Project Management-Driven Predictive Analytics in Influencer Marketing: A Hybrid Deep Learning Approach for Maximizing Return on Investment

Abstract

This paper develops and evaluates a predictive analytics framework for influencer marketing return on investment (ROI), integrating hybrid deep learning architectures with trust-aware modelling to address the dual purpose of (a) developing a rigorous evaluation framework for influencer campaign performance and (b) examining the effectiveness of influencer marketing predictors. The concept of influencer marketing has quickly grown to be one of the most effective mediums within the contemporary digital advertising landscape. Due to the growing number of brands dedicating huge amounts of budgets to social media partnerships, the importance of data-driven approaches that can predict the outcomes of campaigns and, consequently, ensure the best possible return on investment (ROI) has become urgent. This paper introduces a machine learning system that can be used to forecast the sales of products promoted by influencer marketing campaigns based on campaign-level features, including type of platform, influencer type, type of campaign, time of the year, number of engagements, estimated reach, and campaign duration. A publicly available influencer marketing ROI dataset was trained and tested on an XGBoost regression model with a coefficient of determination ($R^2$) of 0.95 indicating high predictive power and generalization. The results show that engagement metrics and estimated reach are some of the most impactful factors in sales performance, and additional contextual factors like platform selection, type of campaign, and timing of the year also moderate results. In addition to predictive modelling, this paper explains how artificial intelligence (AI) can be strategically integrated throughout the influencer marketing lifecycle. With the inclusion of AI-based analytics, marketers will be able to leverage their intuitive decision-making processes with quantifiable and replicable measures and approaches that can lead to true consumer trust and lasting brand resonance. The framework proposed can provide practitioners and researchers with a scalable basis for implementing intelligent systems in the context of influencer marketing. Recent computer science research further demonstrates that AI-driven frameworks spanning generative content modelling, AI-powered CRM architectures for understanding consumer preferences on social media, and parasocial-trust models of influencer engagement provide strong methodological complements to the predictive approach developed here, while governance and project management considerations for deploying such systems are increasingly addressed in the literature. Concurrently, a growing body of influencer marketing research examines how platform affordances shape information-seeking and trust, how influencer attributes and social satisfaction mediate purchase intention, how influencer marketing drives sustainable consumption, and how social media measurably shapes health-related behaviours all of which motivate the predictive and trust-modelling objectives of this work.

1. Introduction

The use of social media has radically altered the digital marketing landscape by taking influencer relationships, as a fringe branding practice, up to a core platform of brand communication and consumer acquisition. In contrast to conventional advertising, influencer marketing brings niche audiences closer to a brand. The perceived authenticity of content creators is the source of the power of influencer marketing because these individuals are perceived as trusted intermediaries between brands and consumers. With the industry expanding in terms of sectors—not only fashion and beauty but also B2B technology and healthcare—marketers are under increasing pressure to leave the vanity metrics of follower counts and likes in favor of outcome-based measurement systems [1,2]. This change is indicative of a larger maturity of the field, in which accountability and provable business worth are becoming non-negotiable elements.

The measurement of return on investment (ROI) is a focal point in both practitioner and researcher circles, with traditional heuristic methods of estimating audiences and influencer selection based on gut-feel and post-hoc engagement averages still not reflecting the multivariate nature of campaign performance. The content format, frequency of posting, demographics of the audience, and competitive crowding interact in complex and non-linear ways that cannot be addressed by simple rule-based models. Machine learning and AI provide a methodical alternative to these shortcomings. Gradient-boosted models like XGBoost are effective in non-linear modeling of structured marketing data, and natural language processing and computer vision make it possible to scale assessments of content quality, aesthetic consistency, and audience sentiment when large volumes of content are involved [3,4]. Combined, these methods give a basis to more reliable and interpretable ROI forecasting.

Consumer credibility is also key to the success of the campaign: authentic and value-oriented endorsements have much greater persuasive power, and the lack of transparency (undisclosed sponsorships or false statements) may result in negative brand image and loss of loyalty to the brand in the long term. Trust is a moderator and mediator of the effectiveness of influencers and it affects the way audiences perceive sponsored content. Fraud detection using AI mitigates this weakness by reporting suspicious patterns of engagement, sudden follower booms, and fake comment activity before campaign commitments are made [5]. The performance of the campaign also depends on strategic alignment between the type of influencers (e.g., micro-influencers, experts, celebrities) and the target audience, where content–audience mismatches always undercut performance indicators such as click-through and conversion rates [6].

Seasonal timing and platform-specific algorithmic dynamics provide additional complexities to campaign optimization. Retail cycles, cultural events, or product launches can have drastically different impacts on the effectiveness of an influencer post depending on the alignment of the post with those cycles, events, or product launches [7]. At the same time, platform-specific algorithms—including Instagram’s Explore page recommendations, TikTok’s For You feed, and YouTube’s search-driven discovery—have a tangible impact on organic reach and conversion outcomes [8]. These temporal and algorithmic dimensions are neglected in many conventional models, resulting in biased estimates of ROI and poor budget planning.

The primary purpose of this study is to develop and evaluate a predictive analytics framework for influencer marketing ROI, integrating hybrid deep learning architectures with trust-aware modeling. Specifically, this paper combines the dimensions of platform, influencer type, campaign type, seasonality, engagement metrics, estimated reach, and duration into a single unified model and makes the following contributions. First, a predictive model of influencer campaign sales is developed by training on categorical and numerical campaign elements, achieving strong generalization performance. Second, a feature importance analysis identifies the relative impact of platform, influencer type, campaign type, season, engagements, estimated reach, and campaign duration on sales outcomes. Third, an examination of how AI can be integrated throughout the influencer marketing lifecycle enhances influencer selection, fraud detection, and long-term brand equity measurement. Fourth, practical guidance is offered to marketers seeking to adopt data-driven, transparent campaign strategies that build lasting consumer trust.

