acceptodds
Under review as a conference paper at ICLR 2027

What If Recommender Systems Were Circuits?

Abstract

Recommender system research largely focused on learning richer user and item representations and designing novel objective functions, while many approaches still model user-item interactions as a simple factorization realized via a dot product. We revisit this factorised component from a probabilistic perspective, asking whether it can support tractable probabilistic inference and learning, without redesigning the underlying recommender architecture. To this end, we introduce CiRe, a plug-in interaction head that replaces the usual dot product with a nonnegative squared circuit, thus delivering a tractable generative model that can be trained with maximum likelihood, avoiding the use of negative sampling and supporting probabilistic queries over user–item interactions. We show that across nine recommender architectures and five datasets, CiRe frequently matches or improves top-k ranking performance with the strongest results combining circuit parametrisation and likelihood training. For simple encoders, removing negative sampling can also reduce training time, whereas richer architectures exhibit different computational trade-offs, which we analyse. Our code is available at https://anonymous.4open.science/r/CiRe-2B56/.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
recommender systems, probabilistic circuits, matrix factorisation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fdLvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes CiRe, which treats the dot product interaction sθ(u,i)=eu⊤eis_{\theta}(u, i) = e_u^{\top}e_i as matrix factorization and improves it using a probabilistic circuit with an NLL loss. This circuited NLL serves as a replacement of BPR by avoiding its costly negative sampling. Authors then apply this circuited NLL to a broad class of classic-matrix-factorization-based, graph-based and content-based recommender models and empirically show that this approach can improve their top-kk performance.

Critical strengths
  1. This paper introduces a novel approach for matrix-factorization-based recommender systems. Many latent factor recommender models are based on matrix factorization (Section 3.6, Aggarwal, 2016) and are typically trained using SGD or ALS. By viewing these models as circuits, we introduce a neural-network-like structure for their training (Loconte et al, 2025), which can potentially improve their performance.

  2. The authors apply this circuit approach to a broad class of classic-matrix-factorization-based, graph-based and content-based recommender models, showing that this approach is general in improving their performance.

  3. The authors provide detailed experiments to evaluate their approach, including performance comparison and ablation study.

References:

Lorenzo Loconte, Antonio Mari, et al. What is the relationship between tensor factorizations and circuits (and how can we exploit it)? TMLR, 2025.

Charu C. Aggarwal. Recommender Systems: The Textbook. Springer, 2016.

Critical weaknesses
  1. The advantages of the circuit view of matrix factorization are unclear. The authors have discussed the benefits of the combination of the circuit and the probabilistic loss (Section 4), but have not discussed the individual benefits of the circuit. I consider a situation when the circuit is applied to a non-probabilistic loss. For example, classic latent factor models may not use BPR or NLL loss. They can be trained using a squared loss ∥∥R−UVT∥∥F2\|\|R - UV^T\|\|_F^2 (Section 3.6.4, Aggarwal, 2016), which is non-probabilistic and does not require negative sampling. In this case, is it still beneficial to view UVTUV^T as a circuit?

  2. The math intuition of the formulations of SQ\text{SQ} and μSQ\mu\text{SQ} is not well-explained. In Section 2,

  • Why must qSQq_{\text{SQ}} adopt complex-valued embeddings rather than real-valued embeddings?

  • Why does qμSQq_{\mu\text{SQ}} need to partition dd complex dimensions into SS disjoint blocks? And what does the term ∑k=1Keguk+hik\sum_{k=1}^K e^{g_{uk} + h_{ik}} mean in Eq. (4)?

  • Why do such formulations of SQ\text{SQ} and μSQ\mu\text{SQ} help achieve better recommendation performance? Are there any specific reasons that such formulations enable more effective information retrieval?

  1. Empirical results are not significant.
  • Sections 4.3 and 4.4 show that the circuit does not always effective in improving performance: "Addressing RQ3: neither conditional likelihood nor circuit parametrisation is consistently beneficial", "Addressing RQ4: additional circuit structure does not systematically improve ranking". These results are not significant, as they are too vague to illustrate the effectiveness of circuit. Authors should make the results explicit by showing under what cases circuit can bring relatively consistent performance improvement.

  • Section 4.2 shows that LNLL\mathcal{L}_{\text{NLL}} and LCNLL\mathcal{L}_{\text{CNLL}} yield similar ranking performance: "Addressing RQ2, joint and conditional maximum likelihood perform similarly for ranking". This result is still not significant, as it does not tell in what aspects we should prefer one loss over the other. Authors need to compare the advantages of LNLL\mathcal{L}_{\text{NLL}} and LCNLL\mathcal{L}_{\text{CNLL}} in some other aspects (e.g., convergence rate, computational cost, etc).

Reference:

Charu C. Aggarwal. Recommender Systems: The Textbook. Springer, 2016.

Clarification: My current assessment is based on the reason that I think authors' circuit view of matrix-factorization-based recommender system is interesting, but they do not provide convincing evidence showing how and why this view is effective in improving recommendation performance.

Questions for authors

See Weaknesses.

10 comments
AuthorsOpenReview ↗

Answer to Reviewer fdLv (1/6)

We thank the reviewer for the thoughtful feedback, recognizing the potential of our approach. We respond to each concern below.

1. “The advantages of the circuit view of matrix factorization are unclear”

Our goal is to revisit the interaction layer of factorisation-based recommenders and turn it into a tractable generative model, without changing the underlying encoder. With maximum-likelihood training, the circuit defines a normalised distribution over user--item interactions and supports probabilistic queries beyond ranking, such as p(u,i)p(u,i), p(u)p(u), p(i)p(i), p(u∣i)p(u \mid i), and p(I∈C∣u)p(I \in C \mid u). The circuit structure makes the computation of the partition function efficient without enumerating the full user--item score table. Joint and conditional normalisation cost O((U+I)d2)O((U+I){d}^{2}) and O((I+B)d2)O((I+B){d}^{2}), respectively, compared with O(UId)O(UId) and O(BId)O(BId) for exact dot-product softmax, where UU and II are the numbers of users and items, BB is the batch size and dd is the embedding width. Our new scalability experiments show that this difference becomes substantial as the catalogue grows (please see “Empirical results are not significant”). We therefore view the main advantage of the circuit formulation as enabling tractable generative modelling and maximum-likelihood learning, rather than as an intrinsic benefit of the parametrisation under a non-probabilistic loss.

Furthermore, we emphasise that our evaluation is not restricted only to recommenders originally trained with BPR. For example, SLMRec uses in-batch InfoNCE and feature-alignment contrastive losses, while BM3 uses a negative-free bootstrap objective. This supports the broader point that the effect of circuitisation cannot be reduced to replacing BPR or removing negative sampling. Consistently, our ablations show that neither circuit parametrisation under BPR nor likelihood training with a conventional dot product is sufficient to explain the gains; the strongest results arise from their combination. A whole-data squared-loss recommender is therefore a relevant non-sampling baseline, but it addresses a different question: it can avoid negative sampling without yielding the same tractable generative model and probabilistic queries. When negative sampling is present, maximum-likelihood training can additionally reduce training cost for lightweight encoders, whereas the trade-off can reverse when propagation or auxiliary objectives dominate computation.

