{"id":"312892a5-8eb4-46db-9cc3-9bc3cd1508c3","arxiv_id":"1908.04811","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Under Poisson content generation and exponential revisit times, the optimal social media access rate is μ* = λ((K/c)^(1/(K+1)) - 1), with λ the content rate, K the timeline size, and c the cost per access.","lead":"The paper derives a closed-form formula for how many new posts a social media user sees per visit, then solves for the visit rate that maximizes utility minus access cost. It validates the model on Facebook data from Brazil's 2018 election and finds that Facebook's News Feed shows fewer new posts than a chronological feed would.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The FIFO-timeline assumption A2 is the load-bearing limit: the paper's own Facebook measurements show the News Feed repeats posts, so the utility (14) overstates real VoA and the optimal rate (17) likely does not apply to ranked feeds.","rationale":"The paper's core derivation (Eqs. (12),(17)) is internally sound: integration of the truncated Poisson over an exponential interval yields Eq. (12), and differentiating the concave utility gives Eq. (17). The concern is not mathematical but scoping: the central claim 'optimal access rate' is only as good as the utility model, and the utility model's Assumption A2 (max new posts first) is contradicted by the paper's own empirical evidence on Facebook News Feed. The reader's weakest_assumption identifies this, and I agree. If a platform uses ranking that repeats posts, the true V̄ is smaller than Eq. (12) for each µ; then the marginal increase of µV̄ with µ is smaller, so the optimal µ* is lower than Eq. (17). Thus the answer to 'how often should I access' would be biased upward for ranked feeds. This is not an ad hominem or consensus disagreement; it is an internal mismatch between the model's assumptions and the paper's own measurements. The concrete test would quantify the bias. A secondary issue: the trace-driven validation (Section III-A) may compare the exponential-access formula (12) to deterministic-interval snapshots; if so, that validation is confounded and should be re-done. This does not change the formal result but further weakens the empirical support. Overall, the reader's CONDITIONAL verdict is appropriate; no change needed.","tokens_in":14885,"tokens_out":10827,"duration_ms":98084,"concrete_test":"From the FTE snapshots (high-sampling bot, taken every 10 min), subsample to emulated access intervals 1/µ = 1, 2, 4, 8, 12, 24 hours and compute the empirical average number of new impressions V_emp(µ) for K=10. Replace V̄ in Eq. (14) with V_emp(µ), maximize U(µ)=µ V_emp(µ) − cµ for c=1, and compare the argmax to Eq. (17) evaluated with the same λ and K. If the argmax is lower by more than 20%, the News Feed's violation of A2 materially changes the recommended access rate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For Eq. (17) to answer the title question, the utility U(µ)=µV̄−cµ must describe real feeds. This hinges on Assumption A2 (Section II-B): the timeline is filled with new posts before any repeats. The paper's own FTE measurements (Section III-C, Figure 2) show Facebook's News Feed violates this: it places repeated posts at the top, so the measured VoA is well below the model's prediction (e.g., K=10 yields about 5 new impressions instead of 10). Because a lower VoA at each µ flattens µV̄, the marginal benefit of extra accesses is overestimated by Eq. (16), so Eq. (17) produces a µ* that is too high for a ranked feed. The paper acknowledges this gap (Section VII: 'do not consider all the impact ... News Feed can have on VoA') but does not adjust the optimal-rate formula. Thus the formal derivation is correct, but the practical recommendation is conditional on the FIFO assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an analytical model for the value of an access (VoA) to an online social network, defined as the expected number of new (previously unseen) impressions in a timeline. Under assumptions of Poisson content generation at rate λ, exponential inter-access times with mean 1/µ, timeline size K, and a timeline that displays new content before repeats (Assumptions A1–A3), the authors derive a closed-form expression for VoA (Eq. (12)) and then solve a utility-maximization problem U(µ)=µV̄−cµ to obtain an optimal access rate µ* (Eq. (17)). The model is parameterized with Facebook data and compared against trace-driven FIFO simulations and against measurements from Facebook bot accounts that use the actual News Feed. The paper reports that the model matches FIFO timelines well but overestimates VoA under Facebook News Feed, which exhibits position bias by repeating posts. The authors conclude that News Feed reduces VoA and that the model provides insights into optimal access rates.","tokens_in":15094,"tokens_out":5751,"duration_ms":58676,"significance":"If the modeling assumptions hold, the paper makes a useful analytical contribution: it provides simple closed-form expressions for the expected number of novel impressions per access and for the optimal access rate, with transparent dependence on content generation rate, timeline size, and access cost. The VoA metric is intuitive and could serve as a building block for comparing personalization algorithms. The empirical observation that Facebook's News Feed repeats posts in top positions is also valuable and consistent with prior work on position bias. The derivation from assumptions to Eq. (12) and Eq. (17) appears correct, and the paper does not fit