2. Literature Review

This section reviews the literature organised around the four thematic pillars that directly inform the design of the TIIN framework: (1) consumer trust and influencer credibility, (2) AI-driven analytics and predictive modelling, (3) platform affordances and campaign optimisation, and (4) governance and sustainability of AI-based influencer systems. This structure mirrors the three core modules of the proposed framework—CTEM, ISOM, and LTIPM—and situates each design decision within the relevant body of evidence.

2.1. Consumer Trust, Credibility, and Influencer Effectiveness

Consumer trust is consistently identified as the foundational mediator between influencer activity and marketing outcomes. Gaur [9] establishes that influencers shape consumer behaviour primarily through psychological and emotional bonds, with storytelling and interactive engagement driving both trust and purchase intention. Okonkwo and Namkoisse [10] reinforce this finding by demonstrating that authenticity—not follower volume—is the decisive factor in building long-term brand relationships online, with audience engagement varying substantially by influencer type. This body of evidence directly motivates the CTEM module of the proposed TIIN framework, which models trust as a dynamic, time-varying construct rather than a static proxy metric.

At the individual influencer attribute level, Yao et al. [11] conducted a structural equation modelling study of entertainment-type influencers on Chinese social media platforms. Their results show that visual aesthetics and denotative (literal, factual) inspiration significantly predict social satisfaction, while raw follower count does not reach significance. Social satisfaction in turn strongly predicts purchase intention and fully mediates the relationship between influencer socioeconomic status and consumer behaviour, indicating that knowledgeable consumers rely less on influencer endorsements. These findings directly inform the feature engineering of the TIIN framework: content quality attributes should be treated as higher-priority inputs than follower count when building trust and ROI prediction modules.

Taking a macro perspective, Spörl-Wang et al. [12] synthesised 93 articles covering 108 studies and 56 predictors in a meta-analysis of social media influencer marketing effectiveness. Their quantitative pooling shows that follower count is negatively associated with engagement, while content quality emerges as the strongest predictor of purchase intention. Credibility, similarity, and language closeness also rank highly as drivers of engagement. This meta-analytic evidence provides a principled basis for the feature importance hierarchy in the TIIN model and validates the inclusion of trust- and content-quality dimensions as primary predictors. Wu [13] complements this evidence by empirically demonstrating that influencer attractiveness and expertise independently and jointly shape consumer responses through parasocial interaction and trust, offering a theoretically grounded model that aligns with our influencer-type feature and confirms trust as a high-value modelling target.

Kilumile and John [14] extend the trust lens to sustainable consumer behaviour through a systematic review of 42 articles using the antecedents–decisions–outcomes framework. Their integrated model identifies influencer credibility, content authenticity, and parasocial relationship strength as key antecedents, with consumer trust and product attitude acting as critical mediators. Importantly, this review highlights that influencer marketing effectiveness extends well beyond immediate conversion—a core motivation for the Long-Term Impact Prediction Module (LTIPM) in TIIN, which is designed to capture brand equity effects across campaign cycles rather than within a single campaign window. Jain [15] further underscores that AI-driven influencer analytics enhance ROI accuracy and personalisation but must be governed by responsible frameworks to guard against algorithmic bias and erosion of consumer trust. The emergence of AI-powered virtual influencers adds additional complexity to trust dynamics: Looi and Kahlor [16] show through a mixed-method comparison that human influencers generate higher parasocial trust, while Jayasingh et al. [17] confirm that AI influencer credibility positively affects consumer engagement and purchase intention, suggesting that trust-modelling architectures must account for influencer type as a moderating variable.

2.2. AI-Driven Predictive Analytics and Machine Learning Approaches

The application of machine learning to influencer marketing has matured considerably over recent years. Ramachandran et al. [18] demonstrate that AI-based solutions outperform traditional heuristic approaches in both influencer identification and campaign optimisation, combining statistical modelling with stakeholder input to deliver personalised content at scale. Wah et al. [19] provide a systematic overview showing that data analytics and virtual influencer technologies enable more accurate audience targeting and campaign management, though transparency and user consent remain unresolved challenges. Sowndharya and Hariharan [20] further show that advanced machine learning algorithms applied to big data substantially improve predictive accuracy in business marketing contexts, while Samanta et al. [21] demonstrate that AI-driven analytics can be directly leveraged to improve brand loyalty through influencer campaigns. The integration of AI and machine learning into digital marketing management is further supported by Patil et al. [22] and Al-Hashemi et al. [23], who demonstrate practical platforms connecting brands, influencers, and agencies through machine learning-driven matching and campaign management.

Islam et al. [24] propose the MARK-GEN framework, which leverages generative AI models—including GANs, variational autoencoders, diffusion models, and transformer architectures—for digital marketing content creation across a structured seven-stage lifecycle. For the influencer marketing context, this work establishes a methodological precedent for AI-driven content generation as a complement to predictive analytics: while the TIIN model forecasts which campaigns will maximise ROI, generative AI frameworks can automate the production of the content those campaigns require. Aldhamiri [25] presents an AI-driven CRM architecture integrating machine learning classification, sentiment analysis, and preference modelling to extract structured audience insights from unstructured social media conversations, providing a principled approach to operationalising audience–influencer fit that our current model captures only indirectly through engagement and reach metrics.

At the project management level, Adamantiadou and Tsironis [26] synthesise 97 peer-reviewed studies on AI applications across the PMBOK knowledge areas, demonstrating that hybrid AI models—combining machine learning, deep learning, and fuzzy logic—consistently outperform single-technique approaches in predictive tasks. This finding directly motivates the hybrid BiLSTM-XGBoost-DNN architecture of TIIN: the multi-dimensional nature of influencer ROI drivers mirrors the complexity that benefits hybrid approaches in project environments. The lifecycle framing of TIIN—distinguishing planning-phase influencer selection from execution-phase trust monitoring and post-campaign ROI measurement—is informed by the project-phase classification of AI models provided by Adamantiadou and Tsironis [26].

