A multi-layer AI decision support system for startup success prediction and risk assessment using knowledge graphs and federated learning

Abstract

Since start-ups have grown so quickly in recent decades, it is more important than ever to determine what elements contribute to their success or failure. Numerous factors, such as market conditions, product differentiation, finance availability, and managerial methods, influence these results. However, precise forecasting is a constant issue due to the intricacy of business ecosystems and the interaction of non-financial and financial aspects. A multi-layer AI-driven prediction model that incorporates early-stage start-ups’ financial and non-financial characteristics is presented in this paper. A Graph Convolutional Network (GCN) creates feature embeddings at Layer 1, whereas a knowledge graph documents the connections between affecting factors. Federated learning is used to safely combine dispersed knowledge while maintaining privacy. Layer 2 uses a deep neural network (DNN) to forecast success or failure based on the fused features. Layer 3 offers risk assessment and interpretability, detecting survival variables including team dynamics and product differentiation. An experimental sample of 20 start-ups was thoroughly examined as part of the model’s evaluation on a dataset collected from Crunchbase that included approximately 623,000 companies, 799,000 founders, and 227,000 funding events. The findings show increased forecast accuracy and emphasize the value of non-financial elements in addition to conventional financial measurements. Through the integration of sophisticated AI approaches with organized domain knowledge, our work connects theoretical frameworks with empirical data. The suggested model contributes to start-up research and the real-world implementation of AI in business analytics by offering entrepreneurs, investors, and legislators a strong decision-support tool.

Introduction

Startups are usually newly formed business models that are linked to great scalability and disruptive innovation. They are frequently seen as potent catalysts for social and economic advancement¹. Due to their significant influence on innovation and technical advancement, as well as their role as a generator of wealth and economic prosperity, startups have emerged as a significant topic in the economic policy of emerging as well as developed nations worldwide. As more universities, governments, and private businesses invest in and encourage people to explore their ideas through these initiatives, startups are flourishing worldwide. The catch is that start-ups have a 90% chance of failing², meaning that many investments will not yield the expected rewards. 90% of start-ups launched in 2023 fail, according to SPA load². There are many entrepreneurs that fail, and it’s critical to determine the elements that contribute to both success and failure. When those elements are discussed and investigated, they will help the ecosystem’s prospective business owners create their successful strategy.

Entrepreneurial failure³ has repercussions that go beyond the startup and affect the economy and employment. The capacity to forecast success is a significant competitive advantage for a number of parties, namely venture capitalists seeking to make investments, since first-rate targets represent industries which have the potential to grow rapidly in the near future. This ultimately enables investors to stay one step ahead of the competition⁴. Investors can determine whether or not to invest in a start-up by using the prediction of start-up success. Machine learning algorithms are now thought to be a useful method for forecasting start-up success. Additionally, deep learning has a lot of potential for start-up prediction in particular and the business domain in general⁵.

Machine learning as well as deep learning, which use algorithms to create models that find patterns in data, provide businesses with insights and predictions to enhance operations, connect with customers, and solve other issues. There are many different algorithms available. These aid in forecasting whether a start-up will turn a profit or not. Examination of difficult real-world issues, such multi-criteria Making decisions and select the best option from a wide range of options calls on a variety of abilities and data analysis techniques⁶. Researchers are interested in and using knowledge graphs because they can combine enterprise data, event data, linked subjects, and other information to create visual enterprise knowledge. Numerous scholars have examined the research issues, methodologies, and state of knowledge graphs in recent years from a variety of angles, including graph building⁷. These models give investors, entrepreneurs, and legislators useful information and allow for more accurate survival forecasts. Though AI-based methods have shown promise, their efficacy depends on the incorporation of theoretical frameworks that take into consideration the dynamic and multifaceted nature of commercial ecosystems as well as on the availability of reliable statistics.

Prior research has mostly concentrated on basic machine learning and deep learning techniques, which frequently rely on sparse time series studies or little financial data^8,9. In particular, the majority of models concentrate on forecasting the future performance of startups using single-dimensional data, such as market trends or previous financial indicators. Although these models occasionally capture the development potential of startups, they frequently ignore important aspects of the larger ecosystem in which they operate, particularly the relationships between businesses that are cooperative and competitive. Startups usually work in resource-constrained contexts, and how they interact with one another greatly influences whether they succeed or fail. Conventional approaches have not made the most of this relational data, which has resulted in subpar startup success prediction, especially when there is sparse data and weak signal strength. In such circumstances, these models find it difficult to accurately detect early success signs. While there are many possible development objectives for startups, going public or being successfully acquired by a large corporation are the most desirable. We try to forecast what will happen to the company next, even though many research and venture capitalists concentrate on forecasting whether a company will successfully go public as a binary classification task^1,8.

Research gaps identified

The majority of current research on forecasting startup financing success involves statistical and regression models and focuses on structured data, including financial factors and characteristics of the project. Even though these techniques show some linear correlations, they are severely limited when dealing with high-dimensional data and complicated nonlinear aspects. Furthermore, these methods frequently ignore how well startups do in terms of market feedback and external information dissemination, especially the potential impact of social media as a vital conduit for contemporary data interchange.

Argaw et al.¹⁰ described that the majority of current models place an unnecessary importance on financial metrics, with very few taking into account non-financial aspects like company culture or team dynamics. Even if financial data is simpler to measure, there is a lacking of rich financial histories in many early-stage firms. Bidgoli et al.¹¹ finds the risk of startups by identifying the features such as staff count and LinkedIn followers using SHAP. These studies improve model prediction, but they frequently overlook more complex non-numerical factors like product differentiation. Maarouf et al.¹² demonstrated the harness language by correctly predicting outcomes using a fused LLM on Crunchbase database. To increase predictive power, Ozince & Ihlamur¹³ used LLM prompting to collect founder traits. Even while these employ sophisticated language models, they frequently do not combine statistical learning and symbolic reasoning, nor do they have organized frameworks like knowledge graphs. In Font-Cot et al.¹⁴ presented a multivariate model that emphasized team and product differentiation by fusing Lipschitz extensions using neural nets and linear regression. Few researches completely integrate conceptual modeling (e.g., KGs) with hybrid neural-symbolic architecture in a layered design, despite the existence of some multivariate and empirical frameworks. KGs are infrequently used in current startup survival models to provide explainable DNN risk assessment or to depict relationships (e.g., between founders, market, and product).