the VoA formula to data, so the central result is not circular. However, the title question concerns real social networks, and the practical answer is limited by the FIFO-timeline assumption that the paper's own measurements show is violated by News Feed; this is a substantial caveat on the applicability of the headline result.","major_comments":[{"comment":"The central practical claim, the optimal access rate in Eq. (17), is derived under Assumption A2 that the timeline displays as many new posts as possible before repeating old content. However, the paper's own Facebook measurements (Section III-C, Figure 2) show that the News Feed violates A2: repeated posts appear in top positions, and for K=10 the measured VoA is about 5 instead of the model's 10. Because the measured VoA is below the model's prediction, the marginal benefit of an extra access (Eq. (16)) is overestimated for ranked feeds, and hence the optimal µ* from Eq. (17) is likely too high for actual Facebook timelines. The acknowledgment in Section VII that the model does not consider all impacts of News Feed does not fix this, because the title question is about real online social networks. I recommend either restricting the domain of the optimal-rate result to FIFO or chronological timelines, or extending the model to account for position bias, e.g., by adding a parameter that captures the fraction of repeated impressions in top positions.","section":"Section II-B (Assumption A2) and Section III-C"},{"comment":"The optimization analysis omits the feasibility condition for the stationary point. The derivative at µ=0 is ∂U/∂µ|₀ = K − c, so the interior maximum given by Eq. (17) exists only when c < K. For c ≥ K, the utility is decreasing for all µ ≥ 0 and the optimal solution is the boundary µ* = 0. As written, Eq. (17) can produce negative or infeasible access rates when c ≥ K, and the numerical examples in Section V use c=1 with K ≥ 2 but do not state the condition. The authors should present the piecewise solution and the condition explicitly; otherwise the optimization claim is incomplete.","section":"Section IV, Eq. (17)"},{"comment":"The empirical validation is weakened by in-sample parameter estimation and by the lack of uncertainty quantification. The value λ=4.487 is estimated from the same Facebook trace against which the model is compared in Figure 1(b), so the close agreement is not a predictive test. No confidence intervals or error bars are reported for the measured VoA values in Figures 1(b) and 2. In addition, the FIFO simulation procedure described in Section III-A is not a faithful implementation of the model's inter-access process: snapshots are built by selecting a reference post uniformly at random from all posts and taking its K−1 immediate predecessors, which forces access times to coincide with post-creation times rather than following the assumed exponential or deterministic interval distribution. This can bias the simulated VoA and the apparent agreement with Eq. (12). The authors should either describe a direct event-driven simulation of the assumed access process or justify why the uniform-reference-post scheme is equivalent.","section":"Section III-A and III-C"}],"minor_comments":[{"comment":"The sentence 'This finding reveals that Facebook Facebook creates a position bias' contains a duplicated word 'Facebook'; it should read 'Facebook creates a position bias'.","section":"Section III-C, paragraph on position bias"},{"comment":"The definition of V would be clearer as V = min(A, K) under the assumption that the timeline is filled with new posts only up to its capacity.","section":"Section II-C, Eq. (2)"},{"comment":"The text writes 'V (ρ, K) = ρ' and 'V (K, ρ) = K' using V for the expectation; for consistency with the notation introduced in Section II, these should be V̄, the expected value of access.","section":"Section V-A, Figure 4 discussion"},{"comment":"The abstract claims the results indicate 'the predictive power of the approach,' but the validation is in-sample and the News Feed measurements show the model overestimates VoA; a more cautious phrasing would better match the evidence.","section":"Abstract and Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The paper's analytical derivation is sound and the identified issues are fixable, but the mismatch between the title question and the FIFO-only validity of the optimal-rate formula is a substantive scope concern. The authors should either clearly reframe the contribution as a model for chronological/FIFO timelines or extend the model to account for position bias. The optimization condition c < K must also be stated. The empirical claims should be scaled back: the current abstract overstates predictive power given the in-sample estimation and lack of error bars."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: Eq. (12) and Eq. (17) are correct, useful, and not something I have seen in the prior work. Under Poisson content generation, exponential inter-access times, and a K-post timeline filled with new posts before repeats, the expected number of new impressions is (lambda/mu)(1 - (lambda/(lambda+mu))^K), and maximizing utility mu*Vbar - c*mu gives mu* = lambda((K/c)^(1/(K+1)) - 1). That is a nice back-of-envelope for screen-time tools, and the derivation is clean.\n\nThe paper earns credit for honesty in the empirical section. The trace-driven FIFO simulation matches the model, and the Facebook News Feed measurements show the model overpredicts VoA. The paper attributes this to News Feed position bias and reports that at K=10 Facebook shows about 5 new posts instead of the 10 the model expects. That is a real, useful empirical finding. The later note that real dropped posts can reappear in subsequent accesses in the 15<K<35 region is a fair self-critique.