To develop an adaptive modulation scheme for flexible high-speed multi-user visible light communication (VLC), automatic modulation classification (AMC) is adopted for monitoring the modulation formats of different subcarrier groups. An AMC scheme based on a joint convolutional neural network (CNN), active learning (AL), and data augmentation (DA) is demonstrated over an orthogonal frequency division multiplexing access (OFDMA) VLC system. The configuration of the diffuse white-light VLC system is combined with a pair integrated transceiver module, a light-diffusing fiber (LDF), and a wireless channel, which can provide white-light illumination and ubiquitous access. Within the forward error correction (FEC) threshold, the data rates of the white-light VLC links can reach 325.5 Mbps with a bit error rate (BER) of 2.163 × 10-3. An experiment with two-user access via the proposed VLC link with an unequal bandwidth allocation was demonstrated. The performance of the AL-aided CNN AMC scheme also shows a classification accuracy rate of 95.48% for the constellation diagrams of different subcarriers of the OFDMA signal over 240 training samples and faster convergence than a CNN-based AMC.