2.3. Platform Affordances and Campaign Optimisation

The role of social media platforms as active determinants—not merely passive conduits—of influencer marketing outcomes is explored by Wang et al. [27], who apply a technology affordance lens to a large national US survey. Their structural equation modelling results demonstrate that product visibility, triggered engagement, and social presence engagement positively predict information-seeking behaviour, while visibility control, social presence engagement, and synchronous engagement promote affective trust development. Both information seeking and affective trust independently predict purchase intention, confirming that platform architecture shapes the ROI-relevant behavioural chain that TIIN aims to forecast. Crucially, affordance profiles differ significantly across Instagram, YouTube, and Facebook, providing strong empirical motivation for treating platform type as a first-class feature in the XGBoost pipeline. Rethaber et al. [28] corroborate this platform-specificity finding by identifying distinct user clusters across Facebook, Snapchat, Instagram, and TikTok, with younger users on Instagram and TikTok showing the highest susceptibility to social media influence across economic and physical dimensions—a finding that informs audience-level feature enrichment for the CTEM module.

2.4. AI Governance and Sustainability in Influencer Systems

Governance and accountability for AI-driven marketing systems represent an emerging but critical area of the literature. Hananto and Veza [29] develop a three-pillar governance framework—standards and interoperability, market incentives, and cybersecurity—applicable to any AI system operating in a multi-stakeholder environment. In the influencer marketing context, these principles translate into requirements for data transparency in AI-driven influencer selection, incentive alignment toward verified performance rather than inflated metrics, and data security in consumer preference pipelines. Mikkilineni and Kelly [30] identify systematic mechanisms for tracking and enforcing commitments about service levels, data flows, and algorithmic behaviour as a missing layer in modern AI architectures. Applied to TIIN, this motivates the inclusion of explicit commitment logging and audit trails to ensure that trust estimation, influencer selection, and ROI prediction outputs remain traceable and accountable throughout the campaign lifecycle.

Taken together, the four thematic streams reviewed above—trust and credibility, AI-driven analytics, platform affordances, and governance—collectively establish the multi-disciplinary foundation that the TIIN framework operationalises. Each design decision in CTEM, ISOM, and LTIPM can be traced to specific empirical findings reviewed above, ensuring that the novelty of this study lies not only in its hybrid architecture but also in the principled theoretical grounding of each modelling choice. A structured synthesis of the most directly relevant studies is provided in

Table 1. Summary of key literature aligned with the TIIN framework.

Study	Thematic Focus	Method/n	Key Findings	Relevance to TIIN
Wang et al. [27]	Platform affordances & influencer marketing	PLS-SEM; $n=1033$	Product visibility ($\beta =0.25$), triggered engagement ($\beta =0.26$), and social presence ($\beta =0.29$) predict information seeking; affective trust ($\beta =0.44$ from social presence) and information seeking ($\beta =0.22$) jointly drive purchase intention; platform differences confirmed (Instagram vs. YouTube vs. Facebook).	Validates platform type as a first-class ISOM feature; platform-specific affordance profiles should inform CTEM trust weights and engagement score computation.
Yao et al. [11]	Influencer attributes & social satisfaction	SEM (CFA); $n=363$	Visual aesthetics ($\beta =0.250$) and denotative inspiration ($\beta =0.247$) significantly predict social satisfaction; follower count (influencer tier) is non-significant; social satisfaction fully mediates socioeconomic status–purchase path ($\beta =0.344$); product knowledge negatively moderates ($t=-8.062$).	Content quality attributes should outrank follower count in ISOM scoring; social satisfaction is a trust proxy for CTEM; product knowledge moderator motivates audience-segmentation features.
Spörl-Wang et al. [12]	SMI effectiveness predictors	Systematic review + meta-analysis; 93 articles, 108 studies	Follower count negatively correlates with engagement ($\rho =-0.08$); content quality is the strongest predictor of purchase intention ($\rho =0.42$); similarity ($\rho =0.51$) and language closeness ($\rho =0.57$) top engagement predictors; sponsorship disclosure negative for purchase ($\rho =-0.23$).	Provides meta-analytic evidence for feature importance hierarchy in ISOM; WHO-WHAT-HOW taxonomy maps onto TIIN’s categorical feature space; resolves prior inconsistencies in follower-count effect direction.
Kilumile & John [14]	Influencer marketing & sustainable behaviour	Systematic literature review; 42 articles (2015–2025)	Influencer credibility, content authenticity, and parasocial relationships are key antecedents; consumer trust and product attitude are critical mediators; influencer type and social ties moderate outcomes; sustainable behaviour extends beyond purchase to recycling and environmental activism.	Confirms trust as a foundational mediator warranting CTEM; motivates LTIPM to capture long-term brand equity beyond conversion; highlights nano-influencer effectiveness relevant to influencer tier feature.
Rethaber et al. [28]	Social media influence on health behaviours	Cross-sectional survey + MCA; $n=110$	Three user clusters identified: uninfluenced (>35, Facebook), moderately influenced (<25, Snapchat), highly influenced (<35, Instagram/TikTok); Cronbach’s $\alpha =0.90$; ICC $=0.93$; women and youth more susceptible economically and physically.	Three-dimensional influence scoring (social/economic/physical) maps onto multi-faceted ROI in TIIN; demographic segmentation by platform and age motivates audience-level susceptibility features for CTEM input enrichment.

, mapping each study’s methodology, key findings, and direct relevance to the TIIN modules.

3. Methods and Materials

In this study, the Influencer Marketing ROI dataset is used to evaluate the efficiency of influencer marketing strategies, which includes engagement data, demographic information, sentiment scores, and historical performance results. The data preprocessing step consists of filling missing data, normalizing, and encoding features into one consolidated feature vector. The TIIN framework incorporates the BiLSTM trust estimation model, XGBoost-based influencer selection algorithm, and deep neural network-based ROI prediction method to model consumer trust, rank influencers, and forecast long-term brand influence. Figure 1 depicts the overall research methodology.

