acceptodds
Under review as a conference paper at ICLR 2027

Learning Where to Look: Geometric Multi-Interest Retrieval for Recommender Systems

Abstract

The retrieval stage of recommender systems must efficiently identify a small set of items relevant to a given user from a large item corpus. Despite their differences, many existing retrieval methods share a common structure: query-based nearest-neighbor retrieval in an item embedding space. From this perspective, retrieval involves two fundamental questions: where to place the query vectors and how far to search. In this work, we develop a geometric framework for multi-interest retrieval. We show that cosine-similarity retrieval corresponds to searching spherical caps on the embedding sphere, and formulate efficient retrieval as a geometric covering problem. Our theoretical analysis shows that multiple retrieval regions can substantially reduce the retrieval cost for users with diverse interests. We further identify two sources of query placement error: sampling error, caused by estimating interests from limited historical interactions, and drift error, caused by changes in user interests over time. To address these errors, we learn multiple retrieval regions directly from user histories, with each region parameterized by a query center and an angular radius. The model is trained to cover future target items while minimizing the total retrieval region. Experiments on four recommendation datasets show that our approach consistently improves top-ranked retrieval quality over strong single-interest and multi-interest baselines, while maintaining compact retrieval regions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
dense retrieval, recommender systems, multi-interest user modeling, multi-vector retrieval, high-dimensional geometry

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yGw9OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper recasts dense retrieval for recommenders as geometry. With item embeddings on the unit sphere, cosine ranking is angular ranking, so a query under a score threshold returns exactly a spherical cap whose area is proportional to how many items are probed. Retrieval becomes a covering problem: cover the future target set with the least total cap area. Theorem 1 bounds an m-cap cover against a single cap by a ratio in which the radius enters to the power d−1 while m enters only linearly, so small radius gains buy large savings. Theorem 2 explains why the natural construction fails: history cluster centers carry sampling error of order θ/√n plus drift error from moving interests, and the cost inflation grows exponentially in dimension again. The method therefore learns the placement: a Transformer over the history emits m heads, each producing a normalized center and a softplus radius, trained with a hinge miss term under a smooth softmin assignment plus an area penalty. On MovieLens-1M/20M, KuaiRand-27k and an internal short-video set it reports the best HR@10 and NDCG@10 on all four, with a Scan metric showing the probed corpus fraction falling from 24.1% at one head to 9.76% at three.

Critical strengths
  • The framing is the contribution and it is correct. Cosine ANN retrieval equals spherical-cap membership, and retrieval cost equals cap area, so multi-query retrieval is a covering problem whose radius advantage is exponential in dimension minus one. I have not seen this written down for recommender retrieval before, and it explains in one line why multiple interest vectors pay off at d = 128 but are much less obviously worth it at d = 32. This is framing that will outlive the specific model, and the connection to min-sum-radii in Footnote 3 is stated honestly, including the NP-hardness that then justifies the surrogate objective.

  • Theorem 2 is used to make a falsifiable prediction and the prediction is tested. The natural instantiation of history clustering, UIH_k-means, is the worst method in Table 1 on MovieLens-1M by a wide margin — HR@10 of 2.91 against SASRec's 28.3 — and ComiRec's clustering-derived variants are also weak on the short-video set. That is exactly what the sampling-plus-drift bound predicts, and it is a far more satisfying use of a theorem than the usual properties paragraph. The local-density check in Section 4.2 is also the right way to defend the area-as-cost proxy rather than waving at the uniformity assumption.

  • Reporting the Scan metric alongside HR and NDCG is the correct measurement decision, and the numbers are informative rather than decorative. Ranking quality alone cannot distinguish a method that finds the target from a method that finds it while probing less of the corpus. Table 4 gives Scan of 24.1% at m = 1, 9.76% at m = 3 and 1.12% at m = 16 on MovieLens-20M, and Table 2 shows that dropping the area term raises Scan from 9.8% to 38.5% while barely changing ranking quality — which is exactly the trade-off Theorem 1 says to manage, measured. The case study in Figure 2 closes the loop at the user level, with cap centers 0.89 radians apart tracking group centers 0.94 radians apart for a two-interest history.

