Machine Learning Model Couples Group Recommendations with Startup Survival Predictions

A computational evaluation system that pairs weighted collaborative filtering with extreme gradient boosting can predict entrepreneurial venture survival and rank promising startup investments for decision-making teams, according to research published in Discover Internet of Things. Investigators Yan Hou and Shuling Yang developed a dual-module framework that addresses the complex dynamics of venture capital syndicates and incubation panels, where individual evaluators hold divergent risk tolerances and preference weights. Rather than isolating recommendation ranking from downstream survival forecasts, the researchers unified both tasks into an interactive analytical pipeline trained on historical enterprise records. When tested on Chinese technology enterprises, the group recommendation engine attained an overall recommendation accuracy of 0.874 alongside an area under the receiver operating characteristic curve of 0.902, while the accompanying binary classification model achieved an accuracy of 0.872 and an area under the curve of 0.912 when forecasting whether enterprises would survive or deregister.

Traditional screening procedures in venture capital and startup incubation frequently depend on subjective expert intuition or isolated predictive algorithms that fail to synthesize group consensus. Venture investment decisions inherently represent collective judgments where multiple partners, investment managers, and technical mentors negotiate competing priorities across team capabilities, technological maturity, and market positioning. Furthermore, individual algorithmic tools that rank business proposals rarely model whether an enterprise possesses the operational resilience to avoid formal cancellation or liquidation over subsequent years. Hou and Yang sought to resolve this disconnect by designing a structured workflow comprising five interconnected phases: enterprise data acquisition, cleaning and multi-modal feature engineering, preference-weighted group recommendation, machine-learning-based survival probability estimation, and dual-track evaluation. Within this architecture, recommendation outputs filter candidate pools to reduce computational overhead for prediction, while probability scores continuously inform and optimize feature weighting parameters within the recommendation engine.

The empirical foundation of the investigation relies on 12,847 validated enterprise records extracted from the IT Juzi startup company database, focusing on companies established between 2015 and 2024 across five high-technology sectors: artificial intelligence, Internet education, medical health, new energy, and consumer goods. Within the aggregate dataset, 8,563 enterprises were identified as active and surviving, representing 66.6 percent of the total sample, while 4,284 firms were classified as cancelled or deregistered, accounting for the remaining 33.4 percent. The distribution across organizational scales showed that 68.3 percent of the enterprises maintained between 1 and 50 employees, 24.1 percent employed between 51 and 200 individuals, and only 7.6 percent exceeded 200 staff members. Financial profiles demonstrated comparable diversity, with 42.5 percent securing less than 10 million yuan in total cumulative capital, 38.2 percent raising between 10 million and 100 million yuan, and 19.3 percent capturing capital exceeding 100 million yuan.

To guarantee strict methodological validity and prevent data leakage across temporal horizons, the investigators implemented a chronological time-series partitioning protocol rather than random sample shuffling. Enterprises founded between 2015 and 2021 formed the training set of 10,278 companies, comprising 6,850 surviving entities (66.6 percent) and 3,428 deregistered firms (33.4 percent). Enterprises founded from 2022 to 2024 constituted the independent test collection of 2,569 businesses, containing 1,713 surviving firms (66.7 percent) and 856 cancelled companies (33.3 percent), thereby maintaining identical class proportions across both splits. The researchers restricted financial variables strictly to transaction records finalized within the initial three years following enterprise registration, excluding later funding rounds to ensure future capitalization milestones did not artificially predict past operational outcomes. The final operational status of every enterprise was established against a fixed census deadline of December 31, 2024, ensuring that all predictive feature timestamps strictly preceded the definitive status determination.

The engineering workflow constructed a multi-source feature matrix integrating internal operational indicators, textual profile semantics, and macro-level environmental variables. Categorical industry classifications were mapped via One-Hot encoding vectors, while numerical variables were normalized and imputed using median or mean replacements with boundary truncation. To process unstructured enterprise descriptions, the authors compared Term Frequency-Inverse Document Frequency vectorization against semantic representations generated by a pre-trained Chinese Bidirectional Encoder Representations from Transformers model. The transformer approach yielded an initial 768-dimensional semantic embedding via mean pooling of final hidden states, which was compressed through a linear projection layer into 128 dimensions using Rectified Linear Unit activation and dropout regularization. This neural embedding strategy improved classification accuracy by 2.3 percent over baseline frequency vectorization when concatenated with structured metrics, forming a dense multi-dimensional input representation for downstream model execution.

External environmental factors were quantified to capture broader macroeconomic fluctuations that directly influence venture viability. An industry prosperity index was derived from three-year average revenue growth metrics published in the National Bureau of Statistics National Economic Industry Classification reports and scaled between zero and one. Government policy support was measured by crawling municipal and central industrial policy documents released during the three-year window prior to each enterprise establishment date, applying keyword extraction to tally occurrences of incentives such as tax relief, specialized subsidies, and direct financial allocations. Market competition dynamics were modeled using five-firm concentration ratios reported by third-party research institutions including iResearch and the Forward Industry Research Institute. High concentration values indicated elevated market barriers and rigorous incumbent competitive pressure, whereas dispersed sectors provided broader entry headroom for newly emerging technological ventures.

