{"id":"50436d8d-23d7-4433-83e7-cea67cc5896e","arxiv_id":"2506.12661","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"INTERPOS improves mobile app next-app recommendation by fusing inter-interaction time intervals into transformer position embeddings via three fusion strategies.","lead":"This paper introduces INTERPOS, a technique that adds time-gap information, called user rhythm, to transformer position embeddings for mobile app recommendation. It reports large gains over existing sequential recommenders on seven app datasets, but the evaluation has several unresolved rigor issues.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates the rhythm signal from added model capacity; a permutation control is needed to support the central claim.","rationale":"The reader's weakest_assumption identifies the same core issue: the lack of an ablation separating the rhythm signal from added parameters. This is the most load-bearing concern because the paper's headline claim is specifically about the value of interaction rhythm, not about the value of extra parameters. If a capacity-matched or permuted-rhythm control reproduces the gains, the proposed mechanism is not supported. The chronological split issue is secondary: even with a fully chronological split, the missing ablation would still undermine the interpretation. The paper has strengths—code is released, results are consistently reported across 7 datasets, and the direction of the effect is plausible for sparse mobile-app sequences—but the missing control is a genuine gap. I therefore keep the reader's CONDITIONAL verdict: the paper should be accepted only after the rhythm-signal ablation is performed and reported.","tokens_in":16902,"tokens_out":4474,"duration_ms":49500,"concrete_test":"Run the INTERPOS-MF variant with LightSANs on the Action dataset (Table II) under identical hyperparameters, but replace each user's true inter-interval sequence Δt_i with a random permutation of those same Δt values, keeping the architecture and parameter count fixed. Compare NDCG@10. If the permuted model retains roughly the reported 0.0386 instead of falling toward the LightSANs baseline of 0.0142, the gains are not attributable to rhythm information.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that injecting interaction rhythm into position embeddings drives the large gains (39–158%) over baselines. However, every INTERPOS variant adds parameters beyond the base model: a rhythm embedding table whose size depends on the maximum normalized Δt, and, for MF and GF, additional MLP layers. The baselines do not have these parameters. No experiment controls for this capacity difference. The reported gains could therefore come from the extra parameters alone, not from the rhythm information. A direct test would be to destroy the rhythm signal while keeping the architecture and parameter count identical—for instance, by permuting the Δt values within each user sequence. If performance remains near the reported INTERPOS numbers, the central interpretation is unsupported; if it drops toward the base LightSANs/SASRec baselines, the rhythm signal is indeed load-bearing. Without such a control, the paper does not establish that rhythm, rather than capacity, causes the improvements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes INTERPOS, a family of positional-morphing strategies that inject user interaction rhythm (time gaps between consecutive interactions) into the position embeddings of autoregressive Transformer-based sequential recommenders. Three fusion mechanisms are introduced: basic additive fusion (BF), MLP-based fusion (MF), and gated fusion (GF). INTERPOS is integrated into LightSANs and SASRec and evaluated on seven mobile app recommendation datasets (six category subsets of MobileRec plus the full MobileRec) using NDCG@K and HIT@K. The authors report large relative gains over strong sequential and time-aware baselines, and they release the source code.","tokens_in":17064,"tokens_out":4172,"duration_ms":48448,"significance":"If the reported gains are attributable to the rhythm signal, the paper would make a useful contribution: it demonstrates that temporal gap information is better injected at the embedding layer for sparse, slow-paced mobile app interaction sequences, where existing time-aware self-attention mechanisms underperform. The work also extends evaluation to a relatively underexplored domain. The paper provides open-source code, evaluates on seven datasets including a large-scale dataset, and includes time-aware baselines such as TiSASRec and FEARec, which is commendable. However, the central interpretation—that the improvements are driven by rhythm information rather than by added model capacity—is not yet supported by the experimental design.","major_comments":[{"comment":"The central claim that interaction rhythm drives the large improvements is not isolated