\n\nThe soft spots are real but not fatal. The biggest is that Assumption A2 is exactly what fails on Facebook's ranked feed, so the optimal rate in Eq. (17) is not yet an answer to the title question for Facebook as it runs. The paper acknowledges this gap in Section VII, but the abstract and Section IV present the result more confidently. Because News Feed lowers the marginal value of frequent accesses, the formula likely recommends checking too often on ranked feeds. That is a load-bearing caveat, but it is openly documented.\n\nSmaller issues: lambda is estimated from the same trace used for validation, the simulation plots have no error bars, and no data or code are released. These are fixable and standard reporting gaps. The citation pattern is fine: the paper builds on the authors' earlier FIFO/TTL timeline work and says so.\n\nVerdict: this deserves peer review. It is a compact, correctly derived idealization with an honest empirical comparison. I would want revision to add artifacts, uncertainty quantification, and a prominent statement that Eq. (17) applies to FIFO/no-repeat timelines, not to ranked feeds as they currently operate.","headline":"A correct and simple closed form for optimal access frequency under FIFO timelines, but the practical answer is conditional because the paper's own Facebook data show real feeds violate the no-repeat assumption.","tokens_in":15624,"tokens_out":2588,"would_cite":false,"duration_ms":27594,"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":"A user's best check-in rate for a social feed is $\\mu^* = \\lambda\\left((K/c)^{1/(K+1)} - 1\\right)$, balancing new posts against access cost; the formula holds for chronological feeds while ranked feeds deliver measurably less value.","keywords":["value of access","optimal access rate","social network timelines","Poisson process","News Feed","user utility","age of information","sampling"],"falsifier":"Run a controlled experiment with a reverse-chronological feed in which posts arrive as a Poisson process with known rate $\\lambda$, the user's inter-access times are exponentially distributed with known mean $1/\\mu$, and each view shows at most $K$ posts; if the measured mean number of new posts per access disagrees with $(\\lambda/\\mu)(1-(\\lambda/(\\lambda+\\mu))^K)$ beyond sampling error, the central claim fails.","tokens_in":14645,"feed_emoji":"📱","tokens_out":7283,"duration_ms":67986,"temperature":0.7,"pith_summary":"Users who want to stay informed face a trade-off: check a social network too often and most accesses return no new posts; check too rarely and posts pile up beyond what one can read. This paper claims that this trade-off has a unique optimal access rate, and that the optimum is given by a simple closed-form expression when content arrives as a Poisson process, inter-access times are exponential, and each timeline has fixed size. The core result is that the utility-maximizing rate is $\\mu^* = \\lambda\\left((K/c)^{1/(K+1)} - 1\\right)$, where $\\lambda$ is the posting rate, $K$ the number of posts visible per access, and $c$ the cost of an access. The authors validate the underlying value-of-access formula against measurements from a large social platform, and find that a chronological feed matches the model while the platform's ranked feed delivers fewer new posts than predicted. If correct, the paper gives users and platforms a quantitative target for how often feeds should be refreshed.","feed_headline":"One formula gives the optimal social-feed check-in rate","feed_subtitle":"New posts versus access cost yields a unique best frequency; ranked feeds fall short of it.","key_machinery":"The engine of the argument is the value-of-access (VoA) metric, defined as the expected number of new impressions per timeline view. The model treats the number of posts created between accesses as a Poisson random variable truncated at $K$, and then integrates over the exponential inter-access distribution to obtain Eq. (12). This closed form is what turns the optimization into a one-variable calculus problem: differentiating $U(\\mu)$ gives $\\partial U/\\partial\\mu = K(\\lambda/(\\lambda+\\mu))^{K+1} - c$, and setting it to zero produces the optimal rate $\\mu^*$ in Eq. (17). The same expression also yields the limiting behaviors the paper discusses: as $K\\to\\infty$, $\\bar{V}\\to\\lambda/\\mu$, and as $\\mu\\to 0$, $\\bar{V}\\to K$.","core_discovery":"The paper's central discovery is a closed-form expression for the expected number of new posts a user sees per access — the 'value of an access' — and the optimal access rate that follows from it. Under Poisson content generation with rate $\\lambda$, exponentially distributed times between accesses with mean $1/\\mu$, and a timeline that shows at most $K$ posts, the value of an access is $\\bar{V} = (\\lambda/\\mu)(1 - (\\lambda/(\\lambda+\\mu))^K)$. Assuming a per-access cost $c$, the utility per unit time is $U(\\mu) = \\mu\\bar{V} - c\\mu$, and the maximizing rate is $\\mu^* = \\lambda\\left((K/c)^{1/(K+1)} - 1\\right)$. The paper's measurements, taken by observing what a set of automated accounts actually saw on a major platform, show that a reverse-chronological feed reproduces the model's values, while the platform's ranked feed shows repeated posts in top positions and therefore yields a lower value