Critical weaknesses
  • The theory and the experiment do not describe the same target set, and the 100%-recall idealization is not innocuous even in the authors' own numbers. Theorems 1 and 2 are about covering a set T of future items at full recall, and Footnote 1 makes that idealization explicit. Section 4.1 then evaluates with full-corpus leave-one-out, stating "exactly one item is held out for each user". When |T| = 1 there is no covering problem at all: one cap always suffices, θ_m = θ_1, and Theorem 1's advantage is identically zero. So the empirical section is not testing the object the theory is about; HR@k on a single positive is a ranking metric, and the geometric cost model never gets exercised as a covering claim on real multi-positive targets. Moreover, Table 2 reports full-corpus HR@200 of 75.6 against cap-restricted HR@200 of 73.0 for the full objective, so the learned union of caps misses on the order of 2.6pp of the targets that the ranking stage would otherwise have found — the coverage assumption the theory rests on does not hold in the deployed model, and the paper does not connect the gap. Figure 2a is the same phenomenon made visible: a dispersed history produces two caps of radius 1.25 and 1.16 radians probing 29.3% of the pool, the target is covered, and it still ranks 437th. The internal short-video dataset construction already uses a future engagement window, so a multi-positive protocol is answerable with a protocol change rather than a new model.

  • There is no variance anywhere in the main results. Table 1 gives seven methods across four datasets as bare numbers, with no seeds, no standard deviations, and no statement of how many runs produced each cell. Large gaps are safe, but the paper's "consistently improves" claim does not hold uniformly as stated: on KuaiRand-27k the method is worse than both single-interest baselines on H@200 (10.4 versus SASRec's 10.9 and HSTU's 11.5) and again on N@200 (2.29 versus 2.35 and 2.33), even though it wins H@10 and N@10 there. A tail regression on the behavioral dataset, in the metric closest to the retrieval budget the paper motivates, needs either error bars or a sentence of discussion rather than silence.

  • The serving argument is the motivation but is not measured. The whole point of minimizing total cap area is that the retriever touches fewer items, yet in practice m caps are executed against an ANN index and the cost is index traversals and candidate generation under a per-request budget. The paper reports neither candidate counts at matched recall in a real index nor latency, and there is no comparison against a cheap alternative at equal budget — for instance, retrieving top-K around each of m cluster centers with K tuned to the same total candidates. Without this, area is a proxy that has not been connected to the quantity an industrial system actually pays for.

  • The radius is set by a regularizer, not by a target, and there is no sensitivity study. The two loss-balancing coefficients λ_area and λ_ntp are fixed at 0.1 and 1 in the implementation paragraph and used unchanged across all four datasets, together with a softmin temperature of 0.05, yet these are the knobs that determine the learned cap widths — the paper's central object. Two consequences are unaddressed. First, a practitioner cannot obtain a required recall level, because the objective trades coverage against area through a coefficient rather than through a stated recall constraint. Second, the same constants are reused across four corpora that differ by more than an order of magnitude in item count and that are trained under different negative-sampling schemes (full-corpus scoring on the 3,883-item MovieLens-1M versus 128 sampled negatives elsewhere), so the reported cap compactness may partly reflect a lucky constant rather than a property of the geometry. A sweep over λ_area showing recall against probed area, ideally with a version that enforces a budget constraint, would fix this cheaply. On head allocation, the paper does analyze the aggregate version: Section 4.4 and Appendix G.2 report that as m grows, supervision concentrates on a subset of caps, leaving others with few assignments and weak gradients, which explains the m = 16 quality drop. What is missing is the per-user version. m is a global constant, so a user whose history is concentrated still emits m caps, and Figure 2b shows precisely that — two centers 0.13 radians apart with a Jaccard overlap of 0.87, which the paper calls safe overlap. It would be useful to know how much of the reported Scan budget goes to redundant heads, since a per-user allocation of active heads (or a sparsity term over head utilization) would presumably buy back area for free.

Questions for authors
  • Can you evaluate on a multi-positive future window and report the metric the theory speaks about — P@R at fixed recall, or the realized total cap area needed to cover the target set — on MovieLens-20M and the internal dataset? This is my main concern and I would raise my soundness score if the covering advantage shows up empirically. (Clarification, rebuttal-feasible.)

  • Why does the method regress on KuaiRand-27k H@200 while improving on H@10? I suspect radius calibration for low-activity users, i.e. heads widening past what a short history supports, but the paper neither reports a per-activity breakdown nor speculates.

  • Table 4 gives Scan as a function of the global m, and Figure 2b shows two nearly identical caps (Jaccard 0.87) for a concentrated history. Can you report, per user, the effective number of distinct caps, the share of total cap area contributed by redundant heads, and the distribution of learned radii? If head utilization were allocated per user rather than fixed at m, the m = 3 versus m = 8 trade-off in Table 4 should improve, which would be a stronger efficiency result than the current constant-m configuration.

