arxiv: 2604.27530 · v1 · submitted 2026-04-30 · 💻 cs.SI · cs.CY

Recognition: unknown

Temporal and Content Coupling Analysis of Social Media User Behavior

Jipeng Tan , Mengye Yang , Zhanghao Li , Yong Min

Authors on Pith no claims yet

Pith reviewed 2026-05-07 07:55 UTC · model grok-4.3

classification 💻 cs.SI cs.CY

keywords news consumptiontemporal dynamicscontent couplinguser behaviorcircadian rhythmspower-law distributionshistorical interestspreference groups

0 comments

The pith

News clicks are driven mainly by historical interests during active hours, with patterns varying by user preference type.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that news consumption arises from the coupling between when users are active and what content they select. It introduces a multi-scale framework that uncovers daily rhythms at the largest scale, power-law gaps between sessions at the middle scale, and exponential patterns within sessions at the smallest scale. Clicks depend heavily on past reading history, but this dependence drops when the content offered is more diverse. The coupling is strongest in peak activity windows, and users split into groups—timeliness and entertainment seekers click often and stick to history, while diversified users click less and respond more to variety. A reader would care because these regularities could improve how platforms time and personalize recommendations without building ever-more-complex profiles.

Core claim

The authors claim that a multi-scale temporal-content framework applied to the MIND and Adressa news datasets reveals clear hierarchical temporal patterns—circadian rhythms via Fourier analysis at macroscale, power-law session intervals with exponent near 1 at mesoscale, and exponential distributions for action counts and intervals at microscale—while showing that historical interests dominate active time periods in driving clicks and that preference groups differ in frequency and sensitivity to diversity.

What carries the argument

The multi-scale temporal-content framework, which decomposes user activity into macroscale circadian rhythms, mesoscale power-law session intervals, and microscale exponential action patterns, then couples these timescales with content analysis to quantify how historical interests interact with time-based activity and diversity.

If this is right

Clicks are driven primarily by historical interests, with this effect weakening as content diversity increases.
Timeliness and entertainment-oriented users click more frequently and rely more on historical interests.
Diversified users click less frequently and are more sensitive to content diversity.
Temporal patterns are hierarchical: circadian at macroscale, power-law at mesoscale, and exponential at microscale.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

News platforms could improve engagement by prioritizing historical content during users' detected active hours.
Segmenting users into preference groups could allow simpler, targeted strategies such as introducing more diversity only to certain segments.
The same temporal-content coupling may appear in other online activities like video consumption or social posting, suggesting broader modeling opportunities.
Testing the framework on additional datasets would clarify whether the patterns are domain-specific to news or apply across social media.

Load-bearing premise

The observed rhythms, distributions, and dominance of historical interests in the MIND and Adressa datasets are assumed to reflect general social media user behavior rather than being specific to these news platforms or their data collection.

What would settle it

Repeating the same multi-scale analysis on a large dataset from a non-news platform, such as a video-sharing or general social network site, and finding no circadian rhythms, no power-law session intervals, or no dominance of historical interests during active periods would falsify the claim that these couplings are general.

Figures

Figures reproduced from arXiv: 2604.27530 by Jipeng Tan, Mengye Yang, Yong Min, Zhanghao Li.