2. “The math intuition of the formulations is not well-explained”

We thank the reviewer for asking for clarifications of the math formulation. The constructions are built on prior works connecting tensor factorisations and tractable circuits and from squared and sum-of-squares probabilistic circuits [Loconte et al., 2023; Loconte et al., 2025a,b]. Our goal is to preserve the factorised user--item computation while turning its output into a nonnegative mass that can be normalised exactly. Complex coordinates play a structural role in this construction. They allow the latent factors to remain unrestricted while, through squaring, producing a real nonnegative interaction, particularly μ\muSQ, where the key operation is the squared modulus. Starting from the complex interaction c(u,i)=∑rzurzirc(u,i)=\sum_r z_{ur}z_{ir}, we define ∣c(u,i)∣2=c(u,i)c(u,i)‾=(ℜc(u,i))2+(ℑc(u,i))2≥0|c(u,i)|^2=c(u,i)\overline{c(u,i)}=(\Re c(u,i))^2+(\Im c(u,i))^2\geq 0.

This is important because the embeddings zuz_u and ziz_i remain completely unconstrained: their entries can take any complex values, while the final interaction is guaranteed to be real and nonnegative. SQ uses a related construction,

qSQ(u,i)=[ℜ∑rzurzir‾]2q_{\mathrm{SQ}}(u,i) = \left[\Re \sum_r z_{ur}\overline{z_{ir}}\right]^2 where nonnegativity follows from squaring the real-valued score.

Thus, the complex parametrisation naturally realises a sum of two real squares while preserving the factorised structure required for tractable normalisation.

To test whether this additional representation capacity drives CiRe’s gains, instead of the whole proposed recipe, we perform a matched-width analysis: since the circuit heads use complex-valued embeddings, an embedding of width dd contains 2d2d real-valued coordinates, compared with dd for the original real-valued embeddings. We therefore independently re-tune all nine original recommenders on all five datasets using real embedding widths {128,512}\{128,512\}, corresponding to the real-valued capacity of circuit widths {64,256}\{64,256\}. Increasing the width improves the original recommenders in 34/4534/45 settings, but the average gain is only 0.00160.0016 Recall@20. We then compare SQ and μ\muSQ against both the original and the wider re-tuned baselines:

AuthorsOpenReview ↗

Answer to Reviewer fdLv (2/6)

ComparatorSQ: better / within-band / worseμSQ: better / within-band / worse
Original recommenders28 / 6 / 1132 / 6 / 7
Wider re-tuned originals24 / 10 / 1129 / 7 / 9

SQ still improves over the wider original in 24/45 settings, while μ\muSQ improves in 29/45; clear decreases occur in 11 and 9 settings, respectively. Thus, increasing the capacity of the original embeddings narrows the advantage in some cases but does not remove it, indicating that the observed ranking gains cannot be explained solely by the additional capacity. In the table, “within-band” denotes an absolute Recall@20 difference smaller than 0.002.

The complete per-setting results are:

EncoderModelClothingDBbookSportsBabyML-1M
BPRMFOriginal0.06200.26370.1004‾\underline{0.1004}0.08560.2997
Original, wider0.06460.2649‾\underline{0.2649}0.10370.08880.3018
SQ0.0649‾\underline{0.0649}0.25790.09340.07400.3019‾\underline{0.3019}
μSQ0.06820.26680.09950.0865‾\underline{0.0865}0.3044
LightGCNOriginal0.06290.26220.09430.09800.2890
Original, wider0.06410.27000.10050.09550.2927
SQ0.0678‾\underline{0.0678}0.2695‾\underline{0.2695}0.1076‾\underline{0.1076}0.0967‾\underline{0.0967}0.2980‾\underline{0.2980}
μSQ0.06910.26250.10780.09560.3025
NGCFOriginal0.05430.25620.09240.07540.3013
Original, wider0.05570.25520.0955‾\underline{0.0955}0.07740.2989
SQ0.0593‾\underline{0.0593}0.2591‾\underline{0.2591}0.09110.0781‾\underline{0.0781}0.3027‾\underline{0.3027}
μSQ0.06750.26720.09940.08100.3047
VBPROriginal0.07050.26860.10660.09440.3014
Original, wider0.07430.2695‾\underline{0.2695}0.10790.09500.2997‾\underline{0.2997}
SQ0.0998‾\underline{0.0998}0.26960.11680.0972‾\underline{0.0972}0.2881
μSQ0.10150.26840.1161‾\underline{0.1161}0.09750.2942
BM3Original0.07110.2586‾\underline{0.2586}0.09850.09600.1748
Original, wider0.0684‾\underline{0.0684}0.26470.10110.0954‾\underline{0.0954}0.1793
SQ0.06750.22920.09610.08030.2912‾\underline{0.2912}
μSQ0.06780.24590.0991‾\underline{0.0991}0.08230.2956
FREEDOMOriginal0.10190.2926‾\underline{0.2926}0.11610.10490.2717
Original, wider0.10250.29560.11770.10530.2761
SQ0.10730.28250.1232‾\underline{0.1232}0.1075‾\underline{0.1075}0.2981‾\underline{0.2981}
μSQ0.1049‾\underline{0.1049}0.28810.12500.11050.2992
PGLOriginal0.09130.2795‾\underline{0.2795}0.11420.10460.2934
Original, wider0.09390.28640.11450.1047‾\underline{0.1047}0.2927
SQ0.10500.26460.12380.10160.3045
μSQ0.0997‾\underline{0.0997}0.27230.1206‾\underline{0.1206}0.11000.3043‾\underline{0.3043}
SLMRecOriginal0.07000.22330.09650.07420.1844
Original, wider0.07170.22700.0971‾\underline{0.0971}0.07600.1827
SQ0.0729‾\underline{0.0729}0.23650.09550.0914‾\underline{0.0914}0.2924
μSQ0.07320.2273‾\underline{0.2273}0.10170.09160.2896‾\underline{0.2896}
SGLOriginal0.06520.26520.1066‾\underline{0.1066}0.07700.2810
Original, wider0.06420.26120.10600.07890.2835
SQ0.0675‾\underline{0.0675}0.27300.10740.0849‾\underline{0.0849}0.2970
μSQ0.07010.2714‾\underline{0.2714}0.10510.08520.2954‾\underline{0.2954}

Bold marks the best and underline the second-best value among the four rows of each encoder, for each dataset. These results show that increasing the embedding capacity of the original recommenders does not eliminate the observed circuit gains.