The group recommendation module employs collaborative filtering grounded in weighted distance aggregation across corporate features, industry alignment, and financing metrics. Pairwise enterprise similarity combines cosine similarity over industry One-Hot vectors, cosine similarity across semantic text embeddings, and inverted Euclidean distances across normalized numerical variables. Following theoretical calibration and verification via grid search maximizing Normalized Discounted Cumulative Gain at rank ten across the interval from 0.1 to 0.8, the authors assigned optimal weight coefficients of 0.25 to industry category, 0.20 to profile text, and 0.55 to financing parameters. Group preference aggregation synthesized individual mentor or investor assessments by computing weighted averages adjusted for organizational seniority, decision-making hierarchy, and historical investment hit rates, transforming historical investment endorsements and rejections into a coherent, normalized prospective portfolio ranking list.

Across ten independent experimental runs using varied initialization seeds on the chronological testing partition, the proposed recommendation framework demonstrated statistically verified performance advantages over conventional baselines. The system achieved a mean accuracy of 0.874 with a 95 percent confidence interval of 0.868 to 0.880, an F1-score of 0.861 with a confidence interval of 0.854 to 0.868, and an area under the curve of 0.902 with a confidence interval of 0.895 to 0.909. Baseline collaborative filtering recorded a lower accuracy of 0.812 with a confidence interval of 0.804 to 0.820, while neural collaborative filtering reached 0.846 with a confidence interval of 0.839 to 0.853. Paired t-tests confirmed statistical significance against traditional collaborative filtering at t equals 3.82 (p less than 0.01) and against neural collaborative filtering at t equals 2.95 (p less than 0.05). In ranking metrics, the proposed model attained a Normalized Discounted Cumulative Gain of 0.853 and Mean Average Precision of 0.841 at cutoff ten, outperforming traditional baselines of 0.782 and 0.768, as well as neural baselines of 0.814 and 0.802.

Survival probability forecasting relied on an extreme gradient boosting binary classification architecture configured with 100 decision trees, a learning rate of 0.1, a maximum tree depth of six, and internal class weight parameters compensating for sample category imbalance. The model yielded a test accuracy of 0.872 with a 95 percent confidence interval of 0.866 to 0.878, an F1-score of 0.846 with a confidence interval of 0.839 to 0.853, and an area under the curve of 0.912 with a confidence interval of 0.906 to 0.918. In comparative benchmark testing against alternative algorithms, logistic regression attained an area under the curve of 0.834 with a confidence interval of 0.826 to 0.842, while random forest generated an area under the curve of 0.876 with a confidence interval of 0.869 to 0.883. Paired t-tests demonstrated significant differences favoring extreme gradient boosting, producing a t statistic of 4.21 (p less than 0.01) against logistic regression and 2.68 (p less than 0.05) against random forest.

Interpretability assessments using Shapley Additive Explanations revealed the relative predictive leverage of specific characteristics across corporate survival trajectories. Total financing volume and funding round counts exhibited the highest positive contributions, averaging Shapley values of 0.124 and 0.108, respectively. Policy support intensity demonstrated a positive average value of 0.087, characterized by a sharp threshold effect where support values below 0.6 produced negative predictive contributions while values exceeding 0.6 strongly supported project longevity. Industry prosperity showed a positive correlation with a mean value of 0.065, whereas market concentration exerted an inverse influence averaging minus 0.038. Organizational size demonstrated non-linear behavior with an average value of 0.052: mid-sized teams spanning 50 to 200 staff achieved the strongest positive associations with survival, whereas workforces exceeding 500 members correlated negatively. Ablation testing confirmed that textual representations contributed 3.1 percent to classification accuracy, external macro-variables added 1.4 percent, and the bidirectional feedback loop connecting recommendation filtering with predictive weights contributed 0.3 percent to accuracy alongside enhanced ranking gains.

Subject of Research: Evaluation and success rate prediction of entrepreneurial projects using group recommendation and machine learning

Article Title: Evaluation and success rate prediction of innovation and entrepreneurship projects using a group recommendation algorithm

Article References: Hou, Y., & Yang, S. (2026). Evaluation and success rate prediction of innovation and entrepreneurship projects using a group recommendation algorithm. Discover Internet of Things, 6(1), Article 150. https://doi.org/10.1007/s43926-026-00434-3

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00434-3

Keywords: group recommendation algorithm, extreme gradient boosting, collaborative filtering, startup survival prediction, venture capital decision-making, entrepreneurship project evaluation, feature engineering, BERT semantic embedding, Shapley additive explanations, macroeconomic variables, industry prosperity index, policy support quantification, time-series cross-validation, IT Juzi database, market concentration ratio, ranking metrics, class imbalance modeling