AI model sharpens entrepreneurial project recommendations from sparse user data

A recommendation system designed for an entrepreneurial service platform improved project matching by combining users’ demographic profiles, changing browsing histories and aggregated behavior statistics with a deep neural network and matrix factorization. In tests on data from 15,342 Chinese entrepreneurs and investors, the proposed IFE-DNN-MF model reduced prediction error and improved top-10 recommendations compared with general-purpose, domain-specific and conventional collaborative-filtering baselines. Its central advantage was to infer preferences that were only weakly expressed in the sparse record of which projects users had viewed, saved, shared, contacted or considered investing in.

The study, published on 18 September 2026 in an open-access research article, focused on a dataset covering January 2022 through December 2023 from a leading entrepreneurial service platform in China. The records included 8,912 projects across 12 industries, among them artificial intelligence, enterprise services, healthcare, new energy and consumer services. Users were approximately 42% investors and 58% entrepreneurs. The platform supplied anonymized registration information, identity type, industry background, preferred financing stage and entrepreneurial experience, together with project details and timestamps for positive interactions.

The scale of the recommendation problem was revealed by the interaction matrix. It contained 1,245,678 positive user-project records, yet 99.09% of its entries were empty. Users interacted with an average of about 81.2 projects, but the median was only 34.5, showing a strongly uneven, long-tail distribution in which a small group of highly active accounts generated a large share of the signals. Researchers converted browsing, bookmarking, sharing, contacting project owners and submitting investment intentions into implicit positive feedback. They then split each user’s chronologically ordered interactions into training, validation and test subsets of 80%, 10% and 10%, respectively.

IFE-DNN-MF addresses the sparsity with two connected components. Its first component extracts a latent representation of each user from three kinds of side information. Basic attributes, including industry background, investment-stage preference and years of experience, pass through an embedding layer for categorical variables and a fully connected network with rectified-linear activation. The historical sequence of project interactions is converted into embeddings and encoded by a gated recurrent unit, or GRU, which is designed to retain information about the order and evolution of user interests. Statistical summaries, such as the distribution of previously viewed project categories and average financing stage, are transformed by another fully connected layer.

The three representations are concatenated and passed through a nonlinear fusion layer to create one implicit user-feature vector. That vector does more than serve as extra input. The model also preserves a basic user embedding learned from the interaction matrix and a project embedding, then routes information through three parallel prediction paths. The matrix-factorization path multiplies user and project embeddings element by element to capture broad linear collaborative patterns. A two-layer deep neural network receives their concatenation to learn nonlinear couplings. The third path explicitly combines the implicit user features with the project embedding, allowing background and behavioral signals to influence the representation of each candidate project.

Outputs from these paths are combined using learnable fusion. One strategy concatenates them before a final linear transformation; another projects the matrix-factorization and neural-network outputs into a common dimension, adds all three and passes the result through a sigmoid function. The design is intended to address three different questions: which project types a user has historically engaged with, how user and project characteristics interact nonlinearly, and which projects fit the user’s background and evolving interests. The authors also describe the architecture as a potential response to cold-start and preference ambiguity, where too few interactions are available for collaborative signals alone.

Across five repeated experimental runs, IFE-DNN-MF produced the strongest overall results. Relative to DIN, identified as the best baseline in the direct comparison, its root mean square error fell by 6.8% and mean absolute error by 7.7%. Its precision at 10 rose by 16.8%, recall at 10 by 16.9% and normalized discounted cumulative gain at 10 by 16.6%. The model reached a recall-at-10 score of 0.2312, which was 46.5% above matrix factorization and 24.7% above neural collaborative filtering. Against the domain-specific CFIL model, its RMSE was 5.7% lower, while its recall at 10 was 11.9% higher. These are offline evaluation results on one platform, rather than evidence that the system has improved investment, financing or venture outcomes.

Ablation experiments indicated that each major component contributed to performance. Removing the entire implicit-feature extraction module raised RMSE from 0.2412 to 0.2598 and reduced recall at 10 by 14.5% to 0.1976. Removing the dedicated feature-enhancement path reduced recall at 10 by 12.5% to 0.2023, while removing the nonlinear neural path lowered recall by 11.2% and increased RMSE by 5.1%. These comparisons support the authors’ claim that simply adding hidden features to a conventional model is less effective than creating a separate route through which those features interact with project representations.

The researchers also tested whether adding project-description text was worthwhile. A supplementary variant encoded introductions and industry descriptions with a pretrained Chinese BERT model, extracted its 768-dimensional contextual representation, compressed it to 64 dimensions and concatenated it with the project identifier embedding. Text produced only a 1.25% to 1.71% gain across the reported indicators, with the largest improvement, 1.54% in recall at five, while adding encoding and inference costs. The authors therefore left text out of the main architecture, arguing that structured fields such as industry, financing stage and team size already captured much of the useful content. A separate comparison of negative-sampling methods found that using other users’ positive projects from the same training batch performed best, raising recall at 10 from 0.2312 to 0.2367, about 2.4% above random sampling.

Two illustrative cold-start cases provide a qualitative view of how the model could use information that conventional collaborative filtering misses. For a new institutional investor with only three browsing records, IFE-DNN-MF recommended artificial-intelligence and big-data projects at angel or Series A stages, aligning with the user’s stated investment preferences, whereas DIN concentrated on enterprise-service projects at later financing stages. For a newly registered entrepreneur with no platform interactions, the proposed model used a profile focused on renewable-energy battery technology, a 10–15-person team and prior photovoltaic and energy-storage experience. Its recommendations included battery-recycling technology, suggesting an inferred interest in the battery value chain. These cases are illustrative rather than controlled proof of better user outcomes, and the underlying platform dataset is not public because of a commercial confidentiality agreement; the authors say the preprocessed data may be available on reasonable request and that source code was submitted as supplementary material.

Subject of Research: Evaluation of an implicit-feature neural recommendation model for entrepreneurial project matching

Article Title: Entrepreneurial project recommendation by implicit user feature extraction and DNN-MF

Article References: Ma, L. (2026). Entrepreneurial project recommendation by implicit user feature extraction and DNN-MF. Discover Artificial Intelligence, 6(1), Article 1177. https://doi.org/10.1007/s44163-026-02167-2

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02167-2

Keywords: recommendation systems, matrix factorization, deep neural networks, entrepreneurial projects, implicit feedback, cold-start users, behavior sequences, gated recurrent units, platform data, Entrepreneurial, project, recommendation