AuthorsOpenReview ↗

Answer to Reviewer fdLv (3/6)

Why the blocks and monotone factor in Eq. (4)? These are structural components of the more general μSOCS family, and their motivation is the expressiveness analysis of Loconte et al. [1]. A single squared circuit and a monotone circuit (one with only non-negative parameters) are incomparable: each represents some distributions compactly that the other can only represent with exponentially more parameters. Summing several compatible squares, a SOCS, can be exponentially more expressive than both (Theorem 2); a complex modulus square is already a sum of two real squares. Multiplying a monotone circuit by a SOCS gives a μSOCS, which contains both classes, since setting either factor to a constant recovers the other. Eq. (4) instantiates this construction for the user–item interaction: the SS blocks give a sum of SS complex squares, and mKm_K is the monotone factor. These results concern circuits over many variables, whereas our interaction has two, the user and the item; they motivate the family rather than guarantee a gain in our setting, which is why we test SS and KK empirically (RQ4).

Concretely, the dd coordinates are partitioned into SS disjoint blocks and the resulting modulus squares are added: ∑s=1S∣∑r∈Bszurzir∣2\sum_{s=1}^{S}\left|\sum_{r\in\mathcal B_s}z_{ur}z_{ir}\right|^2. The main μSQ model used in our experiments has S=1S=1, so no partition is performed. We vary SS only to study richer sum-of-squares structures while keeping the total width dd fixed. Increasing SS therefore changes the interaction structure rather than increasing its width: for example, ∣c1+c2∣2|c_1+c_2|^2 contains cross-block terms, whereas ∣c1∣2+∣c2∣2|c_1|^2+|c_2|^2 does not.

The other factor in Eq. (4), mK(u,i)=∑k=1Keguk+hikm_K(u,i)=\sum_{k=1}^{K}e^{g_{uk}+h_{ik}}, is a positive separable user--item weighting. Multiplying the squared interaction by this factor preserves nonnegativity and admits the weighted Gram-matrix normalisers derived in Section 2. Our main configuration again uses the simplest case, K=1K=1.

Why should these formulations improve recommendation? The complex squared construction is theoretically motivated by expressiveness. A complex interaction satisfies

∣c(u,i)∣2=(ℜc(u,i))2+(ℑc(u,i))2,|c(u,i)|^2=(\Re c(u,i))^2+(\Im c(u,i))^2,

so it naturally realises a sum of real squares rather than a single real square. Prior work [1] shows that such sum-of-squares constructions are provably more expressive than single-square circuits, providing a principled reason to use the complex formulation.

This increased expressiveness does not guarantee better top-kk recommendation in every setting, but it gives the model a richer interaction family while preserving nonnegativity and exact tractable normalisation. Our experiments then test whether this theoretical advantage translates into recommendation performance. We find that the strongest results arise when the circuit interaction is combined with likelihood-based training, whereas circuit parametrisation under BPR alone and likelihood training with a conventional dot product are not consistently beneficial.

We will clarify these aspects in the revised version of our manuscript.

[1] Loconte, L., et al. (2025). Sum of Squares Circuits. Proceedings of the AAAI Conference on Artificial Intelligence, 39(18), 19077–19085.

3. “Empirical results are not significant”

We thank the reviewer for this comment. We agree that the conclusions of RQ3 and RQ4 should state more explicitly when circuitisation is effective, and that RQ2 should clarify the practical distinction between joint and conditional maximum likelihood beyond ranking accuracy.

3a. On the Validity of RQ3 and RQ4

RQ3 identifies which part of the recipe is associated with these gains. In our controlled comparison, each circuit head is trained under BPR and conditional maximum likelihood with the same initialisation within that head: Conditional likelihood improves SQ in 9/9 comparisons and μSQ in 8/9. Conversely, circuit parametrisation under BPR alone is not consistently beneficial. Aiming to address their concern on the effectiveness of CiRe recipe, we extended the analysis of RQ3 to the full set of recommender and dataset:

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

This paper proposes CiRe, a framework for modifying the interaction functions and training objectives of factorization-based recommender systems. Instead of redesigning the user and item encoders, the authors replace the conventional dot-product scoring function with nonnegative squared interactions (SQ/µSQ). Building on existing probabilistic circuit theory, they exploit the algebraic structure of these interactions to compute exact normalization constants through Gram-matrix operations, enabling joint or conditional maximum-likelihood training without negative sampling in the primary recommendation objective.

The paper evaluates these scoring and training formulations across nine recommendation architectures and five datasets. It investigates their ranking performance, the interaction between scoring functions and training losses, joint versus conditional likelihood, circuit complexity, computational efficiency, and probability calibration

Critical strengths

1. Meaningful research direction.

The paper revisits scoring functions and training objectives rather than focusing exclusively on more complex user and item encoders. Investigating how interaction parameterization and learning objectives jointly affect recommendation quality is a worthwhile direction.

2. Mathematically meaningful construction.

The squared interactions enable exact probability normalization through Gram-matrix computations, avoiding explicit enumeration of all user–item pairs in the normalization step. Although the underlying probabilistic circuit theory is established, adapting it to recommendation and deriving the corresponding normalization formulas have practical mathematical value.

3. Broad empirical evaluation with encouraging results.

The experiments cover nine encoders and five datasets under full-catalog ranking evaluation. A circuit-based variant achieves the highest Recall@20 in 37 of 45 encoder–dataset settings, although improvements vary across architectures.

4. Useful preliminary mechanism analysis.

The authors conduct controlled comparisons between Dot, SQ, and µSQ under BPR and NLL, rather than attributing all improvements to the complete method. They also examine joint versus conditional likelihood, circuit complexity, and training costs. These experiments provide useful initial evidence about the interaction between scoring functions and objectives.

Critical weaknesses

1. The core methodological novelty is not sufficiently clear, especially the story under circiuts.

Despite the probabilistic circuit framing, the main methodological changes are replacing the interaction scoring function, changing the training objective, and investigating their combinations. The squared-circuit foundation is largely established in prior work, which the authors acknowledge.

The application is interesting, but it remains unclear what fundamentally new recommendation modeling capability is introduced beyond this reparameterization and its associated normalization technique. The contribution seems closer to a specific scoring-and-training recipe than a new general probabilistic recommendation framework.

2. The source of performance improvements is insufficiently explained.

The central question is why SQ/µSQ combined with MLE improves recommendation quality. Potential explanations include different interaction parameterization, denser full-catalog gradient updates, better sample utilization, reduced sampling variance, additional representation capacity (from real vectors to complex counterparts inceases the dimension by 2 times), or different optimization dynamics.

The paper provides useful factorial ablations, but these do not fully disentangle these possibilities. In particular, it remains unclear whether the main benefit comes from the circuit structure itself or from the learning signals and optimization behavior associated with the new objective.