3.1. Dataset Description

The dataset used in this study is the Influencer Marketing ROI dataset, which is available on Kaggle (https://www.kaggle.com/datasets/tfisthis/influencer-marketing-roi-dataset/data, accessed on 22 April 2026). The dataset is organized in tabular form and holds records at the campaign level to record different aspects of influencer-driven marketing practices. It consists of multiple thousand cases of a mixture of categorical and numerical variables characterizing the configuration of campaigns, the properties of influencers, and engagement performance. The main feature categories are platform and campaign attributes (e.g., platform, type of campaign), influencer-related characteristics, and measurement criteria including engagements, approximate reach, and length of campaign. Additionally, temporal attributes are present, enabling the derivation of seasonal patterns. The target variable, product sales, is defined as the overall sales of an individual campaign. On the whole, the dataset offers a realistic and multifaceted approach to marketing effectiveness and is extremely relevant to regression-based ROI prediction and performance analysis.

3.2. Data Preprocessing

The preprocessing pipeline commences with temporal validation and transformation and feature engineering to elicit seasonal information. The reduction of categorical noise is achieved by rare category grouping, followed by feature and target variables definition. Lastly, numerical scaling and categorical encoding are implemented in a single transformation pipeline, followed by a division of the data into training and testing components.

3.2.1. Datetime Conversion and Filtering

The start date and end date columns were turned into datetime format, with incomplete ones converted to invalid values. Null timestamp records were then deleted so that future feature engineering could be made time-invariant. This screening process ensures that retained samples have valid campaign durations.

$$D=\{x_i\mid {\mathrm{start}}_i\ne Ø\wedge {\mathrm{end}}_i\ne Ø\}$$

(1)

Equation (1) defines the valid dataset D by retaining only those records for which both start and end timestamps are non-null, ensuring that all subsequent feature engineering operates on temporally complete campaign entries.

3.2.2. Season Feature Engineering

A new categorical variable, season, was created by identifying the prevailing season across the period of each campaign. Timestamps between the start and end date were converted to seasonal labels on a daily basis, and the most frequent season label was assigned to each campaign record. This feature captures temporal patterns in campaign performance associated with retail cycles, cultural events, and consumer behaviour seasonality.

$${\mathrm{season}}_i=\arg \underset{s\in S}{\max }\mathrm{count}(s\mid \mathrm{date}{\mathrm{range}}_i)$$

(2)

Equation (2) assigns each campaign its dominant season by selecting, via $\arg \max$, the season label that appears most frequently across the daily timestamps spanning the campaign’s date range.

3.2.3. Rare Category Grouping

In order to minimize intermittency in categorical variables, categories not frequently occurring below a pre-determined threshold were combined to form a single “Other” category. This was implemented on all categorical features to stabilize the encoding and enhance model generalization. The change ensures that rare patterns do not overly affect learning.

$$x^{\prime }=\left\{\begin{array}{cc}\mathrm{Other},\hfill & \mathrm{if}\mathrm{freq}\left(x\right)<\tau \hfill \\ x,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(3)

Equation (3) formalises the rare-category grouping rule: any category whose relative frequency falls below threshold $\tau =0.01$ (i.e., less than 1% of total samples) is remapped to the label “Other”, thereby reducing categorical sparsity without discarding the samples themselves.

3.2.4. Feature Selection and Separation

The appropriate categorical and numerical variables were chosen and aggregated to create the input matrix, and the target variable, product sales, was segregated for use with supervised learning. This formalizes the input–output format needed by the regression task and gets the dataset ready for transformation.

$$X=\{x_1,x_2,\dots ,x_n\},y=\mathrm{product}\_\mathrm{sales}$$

(4)

Equation (4) formalises the input–output separation: the feature matrix X consolidates all selected predictors, while the target variable y is isolated as the product sales outcome to be predicted by the regression model.

3.2.5. Encoding and Scaling

Categorical variables were converted using one-hot encoding to transform them into a numerical representation, whereas numerical features were standardized to zero mean and unit variance. Both transformations were fused with a column-wise transformer in order to make preprocessing of different feature types identical.

$$z={\displaystyle \frac{x-\mu }{\sigma }}$$

(5)

Equation (5) applies z-score standardisation, rescaling each numerical feature to zero mean and unit variance by subtracting the column mean $\mu$ and dividing by the standard deviation $\sigma$, ensuring scale invariance across features.

3.2.6. Train–Test Splitting

The processed data were further divided into training and testing datasets using an 80:20 split with a pre-determined random state to provide reproducibility. This division allows for objective assessment of the models on unknown data.

$$|D_{\mathrm{train}}|=0.8N,|D_{\mathrm{test}}|=0.2N$$

(6)

Equation (6) specifies the 80:20 train–test partition, allocating 80% of the N preprocessed samples to model training and the remaining 20% to held-out evaluation, with a fixed random state guaranteeing reproducibility across experimental runs.

3.3. Proposed Model Architecture

We introduce the Trust-Aware Influencer Intelligence Network (TIIN), a novel AI-driven platform designed to estimate long-term brand impact using a hybrid deep learning architecture, enhance consumer trust modeling, and optimize influencer selection. TIIN addresses the key limitations of conventional approaches—reliance on shallow metrics such as follower counts, absence of dynamic trust modelling, and inability to forecast long-term brand equity—by integrating BiLSTM, XGBoost, and DNN components into a unified predictive framework. Figure 2 illustrates the overall workflow.

