A data-driven short video international communication model based on indicator system communication network and attention BiLSTM neural network

The study addresses how to improve the international dissemination effect of short videos, using TikTok as the research context. Traditional communication dynamics models (e.g., SIR/SEIR) have limitations for short video platforms because short video communication is driven more by content, platform recommendation algorithms, and user behavior than by social ties. The paper proposes combining a key objective indicator system (KOIS) with neural networks to capture implicit relationships among indicators and model short video communication. The outcome indicator for communication power is defined as play volume (Watch_count), chosen as a practical proxy for exposure. The research builds an indicator hierarchy (static vs. dynamic; status vs. outcome), constructs a communication network (KOIS-PN) where indicators act as nodes with hierarchical relations, and integrates this with an Attention-enhanced BiLSTM (AT-BiLSTM) to mitigate long-range forgetting and to appropriately handle both time-varying and static features. The overall goal is to improve prediction of short video plays and provide insights into international communication strategies.

Related research spans humanities/social sciences and natural sciences. Humanities/social sciences studies have examined content characteristics and audience feedback on TikTok (e.g., Zhou’s analysis of Chinese culture-themed videos), and constructed influence indices using likes, comments, plays, and production features (Zhang). Natural science studies include modeling TikTok video popularity and dynamics (Zhong), and applying neural networks (e.g., LSTM) for information dissemination prediction (Dang) and epidemiological forecasting outperforming SEIR (Liu). Indicator system research often fuses objective and subjective metrics; examples include government short video dissemination evaluation via Delphi and AHP (Yu) and fuzzy assessment of TikTok account content quality (Nazarov), as well as public opinion early warning indices under information epidemics (Wang). The paper argues subjective indicators are hard to verify and may shift rapidly, while platform-polished objective indicators (likes, comments, shares, play volume, language, content type) reliably reflect user behavior and propagation rules, providing a sound basis for modeling short video communication.

The methodology comprises three parts: constructing a key objective indicator system (KOIS), building a communication network (KOIS-PN), and integrating it with an Attention-based BiLSTM (AT-BiLSTM). 1) Key Objective Indicator System (KOIS): Indicators are categorized by dynamic/static properties and causal roles: - Static indicators (unchangeable post-release): Content, Pattern (production mode), Language (presentation language), Originality. - Dynamic indicators (vary over time): likes (Dig_count), comments (Comment_count), downloads (Download_count), shares (Share_count), play growth rate (Watch_count_speed), play growth acceleration (Watch_count_acceleration), play volume (Watch_count). Dynamic indicators are further divided into Status indicators (do not directly evaluate communication power) and Outcome indicator (play volume, the direct basis for communication power). Rationale: Play volume approximates exposure (Exposure × click rate); many metrics decompose to formulas involving exposure. Since exposure is hard to collect, play volume is used as the evaluative outcome. 2) KOIS-PN Communication Network: A hierarchical network is constructed where static and status indicators jointly determine current play outcomes; temporal feedback links allow current outcomes to influence future status indicators, resembling an RNN-like structure. 3) AT-BiLSTM Model: Attention is introduced into BiLSTM to connect the final forward LSTM state with information across all time steps, alleviating long-distance forgetting and improving robustness to anomalies (e.g., platform audits temporarily suspending traffic). The model uses residual connections to: - Preserve static indicator information by summing Category I features with AT-BiLSTM outputs (Residual ①), mitigating the loss of non-temporal information. - Directly route Category III indicators (growth rate and acceleration) to the decision layer (Residual ②), ensuring smoother flow of crucial growth dynamics. Indicator categories used in the integrated model: Category I: Content, Pattern, Language, Originality. Category II: Dig_count, Comment_count, Download_count, Share_count. Category III: Watch_count_speed, Watch_count_acceleration. Category IV: Watch_count (outcome). Performance metric: Coefficient of determination (R²). Dataset construction: Data were collected from 38 TikTok creators with stable output, focusing on China-related content (Chinese food or Chinese teaching). From video release time onward, likes, comments, downloads, shares, and plays were crawled hourly. The dataset spans 2022-05-01 to 2022-08-17, contains 2,150 videos and 435,962 records (~80 MB). Key fields include user and video identifiers, description, counts (likes, comments, plays, downloads, shares), creation time, and observation timestamps. Experimental design: Multiple control groups compare baseline models (BiLSTM, Multi-Head Attention, AT-BiLSTM) versus their KOIS-PN-integrated counterparts (KOISPN-BiLSTM, KOISPN-Attention, KOISPN-AT-BiLSTM).