Exact probability normalization is mathematically meaningful, but it does not inherently imply better recommendation learning. More direct evidence on gradient behavior, convergence, and sample efficiency would strengthen the central contribution.

3. The baselines and training-objective comparisons are not sufficiently competitive.

Full-softmax cross-entropy has already been extensively explored in recommendation. A representative example is SASRec+ (Klenitskiy and Vasilev, RecSys 2023), which demonstrates the importance of full-softmax training compared with conventional negative-sampling objectives. Related training strategies are also common in subsequent sequential and generative recommendation research.

Although sequential recommendation is not directly equivalent to the collaborative-filtering setting studied here, these developments illustrate that full-softmax training is an established alternative rather than a new capability unique to circuit-based models.

The paper mainly evaluates traditional collaborative-filtering and multimodal encoders, with limited coverage of more recent strong recommendation approaches. Existing non-sampling recommendation methods are discussed but not included in the main experimental comparisons.

The Dot+NLL comparison is useful, but covers only three encoders and three datasets. Moreover, results obtained with these encoders do not necessarily generalize to architectures designed around full-softmax learning. Stronger, appropriately matched comparisons with established full-softmax (especially comparing with SASRec+), sampled-softmax, and non-sampling approaches are needed to clarify the practical advantages of CiRe and isolate the effect of encoders (it is interesting to see whether the conclusion still holds for sequential encoders which can also be understood as a two-tower model).

4. The independent benefits of additional parameterization remain unclear.

The complex-valued adapters introduce a different parameterization and potentially additional effective representation capacity compared with conventional real-valued dot-product models. Strictly parameter-matched comparisons are missing.

The exponential factors in µSQ also deserve closer investigation. For the main configuration with K=1, the user-side factor cancels from the conditional item distribution, while the item-side factor acts as a global item-level scoring bias.

These factors can also be algebraically absorbed into embedding scales in the unconstrained scoring formulation. Their independent modeling contribution is therefore unclear.

In particular, the item-side bias may capture popularity-related effects. The paper does not sufficiently isolate these factors or examine whether the gains differ across head, torso, and tail items.

5. Practical scalability, retrieval efficiency, and generality are insufficiently established.

Exact Gram-matrix normalization requires statistics consistent with the current user and item representations. Maintaining these statistics during training may become expensive as the user and item populations grow.

The largest evaluated catalogue contains 22,498 items, substantially smaller than typical industrial-scale recommendation settings. Training efficiency is also architecture-dependent: several complex encoders become considerably slower under circuitization.

Another important concern is inference. The paper establishes efficient probability normalization but does not demonstrate how the squared scoring functions can be integrated with efficient top-k ANN/MIPS retrieval, or what their retrieval latency would be at scale.

Furthermore, the approach is restricted to particular factorized interaction structures. Its exact normalization mechanism does not directly generalize to arbitrary MLP-based ranking heads or complex user–item feature interactions.

These limitations do not invalidate the proposed construction, but constrain its practical applicability and the generality of the plug-in framework.

Questions for authors

Please address the concerns in the weakness part.

9 comments
AuthorsOpenReview ↗

Answer to Reviewer qCYV (1/7)

We thank the reviewer for the thoughtful feedback, recognizing the value of our approach, and for identifying the controls needed to assess this approach. Based on their comments, we provide additional results with baselines at increased capacity, full-softmax, sequential encoders, and larger catalogues. The relevant tables and experimental settings are included in this response. We address the five concerns below.

1.”It remains unclear what fundamentally new recommendation modeling capability is introduced”

We thank the reviewer for recognising the value of revisiting recommender interactions and objectives. While CiRe builds on established probabilistic-circuit theory and factorisation-based recommenders, we believe that combining established components can still constitute a novel and useful contribution when the resulting formulation enables new modelling capabilities. In CiRe, this combination turns the shared recommender interaction layer into a tractable probabilistic model without redesigning the underlying encoder.

Our contribution is therefore not a new squared-circuit family, but a new use of these constructions at the interaction layer shared by many factorisation-based recommenders. CiRe retains the encoder and replaces only its final factorised interaction with a nonnegative circuit, so that the resulting recommender defines an explicitly normalised distribution over user--item interactions and can be trained via maximum likelihood.

This changes what the recommender can represent and query. Beyond ranking items for a fixed user, the same model directly supports computing quantities such as $p(u,i)$, $p(u)$, $p(i)$, $p(u\mid i)$, set-valued probabilities such as $p(I \in C \mid u)$, and the probability of a recommendation list: ∑i∈Top20(u)p(i∣u)\sum_{i \in \mathrm{Top}_{20}(u)} p(i \mid u). A traditional recommender can also be adapted to estimate such quantities, e.g., via a softmax layer; however, that requires enumerating the complete user–item score table. In CiRe, the partition function and probabilistic queries can be computed much more efficiently through compact Gram contractions. The computational distinction is important: with $U$ users, $I$ items, batch size $B$, and width $d$, the normalisation costs are

NormaliserCircuit headExponential dot-product softmax
ConditionalO((I+B)d2)O((I+B)d^2)O(BId)O(BId)
JointO((U+I)d2)O((U+I)d^2)O(UId)O(UId)

The circuit expressions follow from the structured factorisation of the interaction: smoothness, decomposability and compatible squaring allow sums over users and items to collapse into Gram-matrix contractions. To demonstrate the effectiveness of CiRe, we added a scaling benchmark on synthetic catalogues (random embedding tables) with 100K users and up to 3M items, batch size 1024, and d=256d=256 complex dimensions, timing the loss and backward pass of one training step. We used a synthetic benchmark as the cost of the head depends only on the numbers of users and items, the batch size and the width, and none of our real catalogues approaches a million items (the largest has 179K, in the sequential experiments):

Head / objective10K items300K items1M items3M items
Dot / joint NLL142 ms4.81 s21.64 s83.6 s
SQ / joint NLL6.7 ms24 ms70 ms190 ms
μSQ / joint NLL7.0 ms25 ms67 ms192 ms
Dot / conditional NLL1.8 ms54 ms229 ms892 ms
SQ / conditional NLL2.1 ms18 ms63 ms184 ms
μSQ / conditional NLL2.7 ms19 ms63 ms187 ms

Thus, the main methodological contribution is a recommender-level generative reformulation of the shared factorised interaction that (i) preserves heterogeneous representation-learning encoders, (ii) yields an exactly normalised joint or conditional model, (iii) supports tractable probabilistic queries beyond ranking, and (iv) exploits circuit structure to make learning and inference substantially more scalable than a softmax.