A multi-layer AI-driven decision support system is proposed that integrates empirical data and conceptual knowledge to closes these gaps. The contribution of this study is stated as follows:

• The first layer generates a knowledge graph to show the complex connections between components such as team experience, market dynamics, and product attributes and GCN to extract the factors influencing the startup risk and success. To illustrate the relationships between important startup characteristics, a knowledge graph is created utilizing both structured (such as funding data and founder profiles) and unstructured (such as firm descriptions and industry news) data. Semantic enrichment and feature extraction that captures latent connections and contextual relevance are made possible by the graph.

Major contribution of this study

• In order to generate a startup vector embedding that represents multidimensional success potential, the GCN will go through the neighborhood of a startup node, including data from connected investors, markets, and founders. The downstream prediction layers get these embeddings as feature vectors. FL securely trains a global startup prediction model by distributing GCN output across clients.

• This symbolic embeddings from GCN is included into a deep neural network (DNN) in the second layer, which uses feature embeddings to predict whether a startup will succeed or fail.

• The third layer performs a risk assessment and evaluation using the outcome of layer 1 and 2, enabling a comprehensive risk factor analysis. It determines each startup’s strengths and shortcomings and calculates the risk involved.

• By integrating financial and non-financial factors and statistical learning, the proposed model offers a thorough and understandable approach to predicting the startup success and failure in an efficient way.

• By improving forecast accuracy and providing insightful information to investors, entrepreneurs, and policymakers, this approach fills a major research gaps.

Novelty of this study

FL and GCN have both been used in different fields in existing researches. Our approach is unusual, though, because we have adapted and integrated them for startup forecasting, which presents different obstacles than other industries. Specifically:

Integration of a domain-driven knowledge graph with GCN: In contrast to general applications, our model clearly captures the relationships between various startup elements (e.g., market conditions, funding events, founder-company links, and team dynamics) using a knowledge graph. In order to learn embeddings from this structured relational data—which are crucial for survival prediction—the GCN is then modified. FL, or federated learning, for startup environments: Inherently distributed and frequently sensitive, startup data raises confidentiality issues for institutions, investors, and founders. To enable efficient training across decentralized data sources without sacrificing privacy, we modified FL to fit this situation by adding safe aggregation and improving communication through selective parameter updates.

In order to improve transparency and decision assistance for stakeholders, our model goes beyond prediction by incorporating a risk appraisal layer that analyzes survival determinants (such as product differentiation and team composition).

Paper structure

The paper is structured as follows: “Related work” section provides the review of the literatures, “Methodologies” section describes the data used with the proposed multi layer DSS for the prediction of startups success or failure using financial and non-financial factors. “Results and discussions” section discusses the experimented results of the proposed model with the identified risk for the failure of startups. “Discussions” section discussed the strength and implications of the proposed model contrast to the existing works. “Conclusion” section concludes the proposed model outcome with the future directions.

Related work

Cot et al.¹⁵ created a multivariate AI-driven model that uses linear regression, neural networks, and Lipschitz extensions to forecast startup survival. Important conclusions emphasize how crucial team dynamics and product uniqueness are in predicting survival odds. The study fills in gaps in the body of literature by fusing conceptual understanding with actual data and provides entrepreneurs, investors, and policymakers with a useful tool for making decisions. Sun et al.¹⁶ presented a revolutionary approach for startup evaluation based on advances in artificial intelligence (AI). Neural networks and clustering algorithms are two examples of machine learning approaches that make it easier to analyze complicated, nonlinear interactions and provide insightful information about startup paths.

Fuentes-Callés¹⁷ evaluated the predictive power of Gradient Boosting, Cox proportional hazards methods, and Random Survival Forests in estimating the length of beginning business activities, discovering that Gradient Boosting offered the best predictive ability. Bidgoli et al.¹⁸ proposed a prediction system using machine learning techniques in order to find profitable businesses for investors. In order to develop the predicting solution and suggest policies, clustering and classification algorithms have been utilized to predict startup success and perform feature importance analysis based on the SHAP and permutation techniques. The performance of four classification algorithms such as Random Forest, Gradient Boost, Multilayer Perceptron, Logistic Regression, and Support Vector Machine is then compared in order to predict business success. The algorithms Random Forest & Gradient Boosting had the highest accuracy, at 82% and 80%, respectively.

Using logistic regression to forecast startup survival, Ghassemi et al.¹⁹ discovered that teams with a variety of backgrounds and a focus on niche markets have a higher chance of success. Their study’s findings demonstrated that in a competitive market, ideas aimed at existing markets have a lower chance of surviving, while teams with a variety of professional and academic backgrounds had a higher chance of surviving. Using the Crunchbase dataset, Corea et al.²⁰ examined venture capital investment choices and found 21 pertinent characteristics for assessing startup performance. Their study highlighted the significance of investment amount, prior investors and their reputations, and social media presence. In a related study, six classifier algorithms were employed as ML models to predict startup success and identify the critical characteristics that are important for generating such forecasts. Crunchbase information organizations make up their dataset (Kim et al.⁴. They employed importance of feature to forecast startup success, and their findings showed that key factors influencing company success include financial backing, industry association level, media exposure, and industry convergence level.