from the increase in model capacity. Every INTERPOS variant adds parameters beyond the base model: a rhythm embedding table whose size depends on the maximum normalized time gap, and for MF and GF one or more MLP/gating layers. The baselines do not have these parameters. Report a capacity-matched control, for example by permuting the Δt values within each user sequence while keeping the architecture and parameter count identical. If performance remains at the reported INTERPOS level after permutation, the gains cannot be attributed to rhythm; if it drops to the base LightSANs/SASRec level, the rhythm signal is shown to be load-bearing. Also test an equal-capacity variant that replaces the rhythm embedding table with a randomly initialized learnable table independent of Δt.","section":"Section V, 'Summary of Results' and Section IV.A"},{"comment":"If the authors keep the significance claim, they must report the test procedure and the resulting p-values; otherwise the claim is unverifiable.","section":"Section IV.A and Section V, 'Summary of Results'"},{"comment":"These inconsistencies undermine the reader's ability to trust the quantitative claims, even though the direction of the results is consistent.","section":"Section V, 'Performance on RolePlaying dataset' and Section I"},{"comment":"This is important because the method is specifically about temporal gaps; a non-chronological split could artificially favor models that use time information.","section":"Section IV.A, 'Experimental Setup'"},{"comment":"The normalization factor (0.2) and clipping threshold (800) for the interaction rhythm are chosen constants, and no sensitivity analysis is provided. The paper should report how performance varies with these hyperparameters over a reasonable range, since they directly control the granularity of the rhythm embedding space. Without this analysis, the robustness of the method to these two free parameters is unclear.","section":"Section IV.A, last paragraph"}],"minor_comments":[{"comment":"In the definition of absolute positions, the set is written as P = {P1, P1, ..., Pn}, where 'P1' is repeated; this appears to be a typo for P2.","section":"Section II"},{"comment":"The list of selected categories reads 'including (Action, RolePlaying, Puzzle, Casual, Simulation, Strategy, and Simulation)' — 'Simulation' is listed twice; the set of six categories plus the full MobileRec should be stated correctly.","section":"Section IV.B"},{"comment":"There is a typo 'INTREPOS-MF' that should be 'INTERPOS-MF'.","section":"Section V, 'Performance on Casual dataset'"},{"comment":"The sentence 'the results of these strategies are statistically insignificant when compared with with another' contains a duplicated 'with' and a grammatical error; also 'another' should be 'each other'.","section":"Section V, 'Summary of Results'"},{"comment":"The percentage values quoted for improvements on the Action dataset in the Introduction appear to be a mixture of NDCG@10 and NDCG@20 values; please separate them clearly and align them with the tables.","section":"Section I and Section V"},{"comment":"Figure 2 reports the percentage of same-day interactions but does not show the actual values or the number of consecutive-interaction pairs per dataset; adding these details would improve reproducibility and interpretability.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a simple, plausible idea and consistent directional results across many datasets, but the experimental evidence does not yet establish that the gains come from rhythm rather than extra capacity. The missing permutation control and the unsupported significance claims are the main blockers. The percentage mismatches are numerous enough that I would ask for a full audit of the numbers before resubmission. The dataset coverage and open-source code are strengths. Fit with a CS-IR venue is acceptable, though the novelty is incremental relative to TiSASRec and FEARec."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core idea—fusing inter-interaction time gaps into position embeddings instead of into the attention layer—is genuinely new relative to TiSASRec and FEARec, and it is simple enough to be useful. Second, the evaluation, as written, does not support the headline gains; the reported percentages are arithmetically off, and the paper lacks the one control that would make the causal claim stick.\n\nWhat's good: The three fusion variants (addition, MLP, gated) are clean. The paper integrates them into both LightSANs and SASRec, evaluates on seven datasets including MobileRec, and the win direction is consistent across tables. Code and data links are provided. That's a solid empirical skeleton.