- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper proposes LIFT, a lifecycle-aware interaction factorization framework that unifies retrieval and ranking in recommender systems. Each interaction is decomposed into ordered Request, Item, Context, and Action tokens, serialized as a causal sequence so that retrieval reads a Request state (pre-candidate) and ranking reads a Context state (post-candidate, pre-action). A shared decoder with Role-Conditioned Attention (shared Q/K, role-specific V/O) and a lightweight Pre-LN role bias enables shared historical modeling while preserving stage-specific information. On ML-20M and Taobao, LIFT achieves the best Joint Score among joint baselines, shows favorable retrieval–ranking Pareto trade-offs, and is supported by ablations on lifecycle design, model components, and scaling analyses.
Critical strengths
- The lifecycle factorization (R→I→C→A) is a clear and principled way to align what information is available at different stages, and to define stage-aligned readouts for retrieval (Request) and ranking (Context).
- Role-Conditioned Attention (shared Q/K; role-specific V/O) is a neat parameter-sharing scheme that matches the intent: common dependency matching across roles while allowing role-specific transformations.
- The Pre-LN role bias is a simple, low-overhead conditioning mechanism that integrates well with a Pre-LN Transformer.
Critical weaknesses
- The framework increases sequence length by 4× per interaction; while training/VRAM costs are analyzed, serving-time latency and throughput under realistic candidate counts are not reported in detail.
Questions for authors
- How are ranking negatives built during training, and are the evaluation ranking candidates exactly the top-K retrieved by LIFT's retrieval head, or a fixed pool per query?
2 comments
A-rPGn-2: Ranking negative construction and evaluation protocol (Q1).
A-rPGn-2: Ranking negative construction and evaluation protocol (Q1).
Thank you for this question. We clarify the general formulation of the ranking loss, the training supervision, and the ranking evaluation protocol.
Ranking negatives during training. The ranking loss in Eq. (3) adopts a general formulation:
= \ell_{\mathrm{rank}} \left( \hat{\mathbf{y}}_{u,\leq T}^{\mathrm{rank}}, \mathbf{y}_{u,\leq T}^{\mathrm{rank}}, \mathcal{N}^{\mathrm{rank}} \right),$$ where $\mathcal{N}^{\mathrm{rank}}$ denotes an optional set of additional negative samples or candidate-sampling information, allowing the framework to accommodate different ranking objectives and negative sampling strategies. In our experiments, we adopt the standard binary cross-entropy (BCE) loss for ranking, with supervision labels constructed directly from observed interactions and their ground-truth feedback: - **ML-20M:** Interactions with ratings greater than 4 are treated as positive samples, while those with ratings of 4 or below are treated as negative samples. - **Taobao:** Purchases are treated as positive samples, while other observed behavior types (page views, favorites, and add-to-cart events) are treated as negative samples. Since BCE directly computes the loss from the predicted score and binary label of each observed interaction, it does not require constructing additional negatives for contrastive or pairwise learning. Therefore, $$\mathcal{N}^{\mathrm{rank}}=\varnothing$$ in our experimental setting. Specifically, the ranking loss is instantiated as: $$\mathcal{L}_{\mathrm{rank}} = -\frac{1}{|\mathcal{S}|} \sum_{i\in\mathcal{S}} \left[ y_i\log p_i + (1-y_i)\log(1-p_i) \right],$$ where $\mathcal{S}$ denotes the set of observed interactions used for ranking supervision, $y_i\in\{0,1\}$ is the ground-truth feedback label, and $p_i$ is the predicted probability of positive feedback. This does not imply the absence of negative ranking examples; rather, negative supervision is already provided by the binary labels derived from observed feedback, without requiring additional negative sampling. In contrast, retrieval is trained using InfoNCE with in-batch negatives and LogQ correction. Thus, the two tasks share historical modeling while retaining task-specific supervision and objectives. **Independent ranking evaluation.** The ranking metrics (AUC and GAUC) in Table 2 are computed on held-out observed interactions and their ground-truth feedback. All models are evaluated on the same test interactions and labels, rather than on the top-$K$ candidates produced by their respective retrieval modules. This evaluation primarily measures the ranking model's ability to predict user feedback. **End-to-end cascade evaluation.** To further assess the effectiveness of the complete recommendation pipeline, we conduct additional end-to-end cascade experiments. Unlike independent ranking evaluation, the ranking candidates in these experiments are exactly the top-$K$ items retrieved by each model's own retrieval module. We evaluate all positive interactions in the ML-20M test set and report Recall@10 and NDCG@10 of the final recommendation lists across 10 candidate set sizes ($K=10$–$100$). LIFT achieves the best average performance across all evaluated candidate sizes. For example, at $K=100$, the results are: | Model | Cascade R@10 | Cascade N@10 | |---|---:|---:| | UniPinRec | .0625 | .0328 | | HSTU | .0989 | .0473 | | LIFT | **.1147** | **.0596** | LIFT improves final-list Recall@10 and NDCG@10 over the strongest baseline, HSTU, by 16.05% and 25.93%, respectively. Please refer to A-tLCs-1 for the complete cascade evaluation results. We will clarify the definition of the negative sample set in Eq. (3), as well as the distinctions between training supervision, independent ranking evaluation, and end-to-end cascade evaluation in the revised manuscript.
A-rPGn-1: Serving efficiency and cache overhead (W1).
We thank the reviewer for recognizing the merits of lifecycle factorization, Role-Conditioned Attention, and Pre-LN Bias. We appreciate the constructive questions regarding serving efficiency and the ranking training and evaluation protocols. We provide further analysis and clarifications below.
A-rPGn-1: Serving efficiency and cache overhead (W1).
We agree that LIFT's finer-grained lifecycle modeling introduces additional serving-time computation and cache overhead, making the quality–efficiency trade-off an important consideration.
Under matched model widths, depths, and cache precision, LIFT requires 4× the historical KV-cache capacity of UniPinRec, while its ranking FFN computation and candidate-history attention computation increase to 2× and 8×, respectively.
However, comparisons at the same model width do not fully characterize the quality–efficiency trade-off. Figure 4(b) shows that LIFT with a hidden dimension of only 64 achieves a Joint Score of .2648, outperforming UniPinRec with a hidden dimension of 256 (.2582).
Under these configurations, the theoretical serving costs are summarized below:
LIFT requires the same historical KV-cache capacity as UniPinRec, while its ranking FFN computation is reduced to 12.5% and its candidate-history attention computation is 2× that of UniPinRec.
These results suggest that LIFT's more effective lifecycle representation enables better recommendation quality with a smaller model width, thereby mitigating the serving overhead introduced by finer-grained sequence modeling.
Please refer to A-tLCs-5 for a detailed analysis of computational complexity, cache overhead, and the quality–efficiency trade-off. We emphasize that these results characterize theoretical model-level costs rather than empirically measured end-to-end latency or throughput in a production environment. We will include this analysis and clarify this limitation in the revised manuscript.