- KOISPN-BiLSTM achieved R² ≈ 0.81 (reported also as 0.8093), significantly outperforming a single BiLSTM baseline (R² ≈ 0.38). - KOISPN-AT-BiLSTM achieved R² ≈ 0.846 (reported as 0.8459), outperforming a single AT-BiLSTM baseline (R² ≈ 0.44). - Single Attention model achieved R² ≈ 0.54; KOISPN-Attention achieved R² ≈ 0.53 (comparable performance). - The introduction of KOIS-PN improved BiLSTM performance by about 113% relative to its baseline and improved AT-BiLSTM performance by about 91%. - KOISPN-AT-BiLSTM outperformed KOISPN-BiLSTM; the attention module contributed an additional improvement (~3.7% reported), helping capture long-range dependencies and handle anomalies such as audit-induced playback suspensions. - Observed play volume characteristics: monotonic increase; non-linear growth; growth concentrated soon after release, then slows and stabilizes.

The findings demonstrate that integrating a domain-specific indicator network (KOIS-PN) with sequence models enhances short video communication modeling. Treating static features separately via residual connections prevents their information from being lost in time series processing, while directly feeding rate and acceleration indicators to the decision layer ensures critical growth dynamics inform predictions. Attention within BiLSTM helps the model reference relevant distant time points, making it robust to platform interventions (e.g., audits) that cause short-term anomalies, and better aligned with the realities of short video dissemination driven by content, recommendation algorithms, and user behavior. These improvements address the research goal of enhancing international communication effectiveness by accurately predicting play volume and clarifying how objective indicators influence outcomes, thereby informing strategies for content creation and platform engagement.

The study proposes a data-driven model for short video international communication by: (1) Building a key objective indicator system and dataset; (2) Constructing an indicator-based communication network (KOIS-PN); (3) Designing an Attention-enhanced BiLSTM (AT-BiLSTM) and integrating it with KOIS-PN; and (4) Validating performance. Experiments show the KOISPN-AT-BiLSTM model is valid and advanced, delivering superior play prediction compared to baselines and clarifying mappings from indicators to international communication effects. The framework can guide short video creation and strategy by analyzing indicator-outcome relationships for different content types and goals. Future work could deepen analysis of indicator contributions, incorporate additional platform signals (e.g., exposure), extend to diverse content and platforms, and explore model generalization across cultural contexts.

- Outcome proxy: The model uses play volume as a proxy for exposure due to data accessibility; true exposure data are not included, potentially limiting precision. - Dataset scope: Data are limited to 38 TikTok creators focusing on China-related content (Chinese food or teaching) over a specific period (2022-05-01 to 2022-08-17), which may affect generalizability to other content genres, platforms, or time frames. - Static vs dynamic features: While residuals aim to preserve static feature influence, the approach may still underrepresent complex interactions between static content attributes and temporal user behavior. - Attention-only models showed limited suitability for this task; performance may depend on hyperparameters and architecture choices not exhaustively explored. - Platform-specific interventions (e.g., audits, manual boosts) introduce non-stationarities that are difficult to model without direct platform signals. - The study focuses on front-end metrics; subjective indicators and richer contextual factors (e.g., audience demographics, network effects) are not incorporated.