of access than the model predicts for the same parameters.","pith_inferences":["An implicit corollary of the optimization is that a platform can manipulate a user's access rate by adjusting how often it repeats posts: lowering the observed value per access raises the $\\mu$ that maximizes utility, which may be exactly why ranked feeds bury novelties.","Because the formula separates $\\lambda$ and $K$ from the user's internal cost $c$, observed access behavior could in principle be inverted to estimate $c$ for a given user, turning the model into a measurement tool for attention-budget studies.","The derivation assumes exponential inter-access times; extending it to general renewal processes would replace the factor $(\\lambda/(\\lambda+\\mu))^K$ with a Laplace-transform term, a testable generalization the paper does not pursue."],"forward_implications":["A user who can estimate the posting rate $\\lambda$, the number of posts $K$ they can actually consume per visit, and the cost $c$ they attach to an access has a concrete number to aim for: $\\mu^* = \\lambda\\left((K/c)^{1/(K+1)} - 1\\right)$.","Because $\\mu^*$ grows linearly with $\\lambda$, busier content streams justify more frequent checking, but the marginal gain of each extra check shrinks as $K$ grows; in the saturated regime, larger timelines lower the optimal check-in rate.","The model predicts that any feed that repeats already-seen posts instead of showing all available novelties will deliver a lower value per access than the user's optimal plan assumes, so the user either scrolls further or checks more often to reach the same utility.","The two operating regimes the paper identifies — where the value of access is limited by the posting rate $\\rho = \\lambda/\\mu$ versus by the timeline size $K$ — give platform designers a criterion for when recommender filters help and when they just add overhead."],"supporting_citations":[{"why":"Supplies the browser extension used to record the impressions the virtual users actually saw.","marker":"[1]"},{"why":"Documents the position bias that the paper invokes to explain why the ranked feed underperforms the model.","marker":"[5]"},{"why":"Provides the exponential-race probability interpretation that makes the closed form intuitive.","marker":"[13]"},{"why":"Supplies the timeline performance analysis the paper builds on for access-rate modeling.","marker":"[16]"},{"why":"Introduces the FIFO/TTL timeline model and the position-bias finding that the validation reproduces.","marker":"[17]"},{"why":"Models the platform's News Feed as a FIFO queue, the comparison baseline for the experiments.","marker":"[18]"}],"fun_headline_variants":["Formula finds your optimal feed-check rate","The math of when to check social feeds","Best social check-in frequency, solved","When to refresh your feed: the formula","Optimal social access rate, in one equation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result assumes a timeline that shows new posts first and repeats old posts only when there are no novelties; a ranked feed that puts repeated posts at the top breaks that assumption, and then the predicted value of access and the optimal rate no longer describe what the user gets.","fun_headline_variants_meta":{"raw":{"variants":["Formula finds your optimal feed-check rate","The math of when to check social feeds","Best social check-in frequency, solved","When to refresh your feed: the formula","Optimal social access rate, in one equation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1333,"prompt_tokens":936,"completion_tokens":397,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":332}},"tokens_in":552,"tokens_out":397,"duration_ms":4899,"temperature":1.0,"reasoning_tokens":332,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:32:37.247236+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled experiment with a reverse-chronological feed in which posts arrive as a Poisson process with known rate $\\lambda$, the user's inter-access times are exponentially distributed with known mean $1/\\mu$, and each view shows at most $K$ posts; if the measured mean number of new posts per access disagrees with $(\\lambda/\\mu)(1-(\\lambda/(\\lambda+\\mu))^K)$ beyond sampling error, the central claim fails.","supporting_citations":[{"cited_title":"Facebook tracking exposed, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the browser extension used to record the impressions the virtual users actually saw."},{"cited_title":"Bias on the web","cited_arxiv_id":null,"evidence_quote":"Documents the position bias that the paper invokes to explain why the ranked feed underperforms the model."},{"cited_title":"Essentials of Stochastic Processes , volume 7 of Springer Texts in Statistics","cited_arxiv_id":null,"evidence_quote":"Provides the exponential-race probability interpretation that makes the closed form intuitive."},{"cited_title":"Perfor- mance analysis of online social platforms","cited_arxiv_id":null,"evidence_quote":"Supplies the timeline performance analysis the paper builds on for access-rate modeling."},{"cited_title":"Fair ness in online social network timelines: Measurements, models and mechanism design","cited_arxiv_id":null,"evidence_quote":"Introduces the FIFO/TTL timeline model and the position-bias finding that the validation reproduces."},{"cited_title":"Bias es in the facebook news feed: a case study on the italian elections","cited_arxiv_id":null,"evidence_quote":"Models the platform's News Feed as a FIFO queue, the comparison baseline for the experiments."}],"review_version":1}