Qiu et al.²¹ combines the company-related social media news data from Twitter with financial data from the Crunchbase platform. In order to capture public sentiments and expectations regarding the company, sentiment analysis is performed on the social media messages using the BERTweet model, which extracts sentiment variables including polarity and intensity. A deep neural network (DNN) is then used to build a decision support system for forecasting financing success by combining financial indicators, social media numerical elements, and sentiment features. According to experimental findings, the DSS that uses social media data performs noticeably better in terms of prediction accuracy than conventional decision support systems. Sentiment elements also improve the model’s capacity to determine a company’s financing performance. Jinze Li²² created a model to categorize startups and examine the critical elements that contribute to their success. The key characteristics that hinder the development of startup businesses were investigated, and some of the characteristics were explained, using Random Forest and Support Vector Machine. Additionally, it contrasted the outcomes of several machine learning techniques, demonstrating how well these techniques work for investors in forecasting business success.

Methodologies

This section discusses the dataset used for this study and proposed a model for predicting the startups success or failure.

Data used

Our data sample was taken from Crunchbase’s August 2018 Daily CSV Export. The complete dataset includes 227,172 tuples describing fundraising round events (mostly from the last 20 years), information about 799,446 firm founders, and 623,232 companies²³.Given the significant failure rate in the early years, for instance, it was roughly 44% in the US at four years and these windows are deemed sufficient for a company²⁴. Research by Arroyo et al. served as the foundation for our data selection process. Initially, startups founded between March 2014 and March 2018 were taken into account. Startups that have advanced to the next level of the C series and have been purchased, closed, or listed on an initial public offering (IPO) during this time have not been included. because early-stage investments don’t appeal to them.

Significant changes in commercial ecosystems are captured during this time, providing a wealth of data for examining survival paths. The authors’ 10 years of working with startups, which gave them exclusive access to context-rich data from both successful and unsuccessful endeavors, influenced the data collection approach.All responses were gathered using structured questionnaires to reduce the possibility of subjective bias. Semi-structured interviews were added, and the results were cross-validated against publically accessible data to guarantee consistency and dependability. Startup managers were given structured questionnaires with 81 items that focused on essential survival traits. Semi-structured interviews with investors, entrepreneurs, and other stakeholders were added to these to provide more qualitative depth. While secondary data, such as financial records, fund-raising histories, and reports on the market ensured a comprehensive perspective, observations made during startup operations offered additional insights. This combination of data sources improves contextual relevance and reliability while adhering to standard practices in startup research. Finding the startups that have the potential to succeed and those who have already done so was the next stage. Investors’ ultimate goal is an IPO. Nevertheless, it is uncommon for businesses to reach the initial public offering (IPO) stage, with 30% of seed-funded businesses exiting through an IPO²⁵. Investors look for businesses that will advance toward new funding infusions and possibly better outcomes. Additionally, they seek for businesses that will invest rapidly in order to grow more swiftly on their own²³. For a more thorough assessment of potential investments, we therefore took into account both partially successful and acquired startups in addition to successful startups. Successful startups are those that made it to the initial public offering (IPO) stage between March 2018 and March 2021. Additionally, businesses who have raised additional capital throughout this time frame are regarded as partially successful start-ups.

The collected attributes are based on five dimensions such as team-dynamics, financial metrics, and conditions of market, service and product differentiation and the vision of strategic. Qualitative responses were transformed into statistical scores using a systematic coding methodology to guarantee the data’s quality and reproducibility. Throughout the procedure, logs were kept that detailed the nature and provenance of each feature. On a scale of 0.0 to 1.0, survival outcomes were represented, where 1.0 denoted completely successful companies that showed consistent growth and market presence. Firms that experienced difficulties with financial stability or market adaptation and had only partial success were given values below 1.0, whilst failing firms that completely stopped operations were given a rating of 0.0.

Model explanation

Figure 1 illustrates the model structure of the proposed multi layer decision support system for predicting the startup survival based on the risk and success. A multi-layer decision support system designed for risk assessment and startup success prediction. As shown in the Fig. 1, the proposed framework combines Federated Learning (FL), Graph Convolution Networks (GCN), and Knowledge Graphs (KG). The process begins with preprocessing the input data, which includes collecting and classifying startup-related financial and non-financial data. A KG is then constructed to identify correlations and semantic links among elements of financial and non-financial factors of startups. A GCN receives this improved representation and uses it to train multidimensional embeddings for each startup by aggregating data from its graph-based neighborhood. These embeddings indicate the startup’s overall success. To preserve data privacy and promote inter-institutional collaboration, a global predictive model is trained among scattered clients via FL.

Fig. 1.

Overview of proposed multi layer decision support system to predict startup survival.

Although we recognize that federated learning (FL) and graph convolutional networks (GCNs) have been used in other fields, our work is innovative in the way these techniques are modified, expanded, and contextualized for the startup forecasting challenge. Startups face particular difficulties, including a lack of organized datasets, confidentiality restrictions, and extremely varied data (finance indicators, team structures, market situations, and product features). Our approach goes beyond a straightforward GCN and FL combo to solve these issues by including a number of methodological improvements which is shown in Table 1. By adapting GCN and FL to the features of startup ecosystems, we are able to go beyond their conventional use cases and provide both methodological originality and usefulness. These improvements allow the model to produce actionable insights for practical decision-making in the entrepreneurial domain in addition to producing accurate predictions.

Table 1.

Comparison of existing GCN + FL applications with the proposed startup forecasting model, highlighting domain-specific adaptations and contributions.

Aspect	Existing GCN + FL applications	Our adaptations for startup forecasting
Domain of Application	Mostly in social networks, IoT, and healthcare.	Startup ecosystems (businesses, founders, financing occasions, and market circumstances)
Input Data	Homogeneous graphs, such as sensor networks and user-item networks.	Heterogeneous knowledge graph connecting industries, financing rounds, enterprises, and founders
GCN Role	Extraction of generic features from graph architectures	Domain-driven extraction of embeddings from relational structures unique to startups (funding dependence, product differentiation, team dynamics)
FL Objective	Privacy-related distributed training.	Cross-institutional startup data training that addresses financial and business data confidentiality
FL Optimization	FedAvg, or standard aggregation	To minimize communication overhead and preserve privacy, use secure aggregation in conjunction with selective parameter changes
Output	Tasks involving classification or prediction (e.g., traffic forecasting, illness risk).	A multi-layered approach that includes: (1) startup survival prediction, (2) risk assessment, and (3) critical factor explanation.
Contribution to Field	Technical settings are used to test algorithms	Business analytics decision-support tool for investors, entrepreneurs, and legislators.