  • How much of the internal-dataset gain is attributable to the fact that its target window contains multiple positives, unlike the leave-one-out academic sets? The largest gains are reported on exactly that dataset, and I would like that confound separated.

  • Have you tried replacing the analytic area penalty with a hard budget — minimize miss subject to Σ_j A(θ̂_j) ≤ c, swept over c? That would give the recall-versus-budget curve an engineer needs and would make Theorem 1's prediction directly checkable.

3 comments
AuthorsOpenReview ↗

Author response

We thank the reviewer for the valuable comments. In response, we have conducted an extensive set of additional experiments and uploaded a revised version of the paper, with all changes highlighted in blue. Our point-by-point responses are provided below.

1. Multi-positive evaluation with precision at a given recall

Following your advice, we have added a new evaluation protocol in which the last 20% of each evaluation user's interaction sequence is used as the target set. We report precision at 50% and 80% recall (Section 4.1-Evaluation metrics; Appendix G.1). Our method consistently outperforms all baselines under this protocol. This provides further evidence of the effectiveness of our covering formulation in the multi-positive setting.

2. Regression on KuaiRand-27k HR@200

We would like to make three points.

  • The performance gap is approximately one percentage point, and our method remains the best at the top of the ranking.

  • As discussed in Section 4.3, HSTU is a particularly strong baseline on this dataset, as it was specifically designed to handle long and non-stationary sequences.

  • We believe this result reflects a trade-off induced by the area regularization. As described in Section 3.3, items are ranked by their minimum slack across predicted caps, while the area term encourages the model to produce compact caps. Compact caps lead to more discriminative rankings for items inside or close to a cap, improving retrieval quality near the top of the ranking. In contrast, items far from all predicted caps tend to receive similarly large slack values, making their relative ordering less discriminative at deeper ranking positions. Thus, the lower HR@200 is a consequence of the same cap compactness that reduces retrieval cost.

3. Adaptive number of caps

We would like to clarify three aspects.

  • We have already identified an adaptive number of retrieval regions as future work in our limitations (Appendix H). Realizing this would require substantial modifications to the model architecture, together with dedicated experiments. We believe this is better addressed as a separate research direction rather than as an incremental extension of the current work, and it was beyond the scope of the rebuttal period.

  • Our objective jointly encourages target coverage and penalizes the total area of the predicted caps. Consequently, heavily overlapping caps are generally discouraged, since they incur redundant area without providing additional coverage. The heavily overlapping caps shown in the case study represent a deliberately selected edge case: a user with a short and highly concentrated interaction history. This example is not representative of the general behavior of our model.

  • Following your advice, we have quantified the redundant fraction of the probed area. Let CiC_i denote the set of items covered by cap ii. We define the redundant fraction as 1−∣⋃iCi∣/∑i∣Ci∣1-|\bigcup_i C_i|/\sum_i |C_i|. Across different numbers of caps, with mm ranging from 2 to 16, this fraction remains between 1.2% and 2.5%. This indicates that the predicted caps overlap only minimally in practice, and that very little of the probed area is counted more than once.

4. The target window on the internal dataset

We believe there may be a misunderstanding here, and we apologize if our description was unclear. The evaluation protocol is identical across all datasets: we use leave-one-out evaluation with a single held-out target per user. The internal dataset does not use a multi-positive target window. A more likely explanation for the larger gains on the internal dataset is that our hyperparameters were selected primarily based on the internal dataset and then reused unchanged on the public datasets. Therefore, the results on public datasets may actually be conservative compared with what could be achieved through dataset-specific hyperparameter tuning.

5. Hard budget versus the area penalty

The hard-budget and penalty-based formulations express the same underlying trade-off between missed targets and retrieval cost. Specifically, the constrained optimization problem min⁡Lmiss\min L_{miss} subject to Larea≤cL_{area} \leq c has the Lagrangian Lmiss+λ(Larea−c)L_{miss}+\lambda(L_{area}-c), which, up to the constant term −λc-\lambda c, gives the penalized objective Lmiss+λLareaL_{miss}+\lambda L_{area}. Thus, varying λ\lambda provides a practical way to trace the recall–budget trade-off associated with different choices of cc. Following your advice, we have conducted this sweep on MovieLens-1M:

λarea\lambda_{area}00.010.11
Cap-restricted HR@20074.673.873.069.2
Scan (% of corpus)38.515.39.768.85

The results illustrate the expected trade-off: reducing the retrieval budget comes at the cost of lower recall.