**Figure 1.** Figure 1: : Schematic diagram of the dynamic time window division for user news view at source ↗

**Figure 2.** Figure 2: : The schematic diagram of the content analysis process for user news con view at source ↗

**Figure 3.** Figure 3: : (a) Analysis and fitting comparison of the temporal pattern of news con view at source ↗

**Figure 4.** Figure 4: : The probability density function distribution of time intervals between ses view at source ↗

**Figure 5.** Figure 5: : ABM simulation of user activity compared to real user activity over a 24- view at source ↗

**Figure 6.** Figure 6: : The difference between the joint probability density distributions of view at source ↗

**Figure 7.** Figure 7: : (a) Probability density distributions of view at source ↗

**Figure 8.** Figure 8: : Randomly selected distributions of (a) view at source ↗

**Figure 9.** Figure 9: : (a) Deviation distribution of daily activity times for users with different view at source ↗

read the original abstract

News consumption behavior is shaped by the coupling between temporal dynamics and content selection. This study proposes a multi-scale temporal-content framework and validates it on two large real-world news datasets, MIND and Adressa. Results reveal hierarchical temporal patterns. At the macroscale, Fourier modeling identifies clear circadian rhythms; at the mesoscale, session intervals follow a power-law distribution with $\alpha \approx 1$; and at the microscale, within-session action counts and inter-action intervals follow exponential distributions with $\lambda \approx 0.3$ and $\lambda \approx 0.02$, respectively. Content analysis shows that clicks are mainly driven by historical interests, while this dependence weakens as content diversity increases. Temporal-content coupling further indicates that users' historical interests dominate active time periods in shaping behavior. Preference groups also differ: timeliness and entertainment-oriented users click more frequently and rely more on historical interests, whereas diversified users click less and are more sensitive to content diversity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives concrete observational fits from two news datasets on multi-scale timing and content coupling, but the alpha≈1 power-law for sessions lacks the usual cutoff and goodness-of-fit details.

read the letter

The main takeaway is that this paper runs a multi-scale breakdown on the MIND and Adressa news logs and reports that clicks are driven more by users' historical interests than by current content diversity, with that dependence strongest during active periods and varying across preference groups like timeliness-oriented versus diversified users. They also extract circadian rhythms at the daily scale, power-law session intervals with alpha near 1, and exponential distributions for within-session actions and gaps with lambdas around 0.3 and 0.02. The combination of temporal hierarchy with content preference splits is a reasonable extension of existing user modeling work and produces some usable numbers for recommendation contexts. The datasets are public, which is a plus for reproducibility. The soft spot is the mesoscale power-law claim. Alpha=1 sits at the boundary where the integral diverges without an explicit upper cutoff, and the abstract gives no xmin, no MLE details, no KS p-values, and no comparison to log-normal or Weibull alternatives. If the full paper does not supply those checks or rule out logging artifacts such as session timeouts, the claimed bridge from macro circadian to micro exponential patterns rests on thinner evidence than it should. The historical-interest dominance result is also tied to whatever definitions they used for interests and diversity, so it may not travel beyond these two platforms. This is the sort of paper that belongs in a computational social science or recommender systems reading group. It supplies empirical anchors rather than a new theory, but the questions are clear and the data accessible. I would send it to peer review so referees can examine the fitting procedures and ask for extra robustness plots.

Referee Report

2 major / 3 minor

Summary. The paper proposes a multi-scale temporal-content framework for news consumption on social media and validates it empirically on the MIND and Adressa datasets. It reports hierarchical temporal patterns—circadian rhythms at macroscale via Fourier analysis, power-law session intervals (α ≈ 1) at mesoscale, and exponential distributions for within-session action counts (λ ≈ 0.3) and inter-action intervals (λ ≈ 0.02) at microscale—along with content analysis showing that clicks are primarily driven by historical interests, with this dependence weakening as content diversity increases. The central claims concern temporal-content coupling (historical interests dominate active periods) and differences across preference groups (timeliness/entertainment users click more and rely more on history; diversified users click less and are more sensitive to diversity).

Significance. If the distributional fits and coupling claims hold after rigorous validation, the work offers concrete empirical grounding for multi-scale user behavior models in social media, with potential value for recommendation algorithms that incorporate both temporal rhythms and content history. Credit is due for grounding the analysis in two named public datasets and reporting specific fitted parameter values rather than purely qualitative observations.

major comments (2)

[Mesoscale temporal patterns] Mesoscale temporal patterns section: the claim that session intervals follow p(Δt) ∝ (Δt)^{-α} with α ≈ 1 is load-bearing for the hierarchical framework that connects macroscale circadian rhythms to microscale exponentials and underpins the subsequent temporal-content coupling conclusions. No xmin, xmax, normalization constant, KS-test p-value, or comparison to alternatives (log-normal, Weibull, truncated power-law) is reported, nor is any reference to standard MLE+KS procedures (Clauset et al.). For α = 1 the integral diverges logarithmically without an explicit upper cutoff; if the tail is an artifact of binning, finite observation windows, or dataset logging rules (e.g., session timeouts), the multi-scale bridge collapses.
[Temporal-content coupling and preference groups] Temporal-content coupling and preference-group analysis sections: the assertion that 'historical interests dominate active time periods' and the reported group differences (timeliness/entertainment vs. diversified users) rest on unspecified quantitative measures. No details are given on the exact statistical models (e.g., regression coefficients, partial correlations, or controls for user activity level, time-of-day, or content popularity), error bars, or robustness checks, making it impossible to assess whether the dominance and group contrasts are load-bearing or sensitive to confounding.