Layer 1: correlation among influencing factors and its extraction using KG-GCNFL

• (i)Preprocessing to handling missing values

The quality of features embeddings in the Graph Convolutional Network (GCN) model may be impacted by the missing values we found in the history records of start-up businesses during our investigation. This study used two imputation methods such as mean imputation and K-Nearest Neighbors (KNN) imputation—to solve this problem. In mean imputation, the mean of the observed values for that feature is used to fill in the missing values. When the data are Missing Completely At Random (MCAR) and the percentage of missing data is low, this approach is simple and efficient. However, if the missing data are not MCAR, it may induce bias and decrease variability²⁶. In contrast, KNN imputation takes into account the similarity between data points and guesses missing values using the values of the ‘k’ nearest neighbors. Because it maintains the underlying data distribution and interactions among features²⁷, this approach is more reliable, particularly when the data are Missing At Random (MAR). We improved the GCN model’s performance and dependability by using these imputation techniques to make sure the feature embeddings input into the model were comprehensive and representational of the underlying data.

• (ii)Startup Knowledge Graph (SKG)

Google was the first to propose the knowledge graph, which is essentially a semantic network that describes objective items and their interactions in the form of graphs, exposing the relationships between entities. The network’s edges are made up of the relationships between the entities, which can be expressed as (head entity, relationship, tail item) triplets. The entities that exist in the objective world are equivalent to the entities in the network.

Given a graph G={V, E}²⁸, where V denote collection of nodes, E denote the set of edges and it is extended for KG as,

1

where V is the set of nodes which is categorized into types called such as founders, investors, markets, products, competitors and so on. is the collection of edges denotes the relationship among the nodes which includes founders has an experience with industry, startup has funding from investor and founder worked at the successful company and so on. R denotes the types of relationship and describes its nature of interactions. is a function which assigns the relationship type r to R of each edge e and is a function assign a type T to each node v. Figure 2 illustrates the KG for the considered data. The entities of the graphs are {Startup, Investor, Founder, Product, Funding Round, Market, Technology, Failure, Success} and their relationships are denoted in edges which includes in Table 2. While manual intervention is mainly restricted to initial schema design and periodic validation, automated data ingestion pipelines that update entities and relationships from external startup databases and funding records can maintain the knowledge graph in operational settings. The knowledge graph format allows for the insertion of region-specific traits or entities to capture local ecosystem characteristics while maintaining the basic entities and linkages across various geographic startup ecosystems.

Fig. 2.

Knowledge graph of startup success and risk prediction using the proposed multi-layer decision support system (MLDSS).

Table 2.

KG relationships among startup factors.

Edges	Relationship
Founder→Startup	Founded
Startup→Product	Develops
Startup→Market	Targets
Startup→Fundinground	Raises
Fundinground→Investor	Funded by
Startup→Technology	Uses
Startup→Success	Leads to
Startup→Failure	Leads to
Founder→technology	Experienced in
Investor→Market	Interested in

• (iii)GCN-FL

When it comes to handling spatial relationships between things, GCNs²⁹ have shown remarkable performance. In order to address the spatial relationships between nodes, they seek to combine the data of nodes and their neighboring nodes. Because of the influence of nearby and distant points, the closer the neighboring points are, the more of an impact they have and each node modifies its state until the ultimate equilibrium. The adjacency matrix contains this pertinent data. The GCN used here is for feature extraction. Suppose, we have the batch of graph that contains N nodes, and each nodes has its own features, that form the matrix , where k is the index number. The nodes relationship forms the adjacency matrix . The GCN aggregates the feature information to form the local neighborhood as,

2

where D is the degree matrix of A which is the adjacency matrix, is the intermediate hidden embeddings ateach layer l which is denoted as (H⁰ is X), W is the weight matrix of layer l. is the ReLU activation function. In this Z = H^(l) is the final node embeddings to represent the potential of startup success.

Figure 3 shows the GCN-FL framework that comprised of multiple SKG for decentralized learning without required the raw or private data. This model comprised of various SKG’s, each denotes the relationship of various factors of startups. To prevent direct sharing of data and considering the privacy and security, FL framework is incorporated to facilitate the global model training.

Fig. 3.

The devloped GCN-FL framework.

Let there are C companies, each with local data set D_c, and local processed embeddings using GCN called Z_c. Each startup trains it owns local model weight as W_c using the embeddings and labels to reduce the loss as:

3

where is the output of local prediction, Z is the embedding of feature i, Y is the output of predicted probability of two nodes and l is the loss function. In order to ensure that customers with more data contribute proportionately to the global model, we specifically use weighted aggregation based on the amount of local samples per client. To lessen distributional discrepancies prior to aggregation, we also apply feature scaling and local normalization to every client. We periodically evaluate local models and dynamically modify aggregation weights to further limit performance loss caused by non-IID data. The central server of FL aggregates the update from all the companies local NN using FedAvg:

4

where is the global model and , . This aggregated model is further redistributed to find the best output until convergence. The aggregated features of both financial and non financial factors are listed in Table 3.

Table 3.

Fused features from GCN-FL.

