MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202611064581 A) filed by Noida Institute Of Engineering And Technology Niet on May 22, 2026, for Synthetic Markov Chain Pre-Trained Transformer System With Adaptive Input Adaptor For Sequential Recommendation.
Inventors include Dr. Sanna Mehraj Kak; and Himanshu Pabbi.
The application for the patent was published on July 17, 2026, under issue no. 29/2026.
Abstract: The present invention provides a Markovian pre-trained transformer system (100) for sequential next-item recommendation addressing the problem of domain-specific retraining requirements and cross-domain transfer limitations in existing recommendation models arising from heterogeneous interaction data and non-transferable learned representations. The system comprises a synthetic Markov chain data generator (101) that samples transition probability matrices from a Dirichlet distribution and generates state trajectories with independently generated orthogonal state embeddings, a transformer backbone (102) pre-trained on next-state prediction to acquire transition probability estimation and last-state attention capabilities, and a lightweight input adaptor (103) comprising an RMSNorm module (104), linear projection layers (105, 107), and LeakyReLU activation (106) that maps semantic item representations into the frozen backbone input space during recommendation fine-tuning. The system achieves cross-domain recommendation accuracy improvements while maintaining inference latency comparable to lightweight domain-specific models, applicable to electronic commerce platforms, digital content services, and online retail recommendation systems.
Disclaimer: Curated by HT Syndication.