minor comments (3)

[Abstract] Abstract and methods: the reported λ values (0.3 and 0.02) and α ≈ 1 should be accompanied by confidence intervals or standard errors; the datasets' temporal coverage and user-sample sizes are not stated, limiting assessment of generalizability.
[Figures and notation] Notation and figures: the power-law and exponential forms should explicitly state their support and any truncation; figure captions for the fitted distributions should include goodness-of-fit diagnostics or overlaid alternative models.
[References] References: the power-law fitting discussion would benefit from citing Clauset et al. (2009) on statistical methods for power-law identification.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback, which has helped us strengthen the statistical foundations and transparency of our multi-scale analysis. We address each major comment point by point below. Where the original submission lacked sufficient methodological detail, we have revised the manuscript to incorporate the requested information, including fitting procedures, model specifications, controls, and robustness checks. These changes preserve the core empirical findings while improving rigor.

read point-by-point responses

Referee: [Mesoscale temporal patterns] Mesoscale temporal patterns section: the claim that session intervals follow p(Δt) ∝ (Δt)^{-α} with α ≈ 1 is load-bearing for the hierarchical framework that connects macroscale circadian rhythms to microscale exponentials and underpins the subsequent temporal-content coupling conclusions. No xmin, xmax, normalization constant, KS-test p-value, or comparison to alternatives (log-normal, Weibull, truncated power-law) is reported, nor is any reference to standard MLE+KS procedures (Clauset et al.). For α = 1 the integral diverges logarithmically without an explicit upper cutoff; if the tail is an artifact of binning, finite observation windows, or dataset logging rules (e.g., session timeouts), the multi-scale bridge collapses.

Authors: We agree that the mesoscale power-law analysis requires explicit reporting of the fitting procedure to substantiate the hierarchical framework. In the revised manuscript we now detail the maximum-likelihood estimation following Clauset et al. (2009), including the estimated xmin (the lower bound of the power-law regime, selected to minimize the KS distance), xmax (set to the 99th percentile of observed intervals, approximately one day, to respect the macroscale circadian boundary), the normalization constant, and the KS goodness-of-fit p-value. We also report likelihood-ratio tests against log-normal and Weibull alternatives, confirming the power-law provides a statistically superior fit within the supported range. Regarding α ≈ 1 and potential divergence, we explicitly state that the distribution is truncated at xmax; this truncation is empirically justified by the daily periodicity already identified at the macroscale and by the finite observation windows of both datasets. We have added a citation to Clauset et al. and a supplementary figure showing the complementary cumulative distribution on log-log axes with the fitted line and bootstrap confidence bands. These additions confirm that the tail is not an artifact of binning or session-timeout rules, thereby preserving the multi-scale bridge. revision: yes
Referee: [Temporal-content coupling and preference groups] Temporal-content coupling and preference-group analysis sections: the assertion that 'historical interests dominate active time periods' and the reported group differences (timeliness/entertainment vs. diversified users) rest on unspecified quantitative measures. No details are given on the exact statistical models (e.g., regression coefficients, partial correlations, or controls for user activity level, time-of-day, or content popularity), error bars, or robustness checks, making it impossible to assess whether the dominance and group contrasts are load-bearing or sensitive to confounding.