Dimensions	Factors	Description	Data type
Team dynamics	Team Size	No. of team members	Integer
	Founder_Exp_year	Avg. experience of founder	Float
	Diversity_Score_Team	The score based on the background and skill	Float in the range 0 to 1
	Prev_Starup_Exits	No. of successful completion by the founders	Integer
	Tech_Background_Founder	Presence of technological background	Boolean (0/1)
Market conditions	Market_Size_Target	Estimated addressable market total	Float
	Density_Competitor	No. of starutps offer the same solution in the market	Integer
	Region_Code	Encoded Geo location	Categorical
	Trend_Score	Degree to which the startup meet the treand	Float (0 to 1)
	Risk Score_Regulatory	Based on regional risk or compliance concerns	Float (0 to 1)
Financial metrics	History of funding	Total amount raised by startup	Float
	Stage_Funding	Encoded stages as Seed, Series A and so on	Categorical
	Quality Score_Investor	Success score of investor from KG	Int
	Burn_Rate	Captial usage in a month of total funding	Float (0 to 1)
	Revenue_Model type	Subscription etc.,	Categorical
Product differentiation	Product stage	Level of development	Categorical
	Deep technology usage	Indicates AI/ML/ Blockchain usage	Boolean(0/1)
	Status_IP	Indicate the protection of patent or IP	Boolean (0/1)
	Tech_Complexity	Computed score of complexity of used platforms	Float (0–1)
Vision of strategic	Cosine similarity to startup success	NLP for startup success visions	Float (0 to 1)
	KG_Influencesocre	Score computed from KG-GCN neighbourd features	Float (0 to 1)
	Connected_success_startups	Graph linked with successful outcomes	Int

We carried out a comparative analysis with a non-FL centralized configuration in order to confirm the efficacy of the FL framework. The Results section reports the specific performance differences in terms of accuracy, F1-score, training duration, and communication overhead.

Layer 2: prediction using DNN

The unified feature vector from the GCN-FL is fed as input to DNN for success or failure prediction since it serve as the major predictive factor of proposed DSS. Deep-level interactions among heterogeneous features are extracted by the DNN’s several nonlinear transformation layers. The last layer produces a binary classification result that shows whether a startup’s funding round is expected to be successful or unsuccessful. The DNN receives the output Z of the embeddings from GCN and outputs the probability as .

Let, is the input matrix from GCN, w^l and b^l are the weight and bias of layer l and L is the number of layers in DNN. The prediction from DNN is expressed as,

5

Final output as success or risk is predicted as,

6

where OL is the output layer value. The framework uses optimization techniques such the Binary Cross-Entropy Loss as a goal function to further improve the forecast dependability of the system. This loss function ensures reliable model training for practical use in decision support scenarios by efficiently measuring the difference between expected probability and actual outcomes.

7

where is the probability of failure (risk score) among the value 0 and 1. This score denotes the startup success or failure of proposed model confidence. A high value of 1 denotes the success and low value 0 denotes the risk of failure. A grid search approach in conjunction with 5-fold cross-validation was used to refine the DNN model in order to guarantee optimal performance and reproducibility. The number of hidden layers (2–5), the number of neurons per layer (64–512), the optimizers (SGD, Adam), the activation functions (ReLU, tanh, sigmoid), and the learning rates (0.0001–0.01) were among the configurations that were examined. Three hidden layers with 256, 128 and 64 neurons each, followed by a ReLU activation and a dropout rate of 0.3 to avoid overfitting, made up the final design. With a learning rate of 0.001 and categorical cross-entropy as the loss function, the Adam optimizer was used to train the model.

Qualitative responses from semi-structured interviews and structured questionnaires were quantified using a predetermined coding methodology in order to methodically include non-financial aspects in the predictive model. Standardized criteria were used to categorize or assign numerical values to each response. For example, open-ended interview inputs were thematically examined and transformed into categories or numerical representations, while team experience, market involvement, and product distinction were categorized into ordinal or binary scales. For thorough startup success and risk prediction, non-financial attributes can be combined with financial measurements using this method, which guarantees consistency and repeatability throughout the dataset (see Table 4 for the coding scheme).

Table 4.

Coding scheme for non-financial startup factors used in the predictive model.

Non-financial factor	Example questionnaire/interview response	Coding protocol / numerical value
Team experience	“Founders have prior startup experience”	1 = Yes, 0 = No
Team size	“Small team of 3 people”	1 = 1–5 members, 2 = 6–10, 3 = 11+
Market engagement	“Regular customer surveys and feedback loops”	0 = None, 1 = Occasional, 2 = Regular
Product differentiation	“Unique product feature compared to competitors”	0 = No, 1 = Yes
Founder-investor network	“Active network of mentors and investors”	0 = Weak, 1 = Moderate, 2 = Strong
Strategic planning	“Formal business plan updated quarterly”	0 = No, 1 = Yes

Risk assessment

The risk identification and assessment are the third layer of the proposed Multi layer DSS system. Within the startups, the risks are need to be identified which is stated in Table 3. The identified risk is analyzed level of risk severity is computation shown as follows:

8

where is the impact of occurrence of risk (1-negligible, 10-catastropic), Occurrence denotes the likelihood of the occurrence of risk (1- rare, 10- frequently) and Detection is the ease of identifying the risk (1- very likely, 10-hard to detect).

Results and discussions

This section discusses the empirical results of this study based on the considered five dimensions and its risk factors. Every experiment was carried out on a workstation that has an Intel Core i9-12900 K CPU, 64 GB of RAM, and an NVIDIA RTX 3090 GPU with 24 GB of VRAM. PyTorch was used to implement the models in Python, and mixed-precision training was enabled to maximize computational efficiency and memory use. For the whole model of GCN + DNN + federated learning: For the entire dataset, the average training time per epoch was about 45 s. 85–95% of the GPU is used for both forward and backward passes. The results integrate the computational parts of considered models such as KG, GCN-FL and DNN to understand the startups success and risk. Our strategy offers a cohesive framework for methodologically addressing the startup survival issue, improving and concretizing current approaches and adding new instruments to increase predictive accuracy. This section discusses the results of the proposed methods adoption to various contexts. While medium-probability firms showed modest qualities but frequently lacked innovation or differentiation, high survival probability startups regularly had robust team dynamics and compelling strategic visions, as seen in Fig. 4. Financial stability and market adaptability were issues for low-probability businesses, highlighting the importance of strategic emphasis and resource allocation. These results are consistent with the IBM Triangular Framework, which predicts success by combining market adaptability, business characteristics, and investment (see, for instance³⁰.

Fig. 4.

Histogram showing the success ratio of the considered sample startups.