\n\nSoft spots: The numbers in the text don't match the tables. For example, on RolePlaying, INTERPOS-GF vs GRU4Rec on NDCG@10 is about 102.9% per Table III, not the 88.62% claimed. The Action section reports NDCG@10 improvements but the abstract says NDCG@20. That makes the 'up to 158%' claims unreliable. Also, there are no error bars or significance tests, despite the paper claiming 'statistically significant' performance. The biggest issue is that no ablation separates the rhythm signal from added parameters. Every INTERPOS variant adds a rhythm embedding table and, for MF/GF, extra MLPs. The stress-test is right: permuting Δt within each user sequence keeps architecture and parameter count identical and would tell you whether the rhythm content actually matters. Without that, the gains could come from capacity alone, not from rhythm. Finally, the split description is thin: 'leave-one-out' without stating that the held-out item is the chronologically last one. In sequential recommendation that is the standard, but it should be explicit.\n\nNone of these are fatal, but together they mean the paper overstates its case. The central direction is plausible and worth checking, but the current evidence is conditional. This paper is for anyone working on time-aware sequential recommendation or mobile app recommendation. It deserves a serious referee if the authors add a permutation control, fix the arithmetic, and report variance. I'd send it to review with major revision, not desk reject.\n\nMy recommendation: engage with it, but require the control.","headline":"New idea for time-aware sequential recommendation, but evaluation needs a permutation control and the arithmetic doesn't add up.","tokens_in":17586,"tokens_out":3381,"would_cite":true,"duration_ms":38531,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"INTERPOS fuses a user's interaction time gaps into positional embeddings, and across seven mobile app datasets its three fusion variants outperform sequential and time-aware baselines on NDCG@K and HIT@K.","keywords":["mobile app recommendation","sequential recommendation","position embedding","user interaction rhythm","inter-interaction time gap","transformer self-attention","next-item prediction","time-aware recommendation"],"falsifier":"Retrain any INTERPOS variant on the same seven splits after replacing each user's inter-interaction gaps with gaps drawn from other users' histories, keeping the rhythm embedding table and fusion-layer sizes identical; if NDCG@20 stays within noise of the reported values, the rhythm content itself is not what drives the improvement.","tokens_in":16704,"feed_emoji":"📱","tokens_out":9330,"duration_ms":103433,"temperature":0.7,"pith_summary":"The paper's central claim is that in mobile app recommendation, the time gaps between a user's consecutive app interactions—what it calls the user's 'rhythm'—carry predictive information that ordinary position embeddings throw away. INTERPOS encodes those gaps as rhythm embeddings and morphs the positional embeddings of a transformer recommender before self-attention, using one of three fusion schemes: basic addition, an MLP projection, or a learned gate. The authors report that across seven mobile app datasets, all three variants beat state-of-the-art sequential and time-aware recommenders by a large margin on NDCG@K and HIT@K, with the largest relative gains on the sparser category-level datasets. If true, this matters because mobile app interaction streams are slower and more irregular than movie or product sequences, and injecting rhythm at the embedding layer may be a cheap way to make autoregressive recommenders sensitive to elapsed time.","feed_headline":"Time-gap 'rhythm' embeddings beat app recommenders by up to 158%","feed_subtitle":"Fusing inter-interaction time gaps into position embeddings improves next-app prediction across all seven datasets.","key_machinery":"The central object is the rhythm embedding layer $f_\\Omega$, which maps each inter-interaction interval $\\Delta t_i$ to a vector $R_i$, together with a fusion function that combines the position embedding $\\Theta$ with the rhythm embedding $\\Omega$. The paper names three fusion strategies: basic fusion ($M = \\Theta + \\Omega$), MLP fusion ($M = h([\\Theta,\\Omega])$ with $h: \\mathbb{R}^{2d} \\to \\mathbb{R}^d$), and gated fusion ($M = W \\odot \\tanh(h_p(\\Theta)) + (1-W) \\odot \\tanh(h_r(\\Omega))$, where $W = \\sigma(h_c([\\Theta,\\Omega]))$). The morphed position embeddings $M$ replace $\\Theta$ in the input to self-attention, so the mechanism's work is to expose the self-attention layers to elapsed-time information one step earlier than interval-aware attention baselines do.","core_discovery":"On the paper's own terms, the discovery is that a user's interaction rhythm—the sequence of inter-interaction intervals $\\Delta t_i = t_i - t_{i-1}$ over an activity window—is a learnable signal that can be fused with absolute position embeddings to make a transformer-based sequential recommender predict the next app more accurately. Rather than making attention itself time-aware, as