Authors: We acknowledge that the original submission did not provide sufficient detail on the quantitative measures and controls used for the temporal-content coupling and preference-group analyses. In the revised manuscript we now specify the exact statistical models: mixed-effects logistic regression predicting click occurrence, with historical-interest similarity (cosine similarity of user embedding vectors from prior sessions) as the focal predictor, user-level random intercepts, and fixed-effect controls for time-of-day (hour dummies), content popularity (log view count), and overall user activity. We report the coefficient on historical interest, its standard error, and the interaction term with content diversity, together with bootstrap-derived confidence intervals. For the preference groups we describe the k-means clustering on user-level features (click rate, diversity index, timeliness preference) and include ANOVA or equivalent tests for between-group differences in click frequency and history reliance. Robustness checks—alternative clustering algorithms, activity-level quartiles, and exclusion of extreme users—are now reported in the main text and Appendix. These additions allow readers to evaluate the load-bearing status of the coupling claims and group contrasts. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical multi-scale analysis grounded in external datasets

full rationale

The paper performs direct statistical analysis on two independent real-world datasets (MIND and Adressa). It reports observed distributions (Fourier-identified circadian rhythms at macroscale, power-law session intervals at mesoscale with α≈1, exponential within-session counts and intervals at microscale) and content correlations (historical interests driving clicks, modulated by diversity). These are fitted to data and interpreted as patterns; no equations define a quantity in terms of itself, no fitted parameters are relabeled as out-of-sample predictions, and no self-citations supply load-bearing uniqueness theorems or ansatzes. The temporal-content coupling claims follow from the empirical measurements rather than reducing to internal definitions or prior author results by construction. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central claims rest on fitting standard statistical models to data and assuming these capture underlying behavioral mechanisms; no new entities are postulated.

free parameters (3)

power-law exponent α = ≈1
Fitted to session interval data at mesoscale
exponential rate λ for action counts = ≈0.3
Fitted to within-session action counts at microscale
exponential rate λ for inter-action intervals = ≈0.02
Fitted to inter-action intervals at microscale

axioms (3)

domain assumption Session intervals follow a power-law distribution
Invoked for mesoscale temporal pattern identification
domain assumption Within-session action counts and inter-action intervals follow exponential distributions
Invoked for microscale temporal pattern identification
standard math Fourier modeling can identify circadian rhythms in user activity
Used for macroscale temporal analysis

pith-pipeline@v0.9.0 · 5465 in / 1636 out tokens · 57747 ms · 2026-05-07T07:55:25.673108+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

44 extracted references · 1 canonical work pages

[1]

C. E. Robertson, N. Proellochs, K. Schwarzenegger, P. Parnamets, J. J. Van Bavel, S. Feuerriegel, Negativity drives online news consumption,Nat. Hum. Behav.7 (2023) 812–822

2023
[2]

Brossard, New media landscapes and the science information consumer,Proc

D. Brossard, New media landscapes and the science information consumer,Proc. Natl. Acad. Sci. U.S.A.110 (2013) 14096–14101

2013
[3]

S. Raza, C. Ding, News recommender system: a review of recent progress, challenges, and opportunities,Artif. Intell. Rev.55 (2022) 749–800. 15

2022
[4]

J. R. Saura, A. Reyes-Menendez, N. de Matos, M. B. Correia, P. Palos-Sanchez, Consumer behavior in the digital age,J. Spat. Organ. Dyn.8 (2020) 190–196

2020
[5]

Ramesh, F

M. Ramesh, F. B. Flohr, Walk-the-talk: Llm driven pedestrian motion generation, in: 2024 35TH IEEE Intelligent Vehicles Symposium, IEEE IV 2024, 2024, pp. 3057–3062

2024
[6]

Jiang, X

H. Jiang, X. Zhang, X. Cao, C. Breazeal, D. Roy, J. Kabbara, Personallm: Investigating the ability of large language models to express personality traits,arXiv preprint arXiv (2024) 2305.02547

work page arXiv 2024
[7]

Z. Xi, W. Chen, X. Guo, W. He, Y. Ding, B. Hong, M. Zhang, J. Wang, S. Jin, E. Zhou, R. Zheng, X. Fan, X. Wang, L. Xiong, Y. Zhou, W. Wang, C. Jiang, Y. Zou, X. Liu, Z. Yin, S. Dou, R. Weng, W. Qin, Y. Zheng, X. Qiu, X. Huang, Q. Zhang, T. Gui, The rise and potential of large language model based agents: a survey,Sci. China-Inf. Sci.68 (2025) 121101

2025
[8]

J. L. Nelson, S. C. Lewis, The structures that shape news consumption: Evidence from the early period of the covid-19 pandemic,Journalism23 (2022) 2495–2512