CiRe’s maximum-likelihood training also removes negative sampling from the principal recommendation objective. For lightweight BPR-based encoders, this can reduce end-to-end training time since the cost of the partition function computation is smaller than the sampling overhead it replaces. We will revise the manuscript to make this distinction explicit: the squared-circuit constructions themselves are established; our novelty lies in using them as a replaceable interaction layer for factorisation-based recommender systems, deriving the corresponding recommender normalisers, and establishing the resulting probabilistic, computational, and recommendation trade-offs across heterogeneous architectures.

AuthorsOpenReview ↗

Answer to Reviewer qCYV (2/7)

2. “it remains unclear whether the main benefit comes from the circuit structure itself”

We agree that circuitisation alone does not imply better recommendation quality. We therefore added new experiments to test two major alternative explanations for the observed gains: additional representation capacity and full-catalogue likelihood training itself.

Additional representation capacity. Because the circuit heads use complex-valued embeddings, an embedding of width $d$ contains $2d$ real-valued coordinates, compared with $d$ for the original real-valued embeddings. We therefore independently re-tuned all nine original recommenders on all five datasets using real embedding widths {128,512}, corresponding to the real-valued capacity of circuit widths {64,256}. Increasing the width improves the original recommenders in (34/45) settings, but the average gain is only (0.0016) Recall@20. We then compare SQ and μ\muSQ against both the original and the wider re-tuned baselines:

ComparatorSQ: better / within-band / worseμSQ: better / within-band / worse
Original recommenders28 / 6 / 1132 / 6 / 7
Wider re-tuned originals24 / 10 / 1129 / 7 / 9

SQ still improves over the wider original in 24/45 settings, while μ\muSQ improves in 29/45; clear decreases occur in 11 and 9 settings, respectively. Thus, increasing the capacity of the original embeddings narrows the advantage in some cases but does not remove it, indicating that the observed ranking gains cannot be explained solely by the additional capacity. In the table, “within-band” denotes an absolute Recall@20 difference smaller than 0.002.

The complete per-setting results are:

EncoderModelClothingDBbookSportsBabyML-1M
BPRMFOriginal0.06200.26370.1004‾\underline{0.1004}0.08560.2997
Original, wider0.06460.2649‾\underline{0.2649}0.10370.08880.3018
SQ0.0649‾\underline{0.0649}0.25790.09340.07400.3019‾\underline{0.3019}
μSQ0.06820.26680.09950.0865‾\underline{0.0865}0.3044
LightGCNOriginal0.06290.26220.09430.09800.2890
Original, wider0.06410.27000.10050.09550.2927
SQ0.0678‾\underline{0.0678}0.2695‾\underline{0.2695}0.1076‾\underline{0.1076}0.0967‾\underline{0.0967}0.2980‾\underline{0.2980}
μSQ0.06910.26250.10780.09560.3025
NGCFOriginal0.05430.25620.09240.07540.3013
Original, wider0.05570.25520.0955‾\underline{0.0955}0.07740.2989
SQ0.0593‾\underline{0.0593}0.2591‾\underline{0.2591}0.09110.0781‾\underline{0.0781}0.3027‾\underline{0.3027}
μSQ0.06750.26720.09940.08100.3047
VBPROriginal0.07050.26860.10660.09440.3014
Original, wider0.07430.2695‾\underline{0.2695}0.10790.09500.2997‾\underline{0.2997}
SQ0.0998‾\underline{0.0998}0.26960.11680.0972‾\underline{0.0972}0.2881
μSQ0.10150.26840.1161‾\underline{0.1161}0.09750.2942
BM3Original0.07110.2586‾\underline{0.2586}0.09850.09600.1748
Original, wider0.0684‾\underline{0.0684}0.26470.10110.0954‾\underline{0.0954}0.1793
SQ0.06750.22920.09610.08030.2912‾\underline{0.2912}
μSQ0.06780.24590.0991‾\underline{0.0991}0.08230.2956
FREEDOMOriginal0.10190.2926‾\underline{0.2926}0.11610.10490.2717
Original, wider0.10250.29560.11770.10530.2761
SQ0.10730.28250.1232‾\underline{0.1232}0.1075‾\underline{0.1075}0.2981‾\underline{0.2981}
μSQ0.1049‾\underline{0.1049}0.28810.12500.11050.2992
PGLOriginal0.09130.2795‾\underline{0.2795}0.11420.10460.2934
Original, wider0.09390.28640.11450.1047‾\underline{0.1047}0.2927
SQ0.10500.26460.12380.10160.3045
μSQ0.0997‾\underline{0.0997}0.27230.1206‾\underline{0.1206}0.11000.3043‾\underline{0.3043}
SLMRecOriginal0.07000.22330.09650.07420.1844
Original, wider0.07170.22700.0971‾\underline{0.0971}0.07600.1827
SQ0.0729‾\underline{0.0729}0.23650.09550.0914‾\underline{0.0914}0.2924
μSQ0.07320.2273‾\underline{0.2273}0.10170.09160.2896‾\underline{0.2896}
SGLOriginal0.06520.26520.1066‾\underline{0.1066}0.07700.2810
Original, wider0.06420.26120.10600.07890.2835
SQ0.0675‾\underline{0.0675}0.27300.10740.0849‾\underline{0.0849}0.2970
μSQ0.07010.2714‾\underline{0.2714}0.10510.08520.2954‾\underline{0.2954}

Bold marks the best and underline the second-best value among the four rows of each encoder, for each dataset.

AuthorsOpenReview ↗

Answer to Reviewer qCYV (3/7)

These results show that increasing the embedding capacity of the original recommenders does not eliminate the observed circuit gains.

Full-catalogue likelihood alone. We further extend the exact dot-product full-softmax comparison to all nine encoders and five datasets, testing whether the improvement can be attributed to CiRe or maximum-likelihood training.

Here, C and J denote conditional and joint maximum likelihood, respectively. Summarising the 45 settings using the 0.002 descriptive band (the per-setting values are given under "Baselines and Training Objectives" below):

ComparisonBetter / within-band / worse
Dot/cNLL vs. original20 / 8 / 17
SQ/cNLL vs. Dot/cNLL29 / 5 / 11
μSQ/cNLL vs. Dot/cNLL29 / 10 / 6
SQ/jNLL vs. Dot/cNLL28 / 5 / 12
μSQ/jNLL vs. Dot/cNLL32 / 7 / 6

The dot-product full-softmax model is therefore not consistently stronger than the original recommender: it improves in 20/45 settings and is lower in 17/45. In contrast, under the same conditional likelihood, SQ and μSQ outperform Dot in 29/45 settings each, with only 11 and 6 losses, respectively. These results are consistent with our controlled score--objective ablation. Under BPR, replacing Dot with SQ or μSQ improves only 2/18 comparisons. Conversely, replacing BPR with conditional likelihood improves the conventional Dot interaction in only 2/9 comparisons.

Taken together, these experiments rule out two important explanations: the gains are not explained by additional embedding capacity, and they are not explained by maximum-likelihood training alone. Instead, they are most consistently associated with the interaction between the circuit parametrisation and CiRe’s loss.