Impact on other ML approaches

The suggested three models in this study such as Knowledge graph, GCN-FL and DNN are applied to find the success or failure of the startups. Each model provides the unique insights to the considered startup datasets. Knowledge graph finds the relation among the factors that influences the startups and GCN-FL executed to find the contextual representations of the factors and FL ensures the security of the system by finding the global model by aggregating the local model without sharing the actual data. DNN has been applied to predict the success and failure of the startups. The combination of these model provides a multi layer decision support system for the analysis of startups survival as illustrated in Fig. 5.The success/ risk rate prediction of the proposed model is compared with the existing ML approaches such as random forest, extreme gradient boosting and artificial neural networks³¹. The industries taken into consideration in the study, such as FinTech, Health, E-Commerce, EdTech, and FoodTech, are displayed on the x-axis. As demonstrated, the suggested KG-GCNFL model outperforms alternative methods in capturing the complex patterns of success and failure, closely adhering to the ground truth across industries. This comparison shows how well knowledge graphs, graph convolutional networks, and federated learning can be combined to predict startup outcomes. It also shows how resilient the model is in a variety of industry scenarios. The ground truth point is similar with the proposed model since it predicts the score with improved accuracy compared to other considered approaches.

Fig. 5.

Success and failure rate assessment of the dataset using the considered approaches.

Performance of the suggested framework in terms of metrics

The input dataset is split into a training dataset (80%) and a test dataset (20%) at random to assess how well the algorithms function. We may evaluate the efficacy of the trained algorithms by comparing their predictions on the test dataset once the data has been divided into training and test datasets. This enables us to ascertain whether the algorithms can produce precise forecasts based on data that has not yet been observed. Figure 6 shows the confusion matrix of the success and failure prediction of startups using proposed multi layer DSS. The model correctly classified 13 success startups and 5 failure startups while it is misclassified the 1 failure startups as success and 1 success startups as failure startups. Based on the matrix, the obtained metrics results are accuracy 93%, precision 94%, recall 83% and F1 score 93%. This model evaluates that the proposed system is effective on identifying the successful startups with improved accuracy and confidence. To evaluate the discriminatory power of the developed model, the RoC is analyzed as shown in Fig. 7. The obtained RoC value is 0.89 that indicates the improved level of separability. It means, the model has 93% of chances of correctly distinguish among the randomly chosen startups success and failure.

Fig. 6.

Performance of the proposed model which shows confusion matrix, illustrating the distribution of correct and incorrect predictions across classes.

Fig. 7.

ROC curve for startup success and failure prediction.

The suggested model’s prediction performance improves with the larger dataset. The accuracy of the test set is 93%, and its AUC is 0.92. 95% confidence intervals in order to characterize uncertainty. Additionally, we used paired t-tests to compare the suggested model with several baseline approaches, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), and a conventional Artificial Neural Network (ANN), using an expanded dataset. Table 5shows that the findings of suggested model continuously beats the baseline, with increases in accuracy and AUC being statistically significant (p < 0.05).

Table 5.

Performance comparison with multiple baselines with statistical significance.

Model	Accuracy (%) (CI – 95%)	AUC (CI-95%)	p-value vs. proposed model
Ground truth	100	1.00	–
Random Forest (RF)	89.5 [87.0–91.8]	0.88 [0.85–0.91]	< 0.05
Extreme Gradient Boosting (XGBoost)	90.2 [87.8–92.5	0.89 [0.86–0.92]	< 0.05
Artificial Neural Network (ANN)	91.1 [88.7–93.2]	0.90 [0.87–0.93]	< 0.05
GCN without KG	88 [84.7–91.2]	0.83 [0.82–0.9]	< 0.05
TransE (KG embedding)	85 [83–88.2]	0.84 [0.81–0.86]	< 0.05
Proposed model	93.0 [90.9–95.1	0.92 [0.89–0.95	–

We used a GCN-FL-DNN Classifier to do a feature importance analysis as shown in Fig. 8 to determine the relative relevance of the several elements that contribute to the success or failure of companies. The suggested GCN-FL-DNN pipeline was used to process the input dataset, which contained important features that were found by building a domain-specific knowledge graph. We used a stand-alone proposed model trained on the same structured feature set to extract the feature importance for interpretability and comparative analysis. “Funding amount” was one of the most significant indicators, according to the training multi layer DSS model, frequently exceeding or ranking closely to traditional measures like money and team experience. This implies that even in situations where the initial funding or market trend alignment is moderate, companies with a well-defined long-term plan and vision have a higher chance of success. A horizontal bar chart was used to display the feature importance scores, which were obtained from the model’s internal Gini impurity reduction measures. The data backs up our theory that early startup longevity is significantly influenced by non-financial and quantitative indicators, especially those pertaining to strategic aim.

Fig. 8.

Feature importance based on GCN-FL-DNN model.

We evaluated the suggested FL framework’s efficacy by contrasting its results with those of a conventional non-FL centralized approach. In addition to guaranteeing data security, the FL-based model produced somewhat superior prediction results, as indicated in Table 6. In particular, FL maintained an acceptable training duration and communication cost while achieving higher accuracy and F1-scores than the non-FL variant. This gain is attributable to FL’s capacity to take advantage of various client data distributions, which improves generalization and resilience.By securely aggregating distributed data, the FL framework outperforms the non-FL configuration while maintaining privacy and improving prediction accuracy, albeit introducing a little communication expense (~ 12 MB per round). Practically speaking, even investors or incubators with modest computational resources may manage the computational and communication overhead of the suggested federated learning system. Instead of using the complete raw dataset, each participating client simply uses preprocessed feature embeddings for local model training, which drastically lowers processing needs. On a typical workstation, the training time per epoch in our experimental setting was roughly 45 s. Additionally, the FL framework has a communication cost of about 12 MB every training round because it only transmits model parameters rather than raw data during aggregation. These design decisions guarantee that the suggested system may be practically implemented across several incubators or venture capital companies without the need for high-performance infrastructure.

Table 6.

