Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting Startup-VC Fund Matches with Structural Embeddings and Temporal Investment Data

Koutarou Tamura

TL;DR

The paper tackles predicting whether a VC fund will include a given startup in its portfolio by formulating it as a fund-specific binary classification task. It introduces a multi-modal startup embedding that fuses textual, numerical, categorical, and structural signals, with structural context derived from a Node2Vec embedding of a bipartite investment graph, and models fund behavior as a sequence of past investments using an LSTM. A binary compatibility scorer predicts inclusion, trained end-to-end on historical fund–startup pairs, with ablations showing the value of structural information and robustness to unseen startups via imputation. On Japanese startup data, the approach yields a notable improvement in F1 score over a baseline and demonstrates practical robustness and potential for broader applicability with interpretability enhancements planned for future work.

Abstract

This study proposes a method for predicting startup inclusion, estimating the probability that a venture capital fund will invest in a given startup. Unlike general recommendation systems, which typically rank multiple candidates, our approach formulates the problem as a binary classification task tailored to each fund-startup pair. Each startup is represented by integrating textual, numerical, and structural features, with Node2Vec capturing network context and multihead attention enabling feature fusion. Fund investment histories are encoded as LSTM based sequences of past investees. Experiments on Japanese startup data demonstrate that the proposed method achieves higher accuracy than a static baseline. The results indicate that incorporating structural features and modeling temporal investment dynamics are effective in capturing fund-startup compatibility.

Predicting Startup-VC Fund Matches with Structural Embeddings and Temporal Investment Data

TL;DR

The paper tackles predicting whether a VC fund will include a given startup in its portfolio by formulating it as a fund-specific binary classification task. It introduces a multi-modal startup embedding that fuses textual, numerical, categorical, and structural signals, with structural context derived from a Node2Vec embedding of a bipartite investment graph, and models fund behavior as a sequence of past investments using an LSTM. A binary compatibility scorer predicts inclusion, trained end-to-end on historical fund–startup pairs, with ablations showing the value of structural information and robustness to unseen startups via imputation. On Japanese startup data, the approach yields a notable improvement in F1 score over a baseline and demonstrates practical robustness and potential for broader applicability with interpretability enhancements planned for future work.

Abstract

This study proposes a method for predicting startup inclusion, estimating the probability that a venture capital fund will invest in a given startup. Unlike general recommendation systems, which typically rank multiple candidates, our approach formulates the problem as a binary classification task tailored to each fund-startup pair. Each startup is represented by integrating textual, numerical, and structural features, with Node2Vec capturing network context and multihead attention enabling feature fusion. Fund investment histories are encoded as LSTM based sequences of past investees. Experiments on Japanese startup data demonstrate that the proposed method achieves higher accuracy than a static baseline. The results indicate that incorporating structural features and modeling temporal investment dynamics are effective in capturing fund-startup compatibility.

Paper Structure

This paper contains 20 sections, 8 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Company Representation via Embedding.
  4. Graph-Based Modeling of Investment Networks.
  5. Combining Network and Temporal Modeling.
  6. Dataset
  7. Overview of the Dataset
  8. Bipartite Investment Graph Construction
  9. Startup Feature Categories
  10. Fund Feature Representation
  11. Transformation of structural features
  12. Proposed Method
  13. Multi-modal Company Embedding with Attention-based Fusion
  14. Temporal Modeling of Fund Portfolios
  15. Inclusion Prediction via Compatibility Scoring
  16. ...and 5 more sections

Figures (2)

  • Figure 1: Overview of the startup embedding module. Multi-modal inputs—textual (BERT-based), numerical, categorical, and structural (Node2Vec-based) features—are independently transformed (with "L" denoting a Linear layer) and then integrated via a multi-head attention mechanism to form the unified startup embedding $\mathbf{e}_c$.
  • Figure 2: Architecture of the inclusion prediction model. A VC fund is represented as a sequence of past startup embeddings $\mathbf{e}_c$, encoded via an LSTM to obtain a fund embedding $\mathbf{f}_f$. This embedding is then combined with a candidate startup embedding to compute the inclusion probability $\hat{y}_{f,c}$ through a binary classifier.