We will therefore clarify the corresponding claim in the manuscript.

3. ”Stronger, appropriately matched comparisons are needed”

We thank the reviewer for this important point. We agree that full-softmax training is an established alternative in recommendation, including in SASRec+ and subsequent sequential recommender.

Sequential recommendation constitutes a distinct prediction setting from the collaborative-filtering formulation studied in our main experiments, since the user representation is conditioned on an ordered interaction history and the task is to predict the next item. Nevertheless, we found the reviewer’s suggestion valuable for testing whether the proposed interaction conversion extends beyond the original setting. We therefore added experiments with GRU4Rec and SASRec, building on the Amazon Reviews 2023 sequential-recommendation setup used by Hou et al. [1], including the All Beauty, Video Games, and Baby Products domains. We retain this task formulation and encoder families while evaluating our interaction heads under full-catalogue ranking. Together with the expanded full-softmax, and sampled-softmax, these experiments provide a broader test of CiRe across both classic and sequential recommender representations.

[1] Hou, Yupeng, et al. "Bridging language and items for retrieval and recommendation: Benchmarking LLMs as semantic encoders." Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Full-softmax across all encoders and datasets. We extend the Dot+NLL control from the original 3×3 study to all nine encoders and five datasets, using an exact conditional softmax over the full catalogue. The resulting Recall@20 comparison is:

Reviewer_f1quOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Summary

This paper suggests to replace the dot-product with a non-negative squared circuit in existing factorization-based recommender systems such as MF, NeuMF, etc. The authors conducted experiments on popular RecSys datasets.

Critical strengths

Strengths:

  • The approach is simple; somewhat creative
  • Various experiments were conducted across 9 baselines and 5 datasets
  • Source code is provided
Critical weaknesses

Weaknesses:

  • Contributions are very marginal, and results are not impressive.
  • According to Table 1, mu_SQ and SQ are not always comparable, and not all results are statistically significant
  • Figure 1 and Table 1 represent the same results?
  • More metrics should be reported, not only Recall@20

This paper is not ready for ICLR this year.

Questions for authors

Please refer to my weaknesses above and defense your paper.

5 comments
AuthorsOpenReview ↗

Answer to Reviewer f1qu (1/4)

We thank the reviewer for recognising the simplicity of the approach, the breadth of the evaluation, and the availability of the source code. We address the concerns below and, based on their comments, added several experiments to strengthen both the methodological positioning and the empirical evidence.

1. “Contributions are very marginal, and results are not impressive.”

In this work, we introduce CiRe, which replaces the final dot-product interaction of factorisation-based recommender systems with a probabilistic circuit. CiRe builds on squared-circuit constructions to convert the original recommendation model into a generative model that can be trained by maximum likelihood, while leaving the underlying representation-learning architecture unchanged.

This probabilistic formulation supports queries beyond ranking, including joint and marginal probabilities such as (p(u,i)), (p(u)), and (p(i)), reverse conditionals such as (p(u\mid i)), category-level probabilities such as (p(I\in C\mid u)), and the probability mass assigned to a recommendation list. Although a softmax model can define the same types of probabilities, the circuit structure enables more efficient exact normalisation through Gram-matrix contractions. In CiRe, conditional and joint normalisation cost (O((I+B)d^2)) and (O((U+I)d^2)), respectively, compared with (O(BId)) and (O(UId)) for the corresponding exact softmax formulations, where (I) is the number of items, (U) the number of users, (B) the batch size, and (d) the embedding dimension.

2. “not all results are statistically significant”

We agree that SQ and μSQ should not be presented as uniformly superior: they are distinct interaction functions and can behave differently across encoders and datasets. Regarding statistical significance, in the submitted results, 33 of the 37 settings in which the better circuit head achieves higher Recall@20 and nDCG@20 are marked significant at the nominal p<0.05p<0.05 level using the reported paired t-test per-user.

3. “Figure 1 and Table 1 represent the same results?”

Yes. Figure 1 is a visual summary of a subset of the ranking results reported numerically in Table 1; it is not an independent experiment. We will make this relationship explicit in the caption to avoid suggesting otherwise.

4. “More metrics should be reported, not only Recall@20”

Table 1 already reports both Recall@20 and nDCG@20 across all nine encoders and five datasets. We will make the nDCG results more prominent in the text. Furthermore, the paper includes a calibration diagnostic, showing that the probabilities of CiRe are better calibrated than baseline recommenders.

Aiming to answer the reviewer’s concern, we substantially broaden the experimental scope of the paper. In particular, we have extended the ablation of RQ3 to the full set of recommenders and all the datasets considered, showing that the trend generalises across recommenders-datasets combinations.

DatasetEncoderOriginalDot+NLLSQ/CμSQ/CSQ/JμSQ/J
ClothingBPRMF0.06200.06100.06320.06870.06490.0682‾\underline{0.0682}
ClothingLightGCN0.06290.06540.06840.07020.06780.0691‾\underline{0.0691}
ClothingNGCF0.05430.06900.06010.0687‾\underline{0.0687}0.05930.0675
ClothingVBPR0.07050.07100.09960.09850.0998‾\underline{0.0998}0.1015
ClothingBM30.07110.05060.06660.06770.06750.0678‾\underline{0.0678}
ClothingFREEDOM0.10190.09920.10470.10380.10730.1049‾\underline{0.1049}
ClothingPGL0.09130.09660.1036‾\underline{0.1036}0.10070.10500.0997
ClothingSLMRec0.07000.06080.07220.07080.0729‾\underline{0.0729}0.0732
ClothingSGL0.06520.06760.06780.0698‾\underline{0.0698}0.06750.0701
DBbookBPRMF0.26370.26100.26380.26890.25790.2668‾\underline{0.2668}
DBbookLightGCN0.26220.2763‾\underline{0.2763}0.27660.27440.26950.2625
DBbookNGCF0.25620.25560.24250.26940.25910.2672‾\underline{0.2672}
DBbookVBPR0.26860.26470.27360.26940.2696‾\underline{0.2696}0.2684
DBbookBM30.25860.22980.23610.23230.22920.2459‾\underline{0.2459}
DBbookFREEDOM0.29260.27080.28260.28640.28250.2881‾\underline{0.2881}
DBbookPGL0.27950.27150.27420.2778‾\underline{0.2778}0.26460.2723
DBbookSLMRec0.22330.21910.22460.22550.23650.2273‾\underline{0.2273}
DBbookSGL0.26520.27670.26970.26850.2730‾\underline{0.2730}0.2714
SportsBPRMF0.10040.10040.09480.09850.09340.0995‾\underline{0.0995}
AuthorsOpenReview ↗

Answer to Reviewer f1qu (2/4)