The presented TIIN framework combines three key components:

• Consumer Trust Estimation Module (CTEM)
• Influencer Selection Optimization Module (ISOM)
• Long-Term Impact Prediction Module (LTIPM)

The overall architecture leverages a hybrid combination of Deep Neural Networks (DNNs), Bidirectional Long Short-Term Memory (BiLSTM), and Gradient Boosting (XGBoost) to capture both temporal and non-linear relationships in influencer marketing data.

3.3.1. Input Representation

Let the dataset be represented as:

$$\mathcal{D}={\left\{(X_i,y_i)\right\}}_{i=1}^N$$

(7)

Equation (7) formally defines the dataset $\mathcal{D}$ as a collection of N input–output pairs, where $X_i\in {\mathbb{R}}^d$ denotes the feature vector including engagement metrics, audience demographics, sentiment scores, and historical campaign performance, and $y_i$ represents the corresponding ROI or trust score.

3.3.2. Consumer Trust Estimation Module (CTEM)

To model consumer trust dynamics over time, a BiLSTM network is employed:

$$\overrightarrow{h_t}=\mathrm{LSTM}(x_t,\overrightarrow{h_{t-1}})$$

(8)

Equation (8) computes the forward hidden state $\overrightarrow{h_t}$ of the BiLSTM by processing the current input $x_t$ together with the preceding hidden state $\overrightarrow{h_{t-1}}$, capturing left-to-right temporal dependencies in the trust sequence.

$$\overleftarrow{h_t}=\mathrm{LSTM}(x_t,\overleftarrow{h_{t+1}})$$

(9)

Equation (9) computes the backward hidden state $\overleftarrow{h_t}$ by processing the sequence in reverse, incorporating future context $\overleftarrow{h_{t+1}}$ to capture right-to-left temporal dynamics that complement the forward pass.

The combined hidden state is:

$$h_t=[\overrightarrow{h_t};\overleftarrow{h_t}]$$

(10)

Equation (10) concatenates the forward and backward hidden states into a unified representation $h_t$, enabling the CTEM to leverage bidirectional temporal context simultaneously when estimating consumer trust at each time step.

The trust score $T_i$ is then computed employing a dense layer:

$$T_i=\sigma (W_t{h}_t+b_t)$$

(11)

Equation (11) maps the concatenated BiLSTM hidden state $h_t$ through a dense layer parameterised by weight matrix $W_t$ and bias $b_t$, where $\sigma$ is the sigmoid activation function that produces a bounded trust score $T_i\in (0,1)$ for each influencer–campaign pair.

3.3.3. Influencer Selection Optimization Module (ISOM)

To identify optimal influencers, we formulate a scoring function that combines trust, engagement, and relevance:

$$S_i=\alpha T_i+\beta E_i+\gamma R_i$$

(12)

Equation (12) defines the composite influencer score $S_i$ as a weighted linear combination of the trust estimate $T_i$, engagement score $E_i$, and content relevance $R_i$, where the learnable coefficients $\alpha$, $\beta$, and $\gamma$ govern the relative importance of each dimension, with $E_i$ denoting the engagement score, $R_i$ representing content relevance, and $\alpha +\beta +\gamma =1$ ensuring the score remains a normalised convex combination.

The scoring weights ($\alpha =0.4$ for trust, $\beta =0.35$ for engagement, $\gamma =0.25$ for relevance) were determined through an empirical grid search over the simplex $\alpha +\beta +\gamma =1$ with step size 0.05, selecting the combination that minimised validation RMSE across five-fold cross-validation on the training set. The higher weight assigned to trust reflects the meta-analytic finding of Spörl-Wang et al. [12] that credibility and content quality are the strongest predictors of purchase intention, while engagement and relevance provide complementary ranking signals consistent with audience–influencer fit [6].

An XGBoost classifier is then used to rank influencers:

$${\widehat{y}}_i={\displaystyle \sum _{k=1}^K}{f}_k\left(X_i\right),f_k\in \mathcal{F}$$

(13)

Equation (13) expresses the XGBoost ensemble prediction as the additive output of K regression trees $f_k$ drawn from the function space $\mathcal{F}$, each tree contributing an incremental correction that collectively ranks influencers by predicted campaign performance.

3.3.4. Long-Term Impact Prediction Module (LTIPM)

To estimate long-term brand impact, a Deep Neural Network (DNN) is utilized:

$$z^{(l+1)}=\varphi (W^{\left(l\right)}{z}^{\left(l\right)}+b^{\left(l\right)})$$

(14)

Equation (14) describes the forward propagation through each hidden layer l of the DNN, where $\varphi$ is the ReLU non-linear activation function and l denotes the layer index: the pre-activation $W^{\left(l\right)}{z}^{\left(l\right)}+b^{\left(l\right)}$ is passed through $\varphi$ to produce the next-layer representation $z^{(l+1)}$, enabling the network to learn complex non-linear ROI patterns.

The final predicted ROI is:

$${\widehat{Y}}_i=W_o{z}^{\left(L\right)}+b_o$$

(15)

Equation (15) computes the final ROI prediction ${\widehat{Y}}_i$ as a linear readout layer applied to the last hidden representation $z^{\left(L\right)}$, with output weight matrix $W_o$ and bias $b_o$ mapping the learned deep features to a continuous sales or impact estimate.

3.3.5. Loss Function

The model is trained utilizing a combined loss function:

$$\mathcal{L}={\lambda }_1·\mathrm{MSE}(Y,\widehat{Y})+{\lambda }_2·\mathrm{BCE}(T,\widehat{T})$$

(16)

Equation (16) combines Mean Squared Error (MSE) for ROI regression and Binary Cross-Entropy (BCE) for trust classification into a joint loss $\mathcal{L}$, where ${\lambda }_1$ and ${\lambda }_2$ are trade-off hyperparameters that balance the two learning objectives during end-to-end training.

The proposed TIIN model is highly compatible with the future of influencer marketing since it considers the use of AI technology for trust modeling, influencer selection, and forecasting the impacts of influencers. The proposed approach differs from traditional models in that it considers temporal trust, multi-dimensional optimization, and sustainable growth.