interval-aware baselines do, INTERPOS injects the rhythm at the input embedding layer, producing morphed position embeddings $M$ that replace the standard position term in $E = \\Psi + \\Theta$. The paper claims this early fusion captures the long, irregular gaps typical of mobile app usage and consistently outperforms all baselines on seven datasets, with relative improvements in NDCG@10 as large as 157–158% over the best baseline on the Action category.","pith_inferences":["Beyond the paper's experiments, the same embedding-stage fusion should transfer to other domains with irregular inter-event times, such as news reading, health tracking, or financial transactions, where order-only positional encodings are equally blind to elapsed time.","Beyond the paper's reported datasets, the gap between the huge relative gains on short category sequences and the modest gains on full MobileRec suggests the benefit scales with sequence sparsity, which could be tested by varying the maximum sequence length and the gap distribution.","Beyond the paper's choices, a continuous rhythm encoding could replace the clipped lookup table (category gaps normalized by 0.2, MobileRec gaps clipped at 800) to test whether the model exploits fine-grained interval differences or just a few coarse regimes."],"forward_implications":["On the six sparse MobileRec category datasets, the LightSANs-based INTERPOS variants produce the largest relative gains, with NDCG@10 improvements of 157–158% over the best baseline on Action.","On the full 48-category MobileRec dataset, gains are smaller but still consistent: INTERPOS-GF improves NDCG@10 by 8.04% and HIT@10 by 12.20% over SASRec.","Because INTERPOS fuses rhythm before self-attention, the same morphed-embedding mechanism can be added to other autoregressive transformer recommenders without changing their attention layers.","Across datasets there is no single winning fusion strategy, so the choice among basic, MLP, and gated fusion should depend on the data distribution and the evaluation metric."],"supporting_citations":[{"why":"Time-interval-aware self-attention baseline that INTERPOS is designed to outperform on every dataset.","marker":"[8]"},{"why":"Frequency-and-time-aware baseline that is the strongest competitor on several category datasets.","marker":"[9]"},{"why":"One of the two transformer architectures that INTERPOS augments with rhythm-morphed position embeddings.","marker":"[10]"},{"why":"The other base architecture, used for both the SASRec baseline and the INTERPOS-SASRec variants.","marker":"[11]"},{"why":"Source of all seven mobile app recommendation datasets used for training and evaluation.","marker":"[29]"}],"fun_headline_variants":["Time-gap embeddings lift app recommender accuracy by 158%","User rhythm in position embeddings boosts next-app prediction","INTERPOS morphs positions with interaction timing for better apps","Mobile app recommenders get 158% NDCG boost from timing embeddings","Learn app usage rhythm to improve sequential recommendations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes the measured gains come from the time-gap information itself, rather than from the extra rhythm embedding table and fusion layers, because no equal-capacity control without rhythm information is evaluated.","fun_headline_variants_meta":{"raw":{"variants":["Time-gap embeddings lift app recommender accuracy by 158%","User rhythm in position embeddings boosts next-app prediction","INTERPOS morphs positions with interaction timing for better apps","Mobile app recommenders get 158% NDCG boost from timing embeddings","Learn app usage rhythm to improve sequential recommendations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1450,"prompt_tokens":1015,"completion_tokens":435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":353}},"tokens_in":631,"tokens_out":435,"duration_ms":5367,"temperature":1.0,"reasoning_tokens":353,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:43:55.337064+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain any INTERPOS variant on the same seven splits after replacing each user's inter-interaction gaps with gaps drawn from other users' histories, keeping the rhythm embedding table and fusion-layer sizes identical; if NDCG@20 stays within noise of the reported values, the rhythm content itself is not what drives the improvement.","supporting_citations":[{"cited_title":"Frequency enhanced hybrid attention network for sequential recommendation,","cited_arxiv_id":null,"evidence_quote":"Frequency-and-time-aware baseline that is the strongest competitor on several category datasets."},{"cited_title":"Lighter and better: low-rank decomposed self-attention networks for next-item recommendation,","cited_arxiv_id":null,"evidence_quote":"One of the two transformer architectures that INTERPOS augments with rhythm-morphed position embeddings."}],"review_version":1}