DatasetEncoderOriginalDot+NLLSQ/CμSQ/CSQ/JμSQ/J
SportsLightGCN0.09430.10540.10710.10610.1076‾\underline{0.1076}0.1078
SportsNGCF0.09240.10610.08980.09730.09110.0994‾\underline{0.0994}
SportsVBPR0.10660.09380.11500.11760.1168‾\underline{0.1168}0.1161
SportsBM30.09850.08160.09700.10020.09610.0991‾\underline{0.0991}
SportsFREEDOM0.11610.11970.12110.1247‾\underline{0.1247}0.12320.1250
SportsPGL0.11420.11910.12520.12140.1238‾\underline{0.1238}0.1206
SportsSLMRec0.09650.08430.0992‾\underline{0.0992}0.09670.09550.1017
SportsSGL0.10660.10670.10850.1074‾\underline{0.1074}0.1074‾\underline{0.1074}0.1051
BabyBPRMF0.08560.08050.08280.08900.07400.0865‾\underline{0.0865}
BabyLightGCN0.09800.09110.09390.09030.0967‾\underline{0.0967}0.0956
BabyNGCF0.07540.0833‾\underline{0.0833}0.07470.08470.07810.0810
BabyVBPR0.09440.08690.09950.09600.09720.0975‾\underline{0.0975}
BabyBM30.09600.06300.07850.0851‾\underline{0.0851}0.08030.0823
BabyFREEDOM0.10490.11150.10780.11030.10750.1105‾\underline{0.1105}
BabyPGL0.10460.11340.10090.10440.10160.1100‾\underline{0.1100}
BabySLMRec0.07420.07560.09310.05460.09140.0916‾\underline{0.0916}
BabySGL0.07700.09370.0853‾\underline{0.0853}0.08480.08490.0852
ML-1MBPRMF0.29970.30000.29770.29760.3019‾\underline{0.3019}0.3044
ML-1MLightGCN0.28900.29220.29620.3013‾\underline{0.3013}0.29800.3025
ML-1MNGCF0.30130.29960.29670.29980.3027‾\underline{0.3027}0.3047
ML-1MVBPR0.30140.28570.2949‾\underline{0.2949}0.29090.28810.2942
ML-1MBM30.17480.27690.28970.29100.2912‾\underline{0.2912}0.2956
ML-1MFREEDOM0.27170.29540.3021‾\underline{0.3021}0.30440.29810.2992
ML-1MPGL0.29340.30600.3083‾\underline{0.3083}0.30970.30450.3043
ML-1MSLMRec0.18440.28520.29650.29100.2924‾\underline{0.2924}0.2896
ML-1MSGL0.28100.28570.28840.28740.29700.2954‾\underline{0.2954}

C and J denote conditional and joint maximum likelihood. Bold marks the best and underline the second-best value in each row. In summary:

ComparisonBetter / within-band / worse
Dot/cNLL vs. original20 / 8 / 17
SQ/cNLL vs. Dot/cNLL29 / 5 / 11
μSQ/cNLL vs. Dot/cNLL29 / 10 / 6
SQ/jNLL vs. Dot/cNLL28 / 5 / 12
μSQ/jNLL vs. Dot/cNLL32 / 7 / 6

Each entry counts the settings in which the first model's test Recall@20 is higher by at least 0.002, within 0.002 (within band), or lower by at least 0.002. Full-softmax improves over the original recommender in some settings, but not consistently. More importantly, under the same conditional likelihood, SQ and μSQ outperform Dot substantially more often than they underperform it. This supports the conclusion that the gains cannot be attributed to full-catalogue likelihood training alone.

Under the same conditional likelihood objective cNLL, μ\muSQ matches or outperforms full softmax in 39 of 45 settings (29 better, 10 within-band). It also requires fewer operations per training step: full softmax scores all II items for each of the BB users in a batch, with cost O(BId)O(BId). In contrast, the conditional circuit first builds an item Gram matrix and then evaluates the batch, with cost O((I+B)d2)O((I+B){d}^{2}). The table below compares these operations for B=1,024B=1,024 and real width d=512d=512, the width selected in most runs.

DatasetItemsDot+NLL (GFLOP)SQ/C, μSQ/C (GFLOP)Ratio
Clothing22,49870.837.01.9×
DBbook5,21716.49.81.7×
Sports18,28557.530.41.9×
Baby7,03722.112.71.7×
ML-1M3,0759.76.41.5×

At real width 128128, the circuit requires 6.06.0–7.7×7.7\times fewer operations than conditional softmax. These counts exclude encoder and optimiser work, which is shared by both methods. In practice, the circuit is still slightly slower at 1010K items, but becomes faster as the catalogue grows: about 1.7×1.7\times at 100100K items, 3.6×3.6\times at 11M, and 4.8×4.8\times at 33M. Joint circuit training is more expensive than conditional softmax on these small datasets because it also normalises over users, but compared with an exact joint softmax it requires 44–28×28\times fewer operations.

AuthorsOpenReview ↗

Answer to Reviewer f1qu (3/4)

Furthermore, we extended the experimental evaluation also to sequential recommendation, which serves as an additional task to test the generalisation of the CiRe: we evaluate GRU4Rec and SASRec on Amazon Reviews 2023 with full-catalogue ranking, following the evaluation protocol of [1], comparing their conventional full-softmax head against SQ and μSQ, and reporting two additional metrics (nDCG@10 and Recall@10 respectively):

DatasetEncoderFull softmax (nDCG@10 / Recall@10)SQ (nDCG@10 / Recall@10)μSQ (nDCG@10 / Recall@10)
All Beauty (~44K items)GRU4Rec0.0015 / 0.00310.0014‾\underline{0.0014} / 0.00310.0015 / 0.0027‾\underline{0.0027}
All Beauty (~44K items)SASRec0.0018‾\underline{0.0018} / 0.0033‾\underline{0.0033}0.0020 / 0.00360.0017 / 0.0029
Video Games (~116K items)GRU4Rec0.0114 / 0.02110.0107 / 0.0200‾\underline{0.0200}0.0110‾\underline{0.0110} / 0.0200‾\underline{0.0200}
Video Games (~116K items)SASRec0.0110‾\underline{0.0110} / 0.02150.0066 / 0.01280.0114 / 0.0211‾\underline{0.0211}
Baby Products (~179K items)GRU4Rec0.0056 / 0.01100.0060‾\underline{0.0060} / 0.0121‾\underline{0.0121}0.0061 / 0.0124
Baby Products (~179K items)SASRec0.0065‾\underline{0.0065} / 0.0132‾\underline{0.0132}0.0027 / 0.00640.0072 / 0.0140

Bold marks the best and underline the second-best value of each metric in each row.