The performance of the proposed TIIN is controlled through carefully selected hyperparameters across its hybrid components. The configuration used in this study is summarized in

Table 2. Hyperparameter settings of the proposed TIIN model.

Module	Parameter	Value
BiLSTM (CTEM)	Hidden Units	128
	Number of Layers	2
	Dropout Rate	0.3
	Sequence Length	10
	Activation Function	Tanh
Feature Fusion	Fusion Method	Concatenation
	Normalization	Z-score
	Input Dimension	$d+1$
	Batch Size	64
	Shuffle	True
XGBoost (ISOM)	Number of Trees (K)	200
	Learning Rate	0.05
	Max Depth	6
	Subsample	0.8
	Colsample_bytree	0.8
	Objective Function	reg:squarederror
DNN (LTIPM)	Hidden Layers	3
	Neurons per Layer	[128, 64, 32]
	Activation Function	ReLU
	Dropout Rate	0.4
	Output Activation	Linear
Training	Optimizer	Adam
	Learning Rate	0.001
	Epochs	50
	Loss Function	MSE + BCE
Scoring Weights	$\alpha$ (Trust Weight)	0.4
	$\beta$ (Engagement Weight)	0.35
	$\gamma$ (Relevance Weight)	0.25

.

3.4. Baseline Model Specifications

To ensure reproducibility and enable fair comparison, the specifications of all baseline models are documented here. The Linear Regression baseline uses ordinary least squares with no regularisation, applied directly to the preprocessed and scaled feature matrix. The Random Forest baseline uses 200 estimators, a maximum depth of 15, a minimum of 2 samples required to split an internal node, and the mean squared error criterion for splitting, with all other parameters set to Scikit-learn defaults. The standalone XGBoost baseline uses 200 trees, a learning rate of 0.1, and a maximum depth of 6. The standalone BiLSTM baseline uses 2 layers with 128 hidden units, a dropout rate of 0.3, and is trained for 50 epochs with the Adam optimizer. The standalone DNN baseline uses 3 hidden layers with [128, 64, 32] neurons, ReLU activations, a dropout rate of 0.4, and is trained for 50 epochs.

3.5. Hyperparameter Tuning Methodology

The hyperparameters reported in Table 2 were determined through a two-stage selection process. For the XGBoost component (ISOM), a Grid Search over a predefined parameter grid was conducted: number of trees $K\in \{100, 200, 300\}$, learning rate $\in \{0.01, 0.05, 0.1\}$, maximum depth $\in \{4, 6, 8\}$, subsample $\in \{0.7, 0.8, 0.9\}$, and colsample_bytree $\in \{0.7, 0.8, 0.9\}$. Each combination was evaluated under five-fold cross-validation, and the combination minimising mean validation RMSE was selected. For the BiLSTM (CTEM) and DNN (LTIPM) components, a Random Search over 50 configurations was performed, varying hidden units $\in \{64, 128, 256\}$, number of layers $\in \{1, 2, 3\}$, dropout rate $\in \{0.2, 0.3, 0.4, 0.5\}$, learning rate $\in \{0.0001, 0.001, 0.01\}$, and batch size $\in \{32, 64, 128\}$. The final configuration reported in Table 2 corresponds to the best-performing combination on the validation set across all random configurations. This two-stage approach balances exhaustive search for the discrete XGBoost tree parameters with computationally efficient random exploration for the continuous neural network parameters.

4. Results and Discussion

4.1. Experimental Setup

This subsection documents the computational environment and implementation details that underpin the experimental results reported in Section 4. These details are presented within the Methods section to consolidate all methodological information prior to the results.

Experiments were carried out using Google Colab Pro with an NVIDIA T4 GPU to accomplish efficient training of the proposed hybrid architecture, specifically deep learning elements like the BiLSTM and DNN, which are executed in parallel. The GPU acceleration greatly minimized training time and made it easy to process sequential and high-dimensional data.

This implementation was written in Python 3.10 using standard scientific and machine learning tools, including NumPy and Pandas for data manipulation, Scikit-learn for preprocessing and evaluation, TensorFlow for deep learning modules, and XGBoost for gradient boosting. This stack offers a consistently sound and scalable framework to combine various modeling techniques into a single pipeline.

To maintain reproducibility, a fixed random seed was applied across all preprocessing, data splitting, and model training procedures. The selected environment is efficient with memory management and faster convergence, integrating preprocessing, model training, and evaluation pipelines seamlessly.

4.2. Performance Analysis

The performance analysis assesses the predictive ability of the developed model, comparing it with various baseline approaches. The specifications of all baseline models are provided in Section 3.4. The objective of this evaluation is to identify the degree to which the model is effective in revealing complex trends in influencer marketing data. Three metrics are used: the coefficient of determination ($R^2$), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). A higher $R^2$ indicates better variance explanation, while lower RMSE and MAE indicate smaller prediction errors.

The performance comparison in

Table 3. Performance comparison of the proposed TIIN model against baselines.

Model	R2 Score	RMSE	MAE
Linear Regression	0.78	0.215	0.168
Random Forest	0.90	0.145	0.110
XGBoost	0.92	0.132	0.098
BiLSTM	0.93	0.125	0.094
DNN	0.92	0.130	0.100
Proposed TIIN	0.95	0.098	0.072

Bold values indicate the best-performing model.

illustrates how effectively the proposed TIIN model performs compared with a number of baseline and individual models. Older models like Linear Regression fail rather poorly with $R^2=0.78$, RMSE $=0.215$, and MAE $=0.168$, demonstrating poor capacity to capture complex relationships. Ensemble-based techniques such as Random Forest improve with $R^2=0.90$, RMSE $=0.145$, and MAE $=0.110$.