Thanks for the additional experiments. The multi-positive protocol with precision at fixed recall is exactly the evidence I asked for, the λ_area sweep plus the Lagrangian argument answers the sensitivity question, and the measured 1.2 to 2.5 percent redundancy rules out my guess that per-user head allocation would recover much area. You also corrected my mistaken assumption that the internal set used a future window, and I accept that. I am raising soundness to 3. Two things are still missing, so the rating stays at 3: no seeds or standard deviations appear anywhere, which leaves "consistently improves" unsupported on cells like KuaiRand H@200, and there is still no index-level measurement of candidates or latency at matched recall, so the retrieval-cost claim remains an area proxy. One further thing I would like to see stated in the paper: if hyperparameters were chosen on the internal dataset and reused on the public ones, then the internal numbers are the optimistic ones and the public ones are conservative, which is the opposite emphasis from the current presentation. Please also say which values were selected that way.

We thank the reviewer for the thoughtful response and for raising the soundness score. We believe that the remaining two “missing things” are already addressed in the current version, and we apologize for not stating them clearly enough. We would be grateful if you would consider raising the rating in light of the following clarifications.

1. Seeds and standard deviations. For the online experiment, Figure 3a reports 95% confidence intervals for every metric, and all eight changes are statistically significant. For the offline experiments, the results in the new Appendix G.1 were already obtained over three random seeds, with standard deviations computed accordingly. We omitted them only to keep the formatting consistent with the other offline tables. We have now added the number of seeds and standard deviations to Table 6 in the latest version. The improvements of our method are statistically significant there as well.

2. Index-level measurement of candidates and latency at matched recall. This is precisely what Figure 3b reports. The top panel shows the number of candidates required by each method at matched recall. At the two marked operating points, one head requires approximately 2.2× and more than 2.5× as many candidates as three heads to achieve the same recall. Reading the bottom panel at these same candidate budgets further shows that the single-head arm has higher ANN index latency. Thus, at matched recall, the single-interest method has higher index latency.

Regarding hyperparameter selection, the selected hyperparameters include the number of Transformer layers, hidden dimension, number of interest heads, λarea\lambda_{area}, λntp\lambda_{ntp}, and learning rate. We would also like to clarify why we do not believe this makes the internal results too optimistic and the public results too conservative. First, the baselines were tuned on the internal dataset using the same procedure, so all methods are compared under the same tuning protocol. Second, the ablations in Tables 3, 4, and 5 were conducted on MovieLens, and the configurations they favor agree with those selected on the internal dataset. This suggests that the chosen hyperparameters are not specific to a particular corpus.

Reviewer_fnY1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a geometric framework for multi-interest retrieval in recommender systems, formulating cosine-similarity retrieval as a spherical-cap covering problem. The authors theoretically analyze the benefits of multiple retrieval regions and identify sampling error and interest drift as limitations of history-based query placement. Based on these insights, they propose a Transformer-based model that jointly predicts multiple query centers and angular radii, optimized through coverage, area, and next-token prediction losses. Experiments on four datasets demonstrate improvements in top-ranked retrieval performance and compact retrieval regions.

Critical strengths
  1. Interesting geometric perspective. The spherical-cap formulation provides an intuitive interpretation of multi-interest retrieval and connects query placement with retrieval efficiency.

  2. Reasonable and well-motivated design. Jointly learning query centers and radii explicitly considers the trade-off between target coverage and retrieval cost.

  3. Encouraging empirical results. The method achieves the best HR@10 and NDCG@10 across four datasets. The ablation studies and case analyses further illustrate the effects of different components and query configurations.

Critical weaknesses
  1. Limited methodological novelty. Multi-interest retrieval using multiple query vectors has been extensively studied. The spherical-cap interpretation follows naturally from cosine similarity, while the proposed architecture largely combines a standard Transformer with multiple prediction heads and auxiliary losses. The distinction from existing learned multi-interest methods is therefore somewhat incremental.

  2. Gap between theoretical analysis and practical assumptions. The theoretical benefits depend on idealized assumptions about interest geometry, item density, and query placement errors. Theorem 1 provides a conditional cost bound rather than guaranteeing that learned queries achieve the required radius reduction, while Theorem 2 relies on restrictive assumptions and worst-case coverage arguments. It remains unclear how strongly these theoretical conclusions hold in realistic embedding spaces.