On Baby Products with SASRec (179K items, batch 2,048) the full-softmax head costs 660 GFLOP per training step (matrix products, forward and backward), μSQ 195 and SQ 98. The circuit heads therefore cut the head's operations by 3.4× (μSQ) and 6.7× (SQ), and the whole step's by 35% and 42%.

These experiments show a more nuanced picture than universal improvement: μSQ remains competitive with full-softmax sequential models across substantially larger catalogues and improves both metrics in most settings, highlighting that circuit interaction can extend beyond collaborative filtering and multimodal encoders, while its effectiveness remains architecture-dependent.

Finally, aiming to better disentangle whether the source of the performance is related to the complex parametrisation, after having disentangled the circuit ones and the contribution of the CiRe’s loss, we run an additional experimental validation: because the circuit heads use complex-valued embeddings, an embedding of width $d$ contains $2d$ real-valued coordinates, compared with $d$ for the original real-valued embeddings. We therefore independently re-tuned all nine original recommenders on all five datasets using real embedding widths {128,512}, corresponding to the real-valued capacity of circuit widths {64,256}. Increasing the width improves the original recommenders in (34/45) settings, but the average gain is only (0.0016) Recall@20. We then compare SQ and (\mu)SQ against both the original and the wider re-tuned baselines:

ComparatorSQ: better / within-band / worseμSQ: better / within-band / worse
Original recommenders28 / 6 / 1132 / 6 / 7
Wider re-tuned originals24 / 10 / 1129 / 7 / 9

SQ still improves over the wider original in 24/45 settings, while (\mu)SQ improves in 29/45; clear decreases occur in 11 and 9 settings, respectively. Thus, increasing the capacity of the original embeddings narrows the advantage in some cases but does not remove it, indicating that the observed ranking gains cannot be explained solely by the additional capacity. In the table, “within-band” denotes an absolute Recall@20 difference smaller than 0.002.

The complete per-setting results are:

EncoderModelClothingDBbookSportsBabyML-1M
BPRMFOriginal0.06200.26370.1004‾\underline{0.1004}0.08560.2997
Original, wider0.06460.2649‾\underline{0.2649}0.10370.08880.3018
SQ0.0649‾\underline{0.0649}0.25790.09340.07400.3019‾\underline{0.3019}
μSQ0.06820.26680.09950.0865‾\underline{0.0865}0.3044
LightGCNOriginal0.06290.26220.09430.09800.2890
Original, wider0.06410.27000.10050.09550.2927
SQ0.0678‾\underline{0.0678}0.2695‾\underline{0.2695}0.1076‾\underline{0.1076}0.0967‾\underline{0.0967}0.2980‾\underline{0.2980}
μSQ0.06910.26250.10780.09560.3025
NGCFOriginal0.05430.25620.09240.07540.3013
Original, wider0.05570.25520.0955‾\underline{0.0955}0.07740.2989
SQ0.0593‾\underline{0.0593}0.2591‾\underline{0.2591}0.09110.0781‾\underline{0.0781}0.3027‾\underline{0.3027}
μSQ0.06750.26720.09940.08100.3047
VBPROriginal0.07050.26860.10660.09440.3014

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Answer to Reviewers' Comments

We thank all reviewers for the careful and constructive assessment. Across the reviews, we identified three recurring questions: what the circuit formulation contributes beyond a scoring reparametrisation, whether the ranking gains survive stronger controls, and whether exact probabilistic modelling provides a practical benefit.

Contribution and interpretation. In this work we propose CiRe, a method where the final dot product in factorisation-based recommender architectures is replaced with a probabilistic circuit. Specifically, CiRe builds on squared-circuit constructions to turn factorisation-based recommendation models into generative models that can be trained via maximum likelihood while retaining the original representation learning architecture.

CiRe allows to compute joint and marginal probabilities such as $p(u,i)$, $p(u)$, $p(i)$; conditional queries such as p(u∣i)p(u\mid i),, category-level probabilities as p(I∈C∣u)p(I\in C\mid u),, and the probability of a recommendation list. Softmax models can also define such probabilities; however, in probabilistic circuits, the advantage is that the partition function can be computed more efficiently via Gram contractions. In CiRe, conditional and joint normalisation respectively cost O((I+B)d2)O((I+B){d}^{2}) and O((U+I)d2)O((U+I){d}^{2}), while with the softmax these raise to O(BId)O(BId) and O(UId)O(UId); here, $I$ represents the item set, $B$ the batch, $U$ the user set, and $d$ the embedding dimensions.

Stronger ranking analysis. To strengthen the significance of our results, we substantially expanded the empirical analysis. First, we independently re-tuned all nine original recommenders by increasing their capacity to match the complex circuit representations. The trends largely persists: in the following we show that, in the vast majority of cases, even increasing the baseline capacity, SQ and μSQ still yield more accurate results (“within-band” denotes an absolute Recall@20 difference smaller than 0.002):

ReferenceSQ/J: higher / within-band / lowerμSQ/J: higher / within-band / lower
Wider re-tuned originals, 45 settings24 / 10 / 1129 / 7 / 9
Dot + sampled softmax, 15 settings7 / 2 / 68 / 4 / 3

We also extend exact Dot+NLL/full-softmax training to all nine encoders and five datasets. Under the same conditional likelihood, SQ is higher / within-band / lower than Dot+NLL in 29/5/11 settings, and μSQ in 29/10/6. These results indicate that neither increased embedding width nor dense full-catalogue likelihood alone explains the observed gains; the accuracy improvements are explained by the interaction of the proposed circuit parametrisations and training losses.

Extension beyond the original encoder families. Following the suggestions by qCYV, we also evaluated GRU4Rec and SASRec on Amazon Reviews 2023 with catalogues ranging from approximately 44K to 179K items. μSQ remains competitive with conventional full-softmax training and improves both nDCG@10 and Recall@10 in different settings.

Scalability. To analyse the scalability properties of CiRe, we benchmarked exact loss and backward computation on synthetic catalogues (random embedding tables) with 100K users and up to 3M items, showing SQ and μSQ maintain their scalability properties:

HeadConditional, 1M itemsJoint, 1M items
Dot229 ms21.64 s
SQ63 ms70 ms
μSQ63 ms67 ms

With 3M items and 1,024 user representations, evaluating ten fixed item groups only takes 64 ms for SQ when summaries are first constructed and 0.62 ms when they are cached.

Scope and presentation. We will clarify several presentation points in the final version of the paper. Table 1 already shows both Recall@20 and nDCG@20, while Figure 1 is intended only as a visual summary on a subset of results, providing an overview of the main empirical trends.

Overall, our additional experiments show that transforming the shared factorised interaction of heterogeneous recommenders into a tractable generative model, while preserving or improving ranking performance under stronger controls and at substantially larger catalogue scales, is a promising research direction. We have incorporated all of these new results and analyses into the revised manuscript, where they are reported in the Appendix and highlighted in blue for ease of review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.