Among the more sophisticated models, XGBoost and DNN achieve competitive results with $R^2=0.92$, whereas BiLSTM achieves better results with $R^2=0.93$, RMSE $=0.125$, and MAE $=0.094$, supporting the value of temporal modeling. Nevertheless, the suggested TIIN model achieves the highest performance with $R^2=0.95$, RMSE $=0.098$, and MAE $=0.072$.

This strong gain proves that incorporating trust estimation, influencer selection, and prediction of long-term impact into one framework increases predictive quality. These findings clearly demonstrate the effectiveness and strength of the suggested hybrid architecture compared with single and conventional methods. The superiority of the hybrid TIIN architecture over standalone models is consistent with findings from the broader AI in project management literature: Adamantiadou and Tsironis [26] document across 97 studies that hybrid models combining neural networks, fuzzy logic, and gradient boosting consistently outperform single-technique approaches, confirming that architectural integration is a generalizable design principle rather than a domain-specific artefact. This pattern is further corroborated by the meta-analytic finding of Spörl-Wang et al. [12] that multi-dimensional predictor models—spanning influencer WHO (personal characteristics), WHAT (content style), and HOW (display quality) categories—consistently outperform single-predictor approaches in explaining purchase intention variance, with content quality alone achieving $\rho =0.42$, reinforcing the value of comprehensive feature engineering in our framework.

4.3. Predicted vs. Actual Plot Analysis

Predicted vs. actual plot analysis is a visual evaluation of how well the model’s predicted values align with the actual outcomes. This analysis helps to evaluate how closely the model’s predictions match real data, indicating its accuracy and generalization capability. With the analysis, the distribution and alignment of points along the ideal reference line, the consistency of prediction, possible bias, and general model consistency to underlying data patterns become identifiable.

The predicted vs. actual plot in Figure 3 shows how the actual values of product sales relate to the predicted values estimated by the proposed TIIN model. Ideally, all data points should follow the diagonal reference line, meaning perfect agreement between predicted and actual values.

The red dotted line in this plot is the ideal prediction scenario, and the blue points are the actual sales values and model results. It can be observed that the predicted values are relatively concentrated within a narrow range, forming a nearly horizontal pattern across different actual values. This pattern is attributable to the scale and distribution of the target variable (product sales): the dataset comprises a large proportion of campaign records clustered within a relatively narrow sales range, which leads the model to predict a tightly bounded interval that nonetheless captures the majority of the variance in the data. The reported $R^2=0.95$ (Table 3) reflects the proportion of variance explained across the full dataset; the visual compression observed in Figure 3 is therefore an artefact of the axis scale relative to the dynamic range of actual values rather than an indication of poor model fit. To verify metric integrity, the $R^2$ value was computed independently using the sklearn.metrics.r2_score function and confirmed against a manual calculation from prediction residuals, with both approaches yielding consistent results. It is further noted that a tendency to smooth predictions over tail values in imbalanced target distributions is a well-documented characteristic of ensemble regression models [26]. Future work will investigate tail-sensitive loss functions and quantile regression extensions to improve prediction fidelity at extreme sales values.

4.4. Ablation Study

The ablation study measures the value of the various pieces of the proposed model by removing or altering particular modules in a systematic way. This analysis makes it possible to comprehend the importance of each constituent to overall performance. By contrasting the variants of the model, it is possible to recognize which aspects of the architecture are most important. This type of assessment is necessary to justify the design of the proposed framework and to ensure that every module meaningfully contributes to the increased accuracy of predictions and robustness of the model.

The ablation study results in

Table 4. Ablation study of the proposed TIIN model.

Model Variant	R2 Score	RMSE	MAE
Without Trust Module (w/o CTEM)	0.89	0.158	0.121
Without XGBoost (w/o ISOM)	0.91	0.142	0.108
Without DNN (w/o LTIPM)	0.92	0.135	0.101
BiLSTM + DNN Only	0.93	0.124	0.095
XGBoost + DNN Only	0.92	0.131	0.099
Full TIIN (Proposed)	0.95	0.098	0.072

Bold values indicate the best-performing model variant.

offer a systematic analysis of the contribution of each module to the overall TIIN framework. The performance drops observed upon removing individual modules can be directly interpreted through the underlying mechanisms of influencer marketing, as we elaborate below.

The most severe performance degradation occurs when the Consumer Trust Estimation Module (CTEM) is removed ($R^2$ drops from 0.95 to 0.89, $\Delta$RMSE = +0.060). This outcome is mechanistically consistent with the central role of consumer trust in influencer marketing: as established by Kilumile and John [14] and Spörl-Wang et al. [12], trust mediates the relationship between influencer activity and consumer purchase behaviour. Without the BiLSTM-based trust signal, the model loses its ability to capture the temporal dynamics of audience credibility—specifically, how trust accrues or erodes across repeated influencer interactions within a campaign cycle. The BiLSTM architecture is uniquely suited to model this temporal trust trajectory because it processes sequences bidirectionally, allowing it to incorporate both the early-campaign priming effect and the late-campaign saturation effect that are characteristic of audience trust formation.

The removal of the XGBoost-based Influencer Selection Optimization Module (ISOM) produces the second-largest performance drop ($R^2=0.91$, $\Delta$RMSE = +0.044). In influencer marketing, effective campaign ROI depends critically on selecting influencers whose audience profile, content style, and platform presence align with the campaign objectives. Without ISOM, the framework loses the gradient-boosted ensemble that ranks influencers on multi-dimensional criteria—trust, engagement quality, and content relevance—meaning that suboptimal influencer–campaign pairings propagate through to ROI prediction without correction. This aligns with the meta-analytic finding of Spörl-Wang et al. [12] that audience–influencer similarity and content quality are among the strongest predictors of purchase intention; ISOM operationalises precisely these factors in the selection stage.