  3. Insufficient validation of retrieval efficiency. Although reducing retrieval cost is a central motivation, most experiments evaluate full-corpus ranking performance, while actual ANN serving is only described. Direct comparisons of retrieval latency, candidate budgets, and recall–cost trade-offs against baselines are missing, making the practical efficiency improvements difficult to assess.

  4. Experimental evidence could be more convincing. Improvements on public datasets are relatively modest, and the proposed method does not consistently outperform HSTU at larger retrieval cutoffs. The substantial underperformance of existing multi-interest baselines also raises questions about baseline competitiveness. Furthermore, the largest gains come from a proprietary dataset, and there is limited direct evidence that the proposed approach specifically mitigates interest drift rather than benefiting from stronger sequence modeling and additional training objectives.

Questions for authors

See my strong and weak points.

3 comments
AuthorsOpenReview ↗

Author response

We thank the reviewer for the valuable comments. In response, we have conducted an extensive set of additional experiments and uploaded a revised version of the paper, with all changes highlighted in blue. Our point-by-point responses are provided below.

1. Novelty

Existing multi-interest methods (MIND, ComiRec, and PinnerSage) represent a user's interests as mm points (i.e., interest vectors), whereas our method represents them as mm regions, each defined by a center and an angular radius. This distinction is fundamental: explicitly modeling the radius allows retrieval cost to be incorporated as a differentiable quantity in the training objective. In contrast, existing methods primarily optimize retrieval accuracy without explicitly controlling retrieval cost, while our method jointly optimizes both.

2. Connection between theory and practice

We would like to clarify four aspects of this connection:

  • The role of assumptions. Recommendation is largely an empirical field, and much of its progress has been driven by empirical methods. Establishing quantitative theoretical results in this setting inevitably requires precise definitions and explicit assumptions. We believe making these assumptions transparent is essential for ensuring that theoretical claims are both interpretable and verifiable.

  • Empirical validation of assumptions. Each assumption introduced in our theoretical analysis is accompanied by an empirical check. Specifically, Section 4.2 investigates the assumptions underlying both theorems, and the results suggest that these assumptions are reasonable in our experimental setting.

  • Realized retrieval cost. Theorem 1 establishes a conditional bound on retrieval cost. We agree that this theoretical bound alone does not guarantee that the learned queries achieve the same bound in practice. We therefore report the realized retrieval cost in Table 5: as the number of predicted interests increases from 1 to 16, the fraction of the corpus scanned decreases from 24.1% to 1.12%.

  • How the theory motivates the method. More importantly, the value of the two theorems lies not merely in the mathematical form of their bounds, but in the design insights they provide. Theorem 1 shows that representing multiple interests can substantially reduce retrieval cost, motivating the use of multiple interests in the first place. Theorem 2 shows that predicting future interests based solely on an individual user's interaction history incurs substantial error, motivating cross-user learning to improve interest placement. These two implications were explicitly highlighted in the italicized remarks following the theorems (line 142, line 181).

3. Validation of retrieval efficiency

Following your advice, we have added online evaluation results. Figure 3b in Section 5 reports both the recall–cost trade-off and ANN retrieval latency. Compared with the single-interest method, the multi-interest formulation achieves higher recall with negligible additional retrieval latency.

4. Strength of the experimental evidence

We have added three components to strengthen the experimental evidence:

  • New dataset: Amazon Books (Section 4.1-Datasets; Tables 1 and 2; Appendix D). We note that multi-interest baselines perform strongly on this dataset, with several outperforming SASRec and HSTU. This provides further evidence that the baselines are competitive.
  • New baseline: DPCML, an additional metric-learning baseline (Section 4.1-Baseline methods; Tables 1 and 2; Appendix E).
  • New evaluation protocol: We use the last 20% of each evaluation user's interaction sequence as the target set, rather than holding out a single item, and report P@R50 and P@R80 (Section 4.1-Evaluation metrics; Appendix G.1).

Our method achieves consistent improvements across all three additions, providing further evidence of its effectiveness and addressing the concern that its gains on public datasets are relatively modest.

Regarding whether our method “specifically mitigates interest drift rather than simply benefiting from stronger sequence modeling or additional training objectives”, we would like to clarify that these components are integral mechanisms through which our method addresses interest drift, rather than independent sources of improvement that can be separated from the method itself. Their individual contributions are examined through ablation studies in Tables 3 and 4.

I have no further concerns and have raised the score. Good luck to this good work : )

Thank you so much for your kind support! We really appreciate your time and consideration. Please feel free to let us know if you have any further questions or concerns. We would be happy to address them.

