is not an independent core innovation of this paper. It should be noted that an ablation experiment w/o LLM has been added in Section 4.4 of the revised manuscript: randomly initialized, learnable behavior ID embeddings are used to replace the semantic encoder BAAI/bge-m3, while keeping the downstream structure of SDBC-RO unchanged. The performance of this variant is inferior to that of the full model on both datasets, which demonstrates that the pretrained textual semantic prior yields steady improvements over learnable ID representations. 4: We thank the reviewer for the valuable suggestions. We have made corresponding revisions in the revised manuscript, as detailed below: Figure 1 has been reorganized around four core stages: SDBC construction, two-level heterogeneous hypergraph learning, behavior semantics knowledge graph optimization, and cross-chain fusion. We have streamlined redundant notations and dense arrows, and unified the legend symbols for users, items, categories, and behavior semantics. Figures 2 and 3 have been fully redrawn as vector graphics with consistent font styles and sizes across all figures. Meanwhile, we have conducted a systematic review of the notation system in the methodology section, ensuring that the edge set of the behavior graph, TransE relation definitions, representation dimensions, loss functions, and negative sampling rules are all fully consistent with the actual implementation. In addition, we have added brief intuitive introductions at the beginning of each core methodology subsection, allowing readers to first understand the modeling objectives and overall information flow before delving into the formula details.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.