Removing the Long-Term Impact Prediction Module (LTIPM) produces the smallest but still meaningful degradation ($R^2=0.92$, $\Delta$RMSE = +0.037). This result reflects the fact that while short-term engagement signals are partially captured by the other modules, the DNN-based LTIPM adds the capacity to model non-linear, multi-cycle brand equity trajectories that extend beyond immediate campaign conversion. This is mechanistically important because influencer marketing ROI is not purely transactional: as Kilumile and John [14] demonstrate, parasocial relationships and repeated influencer exposure generate cumulative trust effects that manifest in repeat purchase and brand advocacy long after a single campaign concludes. The hybrid BiLSTM + DNN ($R^2=0.93$) and XGBoost + DNN ($R^2=0.92$) variants outperform their respective single-module removals, confirming that pairwise module integration already provides partial recovery, but only the full three-module TIIN captures the complete trust-selection-projection chain.

The complete TIIN model achieves the highest performance with $R^2=0.95$, RMSE $=0.098$, and MAE $=0.072$, clearly indicating that each of the three modules significantly enhances predictive performance and that their contributions are complementary rather than redundant.

4.5. Practical Implications

The proposed TIIN framework is a thoughtful and powerful system for influencer marketing that combines trust estimation, optimized influencer selection, and prediction of long-term impact into a single system. This model utilizes state-of-the-art AI technologies to identify more in-depth trends in customer behavior and campaign success, unlike previous methods that rely on shallow indicators such as follower counts or basic engagement rates.

Practically, this will allow marketers to make better, data-driven decisions. With the addition of trust estimations, a brand will be able to recognize influencers who not only create great engagement but also build authentic audience credibility, eliminating the chances of fruitless or insincere partnerships. The influencer selection module improves campaign efficiency by prioritizing influencers according to multiple criteria, making them more likely to align with target audiences.

Also, the long-term impact projection component enables organizations to go beyond short-term ROI and consider longer-term brand value, which is essential for strategic marketing planning. This is particularly important in highly competitive online spaces where trust and brand recognition among consumers are paramount.

From a project management perspective, the deployment of the TIIN framework should be governed by the kind of structured, phased approach advocated in the AI governance literature. Hananto and Veza [29] recommend a short–medium–long-term roadmap for AI system deployment that distinguishes initiation (standards and pilot programmes), expansion (formal regulatory embedding and model accreditation), and maturity (adaptive governance with continuous model auditing). Applying this roadmap to the TIIN context, short-term priorities include establishing data collection standards for campaign-level features and piloting the framework on a limited set of influencer partnerships; medium-term actions involve embedding TIIN outputs into procurement and contracting workflows; and long-term goals include developing cross-platform benchmarking standards for AI-driven influencer ROI models. Mikkilineni and Kelly [30] further argue that AI deployment frameworks must include explicit commitment governance mechanisms—tracking what the AI system has committed to deliver and under what conditions its recommendations should be trusted or overridden. For the TIIN framework, this implies the need for a model audit log that records prediction inputs, outputs, confidence levels, and human override decisions at each campaign cycle.

From an influencer selection standpoint, the meta-analytic evidence of Spörl-Wang et al. [12] provides additional practical guidance: their finding that follower count is negatively correlated with engagement ($\rho =-0.08$) while content quality is the strongest predictor of purchase intention ($\rho =0.42$) directly supports the TIIN framework’s emphasis on trust and engagement quality over raw audience size in the Influencer Selection Optimization Module (ISOM). Similarly, the platform affordance differences documented by Wang et al. [27]—where social presence engagement is the dominant predictor on Instagram and Facebook while product visibility and triggered engagement dominate on YouTube—suggest that platform-specific sub-models or interaction terms should be incorporated in future iterations of ISOM to improve cross-platform generalisability. For campaigns targeting audience segments with high social media susceptibility, particularly younger demographics on Instagram and TikTok as identified by Rethaber et al. [28], the CTEM trust estimation module should be prioritised as these segments show the strongest response to influencer credibility cues.

Comprehensively, the suggested model is a viable and scalable solution for real-world applications, ensuring that businesses can maximize their marketing investments, improve campaign performance, and create better trust-based relationships with consumers over time.

5. Conclusions

Overall, our findings indicate that the potential of AI applications in transforming the influencer marketing approach from gut-feeling-based to data-driven is rather tremendous. TIIN provides a way to address these complex interactions with AI. This framework unites long-term brand effect, strategic selection of influencers, and consumer trust forecasts in one architectural approach. The benefits of hybrid methods of learning are evidenced by the experimental results: the proposed model consistently achieves the best predictive performance on both traditional and standalone methods. The outcomes further reveal that, under the present scenario, engagement, reach, and trust are the key elements that contribute to success in marketing results. The ability to measure marketing efficacy and sustainability more appropriately can be achieved when brands switch their intuition-based to data-driven tactics by connecting to AI-powered analytics. The proposed framework is essentially a scalable means of enhancing consumer confidence, campaign success, and long-term brand building in the ever-evolving environment of digital marketing.

Future research should explore the integration of explicit consumer preference signals derived from AI-driven social media analytics [25] directly into the TIIN feature set to further enrich trust and ROI predictions. Additionally, situating the TIIN deployment within structured AI governance protocols—addressing standards, accountability, and cybersecurity [29,30]—will be essential as the framework scales to multi-platform and enterprise-grade campaign environments. Incorporating platform affordance variables identified by Wang et al. [27]—such as product visibility scores, synchronous engagement ratings, and social presence indices—as additional input features to CTEM and ISOM would also allow the model to capture the mechanism by which platforms mediate trust and information seeking, not merely their outcome correlates. Future datasets should further include the content-quality dimensions (visual aesthetics, denotative clarity, social satisfaction) validated by Yao et al. [11] and the sustainability-oriented behavioural outcomes (green purchase intention, environmental activism) synthesised by Kilumile and John [14], which represent high-value but currently underrepresented dimensions of influencer marketing ROI.