2022
[9]

Diaz-Campo, F

J. Diaz-Campo, F. Segado-Boj, E. Fernandez-Gomez, User habits and social media type as predictors of news consumption and sharing,Profes. Inf.30 (2021) e300417

2021
[10]

Y. Chen, W. Ye, C. Lin, Y. Chen, Unbiased news recommendation model time and con- tent,Expert Syst. Appl.256 (2024) 124864

2024
[11]

Y. Ding, B. Wang, X. Cui, M. Xu, Popularity prediction with semantic retrieval for news recommendation,Expert Syst. Appl.247 (2024) 123308

2024
[12]

Z. Ou, Z. Han, P. Liu, S. Teng, M. Song, Siir: Symmetrical information interaction mod- eling for news recommendation,IEEE Trans. Neural Netw. Learn. Syst.35 (2024) 17111– 17122

2024
[13]

Q. Meng, H. Yan, B. Liu, X. Sun, M. Hu, J. Cao, Recognize news transition from collective behavior for news recommendation,ACM Trans. Inf. Syst.41 (2023) 93

2023
[14]

C. Lim, S. Kim, Examining factors influencing the user’s loyalty on algorithmic news recommendation service,Humanit. Soc. Sci. Commun.11 (2024) 10

2024
[15]

A. Iana, M. Alam, H. Paulheim, A survey on knowledge-aware news recommender systems, Semant. Web15 (2024) 21–82

2024
[16]

Zhang, X

S.S. Zhang, X. Liang, Y.D. Wei, X.P. Zhou, J.C. Zhao, On statistical social behaviors from millions of cell phone users,Physica A-Stat. Mech. Appl.512 (2018) 456–467

2018
[17]

Zhang, H.Y

S.T. Zhang, H.Y. Yuan, L.L. Duan, Analysis of human behavior statistics law based on wechat moment,Physica A-Stat. Mech. Appl.540 (2020) 122854

2020
[18]

Kwon, W.S

O. Kwon, W.S. Son, W.S. Jung, The double power law in human collaboration behavior: The case of wikipedia,Physica A-Stat. Mech. Appl.461 (2016) 85–91

2016
[19]

D. Roy, M. Dutta, A systematic review and research perspective on recommender systems, J. Big Data9 (2022) 59

2022
[20]

L. I. Labrecque, J. vor dem Esche, C. Mathwick, T. P. Novak, C. F. Hofacker, Consumer power: evolution in the digital age,J. Interact. Mark.27 (2013) 257–269

2013
[21]

Sagkaya Gungor, T

A. Sagkaya Gungor, T. Ozansoy cadirci, Understanding digital consumer: A review, syn- thesis, and future research agenda,Int. J. Consum. Stud.46 (2022) 1829–1858

2022
[22]

T. G. Kormelink, I. C. Meijer, A user perspective on time spent: temporal experiences of everyday news use,Journalism Stud.21 (2020) 271–286

2020
[23]

H.S. Sheu, Z. Chu, D. Qi, S. Li, Knowledge-guided article embedding refinement for session-based news recommendation,IEEE Trans. Knowl. Data Eng.33 (2022) 7921– 7927

2022
[24]

Y. Wu, K. Li, G. Zhao, X. Qian, Personalized long- and short-term preference learning for next poi recommendation,IEEE Trans. Knowl. Data Eng.34 (2022) 1944–1957

2022
[25]

B. Koo, H. Jeon, U. Kang, Accurate news recommendation coalescing personal and global temporal preferences, in:Advances in Knowledge Discovery and Data Mining, PAKDD 2020, PT I, Vol. 12084 of Lecture Notes in Artificial Intelligence, Singapore Tourism Board, 2020, pp. 78–90

2020
[26]

Viana, M

P. Viana, M. Soares, R. Gaio, A. Correia, Consumer attitudes toward news delivering: an experimental evaluation of the use and efficacy of personalized recommendations,Inform. 16 11 (2020) 350

2020
[27]

Sakagami, T

H. Sakagami, T. Kamba, Learning personal preferences on online newspaper articles from user behaviors,Comput. Netw. ISDN Syst.29 (1997) 1447–1455

1997
[28]