Performance comparison between FL and Non-FL versions of the model.

Model variant	Accuracy (%)	F1-Score	Training time (epochs)	Communication overhead
Non-FL (Centralized)	92.8	0.91	100	None (0 MB)
FL (Proposed)	93.2	0.93	110	Moderate (12 MB per round)

Significant values are in [bold].

Impact on risk evaluation

The heatmap of the startup failure prediction using RpN is shown in Fig. 9. Based on this heatmap, the factors that have high risk on influencing the startup failure are computed and shown in Table 7. The severity, occurrence, and detectability components—all of which are particularly defined for start-ups—are multiplied to determine the RPN for each factor.

Fig. 9.

Heatmap of high risk factor identification for startup failure prediction.

Table 9.

Ablation study evaluating the contribution of individual components in the proposed start-up risk prediction model.

Scenario	Removed/modified component	Accuracy (%)	F1-score
Full Model	None (All layers active)	93	0.92
A	GCN embeddings replaced with raw features	86	0.85
B	Knowledge graph removed	88	0.87
C	Federated learning disabled (centralized training only)	90	0.89
D	DNN replaced with a simple logistic regression classifier	82	0.81

Expert topic knowledge and historical data from Crunchbase are used to determine the factor’s severity, which indicates the possible impact on the start-up’s survival or failure. For instance, a poor team diversity or a high founder turnover rate intensifies the severity.

The factor’s frequency or prevalence among the start-ups in the dataset provides the basis for occurrence. For example, early-stage companies are more likely to have a high burn rate or a competitive market density.Given the availability of both financial and non-financial indicators, detectability refers to how quickly the risk may be identified or reduced.The top five high risk factors of every startups includes Burn rate, Founder_Exp_year, Product stage, History of funding and Quality Score_Investor. The reason behind this risk is described in Table 8.

Table 8.

Description of high-risk factors affecting startups.

High risk factors	Reason
Burn rate	When a startup has a high burn rate, it is consuming money far more quickly than it is making money. Without consistent funding or sales growth, the business runs the risk of rapidly running out of cash, which is one of the main reasons why startups fail.
Founder_Exp_year	Inexperienced founders frequently lack the network, operational skills, and strategic vision necessary to grow a firm. They might make preventable errors in market entry, product timing, employment, or fundraising.
Product stage	It is extremely difficult for a product to stand out in crowded markets if it has distinctive value propositions. It frequently results in low investment interest and poor user retention
History of funding	Hiring, product development, marketing, and expansion are all hampered by inadequate capital. Startups may have to quickly change course or fail due to a lack of funding.
Quality Score_Investor	Poor investors might not have the networks, domain knowledge, or strategic value necessary to mentor firms. This reduces market credibility and future funding availability.

Table 7.

High-risk factors contributing to startup failure.

Dimensions	Factors	Risk Description	RpN
Team Dynamics	Founder_Exp_year	Avg. experience of founder	393
	Diversity_Score_Team	The score based on the background and skill	253
	Prev_Starup_Exits	No. of successful completion by the founders	254
Market conditions	Market_Size_Target	Estimated addressable market total	217
	Density_Competitor	No. of starutps offer the same solution in the market	181
	Risk Score_Regulatory	Based on regional risk or compliance concerns	211
Financial Metrics	History of funding	Total amount raised by startup	337
	Quality Score_Investor	Success score of investor from KG	295
	Burn_Rate	Captial usage in a month of total funding	505
	Revenue_Model type	Subscription etc.,	98
Product differentiation	Product stage	Level of development	337
	Deep technology usage	Indicates AI/ML/ Blockchain usage
	Status_IP	Indicate the protection of patent or IP	100
Vision of strategic	Semantic similarity to success	NLP for startup success visions	141
	KG_Influencesocre	Score computed from KG-GCN neighbourd features	100
	Connected_success_startups	Graph linked with successful outcomes	95

Significant values are in [bold].

Discussions

The findings of this analysis highlights the versatility of the proposed multi layer DSS in evaluation the startup risk and success. This model addresses the research gaps from the literatures with the integration of empirical models and offers a tool for the entrepreneur ecosystems. This section elaborates the strength, limitation and implications of the developed frameworks.

This study’s time frame (2010–2024) encompasses a number of noteworthy macroeconomic occurrences that affected startup dynamics. Notably, the COVID-19 pandemic (2020–2022) significantly impacted startups in a variety of industries in terms of funding, market demand, and operational capacities. Many entrepreneurs had to quickly change their business models, deal with delayed investment rounds, and deal with changes in consumer behavior. In order to guarantee that the model represents actual circumstances and takes exogenous shocks into account, these factors were recorded in our dataset and taken into account during analysis. By taking into account these macroeconomic variances, the model’s predictive performance was assessed throughout times of crisis as well as normal market conditions, proving its resilience and suitability for a range of economic situations.

The effect of COVID-19 on startups in various industries is depicted in Fig. 10. The line graph illustrates the drop in fundraising rounds during the pandemic, while the stacked bars reflect the number of firms that changed their business plans and encountered operational difficulties. This graphic illustrates how startup operations and funding were impacted by macroeconomic shocks and shows how the predictive model takes these actual disruptions into account to provide reliable predictions of startup success and failure.

Fig. 10.

Impact of COVID-19 on startup operations and funding.

Ablation study

By methodically eliminating or altering important architectural components, we carried out an ablation study to comprehend the impact of each component of the suggested AI-driven start-up prediction model in Table 7. The Crunchbase-derived dataset is used in the study to assess the influence on prediction performance (accuracy). Each component makes a significant contribution to the model’s performance, as confirmed by the ablation investigation. To achieve the high 93% accuracy in start-up risk prediction, federated learning, deep neural network modeling, knowledge graph integration, and GCN-based embeddings are essential. The findings also support how crucial it is to record both non-financial and financial aspects of early-stage start-ups.