Reviewer_Q9kbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Summary

The paper concentrates on the retrieval component of a recommendation pipeline and introduces a geometric approach to multi-interest retrieval that addresses both where to position query vectors and how wide the search should be. It casts retrieval as a geometric covering task, learning multiple retrieval regions directly from users’ historical interactions, where each region is defined by a query center and an angular radius. Training aims to ensure these regions capture future target items while keeping the overall covered region as small as possible.

Critical strengths

Strengths

  • This is an interesting and practical problem to solve, especially framing it as a geometric covering problem.
  • The authors introduce sampling error and drift error to address the 'center challenge'.
  • Since the problem is NP-hard, the authors propose optimizing the predicted caps directly to cover the targets while keeping the total area small.
  • The training and serving pipeline are introduced.
Critical weaknesses

Weaknesses

  • According to Table 1, it appears that the proposed method does not always outperform prior baselines/use cases.
  • Internal Short-Video appears to have the highest lift, but it would be great to report results in production and/or from an A/B test (e.g., treatment vs. control). Although Appendix D includes statistics, I believe showing only lifts compared to the (old) baselines is not enough.
  • Similar to the above, the large-scale online serving section needs more detail.
  • Complexity and runtime should be provided in more detail.
  • The case study in Section 4.5 is interesting; however, drawing learnings from Internal Short-Video would be more helpful than using MovieLens-1M.
Questions for authors

Questions

  • When our user interests evolve over time continuously, how do you address this problem? What are the edge cases?
  • Can we visualize Figure 2 in 3D instead of 2D (such as Figure 1a)?
  • Regarding the clustering effect in MovieLens-1M, it would be great to also consider other metric-learning approaches.
  • Please refer to my other comments in the Weaknesses section.
1 comment
AuthorsOpenReview ↗

Author response

We thank the reviewer for the valuable feedback. In response, we have conducted an extensive set of additional experiments and uploaded a revised version of the paper, with all changes highlighted in blue. Our point-by-point responses are provided below.

1. Effectiveness of the proposed method

We have added three components to strengthen the effectiveness of the our method:

  • New dataset: Amazon Books (Section 4.1-Datasets; Tables 1 and 2; Appendix D).
  • New metric-learning baseline: DPCML (Section 4.1-Baseline methods; Tables 1 and 2; Appendix E).
  • New evaluation protocol: We use the last 20% of each evaluation user's interaction sequence as the target set, rather than holding out a single item, and report P@R50 and P@R80 (Section 4.1-Evaluation metrics; Appendix G.1).

Our method demonstrates consistent improvements across all these additional evaluations.

2. Online A/B test and serving details

We have added Section 5 to describe the design and results of our online experiment. The A/B test demonstrates that the multi-interest formulation delivers significant improvements in online engagement over the single-interest variant, with negligible additional serving cost.

3. Complexity and runtime

Figure 3b in Section 5 reports the impact of the number of interests on ANN retrieval latency. Latency remains essentially unchanged, and for larger candidate-set sizes, the multi-interest variant can even be faster than the single-interest variant.

4. Case study

Following your advice, we have moved the case study to the Internal Short-Video dataset and replaced the 2D projection with 3D sphere projection. Further details are provided in Section 4.5.

5. Continuously evolving interests

We understand the question as how our model handles users whose interests continuously evolve and who keep exhibiting new interests. Predicting an entirely new interest before the user has shown any indication of it is inherently challenging. We address this through two aspects:

  • Collaborative learning of interest transitions. Our model is trained on data from all users and can leverage collaborative signals to learn patterns of interest evolution from users with similar interaction histories. Consequently, even if a particular interest has never appeared in an individual user's own history, the model may still anticipate it based on patterns learned from other similar users.

  • Adaptation to newly observed interests. An edge case arises when a user's interest changes are entirely random, with no similar users or recurring patterns to learn from. In this case, the model cannot anticipate the new interest. More generally, no machine-learning model can reliably predict genuinely random user behavior. However, once the user exhibits the new interest, our model can quickly adapt to the updated history and produce a cap corresponding to that interest.

6. Metric-learning baseline

We have added DPCML as an additional metric-learning baseline. Please see our response to Point 1 above for an overview, with implementation and evaluation details provided in Section 4.1-Baseline methods, Tables 1 and 2, and Appendix E.

Comments on the paper

1 comment

Dear Authors,

We are approaching the end of the rebuttal period, in order to allow the reviewers to respond to your rebuttal please make sure to submit it as soon as possible.

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.