Mitova, S

E. Mitova, S. Blassnig, E. Strikovic, A. Urman, A. Hannak, C. H. de Vreese, F. Esser, News recommender systems: a programmatic research review,Ann. Int. Commun. Assoc. 47 (2023) 84–113

2023
[29]

Z.D. Zhao, Z. Yang, Z. Zhang, T. Zhou, Z.G. Huang, Y.C. Lai, Emergence of scaling in human-interest dynamics,Sci. Rep.3 (2013) 3472

2013
[30]

Fortunato, A

S. Fortunato, A. Flammini, F. Menczer, A. Vespignani, Topical interests and the mitiga- tion of search engine bias,Proc. Natl. Acad. Sci. U.S.A.103 (2006) 12684–12689

2006
[31]

C. Feng, M. Khan, A. U. Rahman, A. Ahmad, News recommendation systems - accom- plishments, challenges and future directions,IEEE Access8 (2020) 16702–16725

2020
[32]

X. Yu, M. Haroon, E. Menchen-Trevino, M. Wojcieszak, Nudging recommendation al- gorithms increases news consumption and diversity on youtube,PNAS Nexus3 (2024) pgae518

2024
[33]

Espeland, The dynamics of political interest and news media avoidance: A generational and longitudinal perspective,Journalism Stud.25 (2024) 1476–1497

E. Espeland, The dynamics of political interest and news media avoidance: A generational and longitudinal perspective,Journalism Stud.25 (2024) 1476–1497

2024
[34]

Karnowski, A

V. Karnowski, A. S. Kuempel, L. Leonhard, D. J. Leiner, From incidental news exposure to news engagement. how perceptions of the news post and news usage patterns influence engagement with news articles encountered on facebook,Comput. Hum. Behav.76 (2017) 42–50

2017
[35]

Scheffauer, T

R. Scheffauer, T. Durotoye, H. G. de Zuniga, Incidental news exposure and political consumerism - exploring nuances,Telemat. Inform.84 (2023) 102044

2023
[36]

S. Park, J. Y. Lee, Incidental news exposure on facebook and its relation to trust in news, Soc. Media Soc.9 (2023) 20563051231158823

2023
[37]

Kleppe, M

M. Kleppe, M. Otte, Analysing and understanding news consumption patterns by tracking online user behaviour with a multimodal research design,Digit. Scholarsh. Humanit.32 (2017) 158–170

2017
[38]

T. G. Kormelink, I. C. Meijer, What clicks actually mean: Exploring digital news user practices,Journalism19 (2018) 668–683

2018
[39]

Tenenboim, A

O. Tenenboim, A. A. Cohen, What prompts users to click and comment: A longitudinal study of online news,Journalism16 (2015) 198–217

2015
[40]

F. Wu, Y. Qiao, J.H. Chen, C. Wu, T. Qi, J. Lian, D. Liu, X. Xie, J. Gao, W. Wu, M. Zhou, Mind: A large-scale dataset for news recommendation, in:58th Annual Meeting of the Association for Computational Linguistics (ACL 2020), Assoc. Comput. Linguist., 2020, pp. 3597–3606

2020
[41]

J. A. Gulla, L. Zhang, P. Liu, O. Ozgobek, X. Su, The adressa dataset for news rec- ommendation, in:2017 IEEE/WIC/ACM International Conference on Web Intelligence (WI 2017), IEEE; WIC; ACM, 2017, pp. 1042–1048

2017
[42]

Y. Lei, Z. Wu, Time series classification based on statistical features,EURASIP J. Wirel. Commun. Netw.2020 (2020) 46

2020
[43]

D. He, A. G¨ oker, Detecting session boundaries from web user logs, in:Proceedings of the 22nd Annual Colloquium on Information Retrieval Research (IRSG 2000), Cambridge, UK, 2000, pp. 57–66

2000
[44]

T 10-5 10-4 10-3 10-2 P(

Appendices Appendix A. Random Click Behavior Experiment Section 3.2 of the main text has verified the semantic similarity distribution, revealing that real user click behavior exhibits a historical interest dependency characteristic (i.e., the ”right-skewed phenomenon”), where the semantic similarity between clicked 17 content and historical interests is ...

2023