Strengths

The proposed KG-GCN-GL based Multi layer DSS to address the challenge of startup risk or success prediction and it has the strength as follows:

• Compare to the traditional approaches, the proposed framework considers both quantitative factors related to finance and qualitative factors such as team and vision that shows the complete picture about the startups.

• With the implementation of KG, the relationship among the startup attributes are structured that are effectively encodes the knowledge about the startups domain. GCN enables the relational pattern learning of the features.

• The FL ensures the data security that provides collaborative training by without sharing the raw data which will enhances the system scalability and decentralization that is important for the initial stage of investments.

• The model obtained the 93% of accuracy with AUC score of 0.9 shows the improved discrimination among the startup failure and success. The risk assessment section clearly explain the high risk factors and discusses the reason behind the risk of the concerned factors which will be helpful for the startups to concentrate on this factor carefully before they initiate. Among the considered five dimensions burn rate of the financial metrics have the high risk factor which decides the startup success or failure.

Limitations

Even though the framework demonstrates the novel framework with robust performance, it also has some limitations such as:

• The experiment considered 20 startups of small dataset. Although the performance of the model is promising, a larger data is needed to provide the stronger generalization.

• The classification of the model is binary that differentiate the success or failure of the startups. It would be better to consider the intermediate or pivoting stages of the startups.

• The inclusion of vision strategic introduces bias since it is encoded manually which affects the model consistency.

Implications

This study has several practical and theoretical implications. The study contributes to the conceptual understanding of how startup attributes’ structural and semantic linkages can affect predictive modeling by combining a knowledge graph with a GCN-FL-DNN architecture. This method shows that adding relational data (by knowledge graphs) along with distributed learning methods (via federated learning) can produce more reliable and context-aware insights than classic linear or tree-based models, which mainly rely on static financial indicators. The model emphasizes the theoretical significance of qualitative, non-financial characteristics, like team dynamics and strategic vision, in predicting company success. Additionally, in an era of decentralized analytics and data governance, the implementation of federated learning offers a new paradigm for creating data-driven entrepreneurial strategies that respect privacy and data ownership. This is conceptually significant. All things considered, the study adds to the growing body of research on hybrid AI frameworks that combine deep neural inference, privacy-preserving cooperation, and structural graph reasoning for predictive decision-making in socioeconomic domains. Predictions based on past data may be momentarily impacted by extreme macroeconomic events like financial crises or pandemics. Nonetheless, the suggested approach facilitates regular retraining and ongoing updates via the federated learning process, enabling the model to adjust to changing market conditions and new trends in the startup ecosystem.

The research conclusions have important practical implications for a variety of startup ecosystem participants, such as investors, incubators, accelerators, and legislators. The suggested methodology reduces dependence assessment by acting as a data-driven decision support tool that can accurately forecast startup success or failure. While the risk score emphasizes important elements contributing to probable failure, the predicted success probability shows the likelihood that a startup would achieve good outcomes like sustainable growth or successful funding development. In real-world applications, investors can view probability above 0.70 as having great potential, values between 0.40 and 0.70 as needing further research, and values below 0.40 as signifying a higher level of investment risk. Rather than being rigid standards for financing decisions, these thresholds function as flexible decision-support recommendations.

The methodology can accelerate screening for investors and venture capital organizations by spotting high-potential startups early on, especially by assessing team capabilities and strategic vision, two aspects that quantitative models sometimes ignore. The tool can be used by investors and accelerators to customize support programs according to the risk profiles that the model produces. The incorporation of federated learning guarantees that sensitive startup data may be evaluated collaboratively across organizations without compromising data privacy, making the system scalable and consistent with data protection rules. With only minor infrastructure changes, the suggested multi-layer DSS can be practically implemented into current venture capital or accelerator review procedures. Through API-based interfaces, the system can function as a modular analytics layer linked to assessment dashboards or current deal-flow management systems. With minimal preprocessing, startup data that is already gathered during application screening—such as founder biographies, funding histories, market characteristics, and product attributes—can be easily translated to the model’s input properties. By removing the necessity for centralized data repositories, the federated learning framework further permits cooperative model training across several enterprises while protecting data privacy. Because of this, the DSS may assist investors in making decisions throughout the screening, due diligence, and portfolio monitoring phases without interfering with ongoing business operations. Furthermore, by monitoring longitudinal success trends, government organizations and funding agencies can use the model to evaluate the effects of grant programs or policy measures. When taken as a whole, these useful results demonstrate how the suggested system might improve the effectiveness, equity, and openness of startup assessment procedures.

Conclusion

In this study, a novel multi-layered decision support system combining Deep Neural Networks (DNN), Federated Learning (FL), Graph Convolutional Networks (GCN), and Knowledge Graphs (KG) is presented to predict the startup success and failure risks. Unlike usual models that primarily focus on financial indicators, the proposed technique includes both non-financial factors variables. The approach preserves data privacy while capturing the inter-dependencies of the startup ecosystem using graph-based learning. With an accuracy of 93% and an AUC of 0.90, the model showed excellent discriminatory power and high predictive performance. Furthermore, SHAP analysis and feature importance showed that team relationships and strategic clarity are critical to startup survival. This study provides a useful, privacy-conscious decision-making tool for investors, incubators, and policymakers in addition to providing a theoretical contribution by broadening the methodological arsenal for entrepreneurship analytics. To further increase real-world relevance, future work will concentrate on improving interpretability, automating qualitative feature extraction, and growing the dataset with multi class classification. Although SHAP based analysis are to be incorporated to the current model since it remains a challenge in the sensitive decision making platforms.

Author contributions

C.P. and X.P. conceptualized and designed the study. C.P. developed the methodology and carried out the experiments. L.S. contributed to data preprocessing and interpretation. X.P. supervised the research and critically revised the manuscript. All authors contributed to manuscript writing and approved the final version.

Funding

This work was supported by the Research Project of Zhejiang Federation of Humanities and Social Sciences (Grant No. 25HQZZ083YB): “A Study on the Construction of Entrepreneurial Risk Awareness Indicators for Dual-Innovation Education of Vocational College Students in Zhejiang Province.”

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.