MUMBAI, India, Oct. 26 -- Intellectual Property India has published a patent application (202611107131 A) filed by Kulvinder Singh; Sanjeev Dhawan; and Saumya Goyal on September 06, 2026, for An Intelligent Hardware-Aware Multi-Objective Optimization Framework For Large-Scale Personalized Recommendation Systems.
Inventors include Kulvinder Singh; Sanjeev Dhawan; and Saumya Goyal.
The application for the patent was published on October 23, 2026, under issue no. 43/2026.
Abstract: With the emergence of large-scale online platforms that manage millions of user-item interactions, there is a growing need for recommender systems that can deliver high accuracy, computational efficiency, and scalability. Although much work has addressed individual optimization methods, there is still no comprehensive empirical knowledge of the interactions between architectures, hardware acceleration, and sampling strategies for various data scales. This discrepancy complicates the determination of whether an optimization technique is genuinely superior or merely appears more effective due to specific factors such as dataset size, hardware, or testing configuration. This work carries out a thorough empirical comparison of twelve basic optimization techniques. These include six parallel processing architectures (ALS, Async-SGD, Sync-SGD, DDP, Parallel Knowledge Enrichment and Embedding Sharding + Cache) and six sampling strategies (Uniform, Group- Specific, Popularity-Based, Hard Negative, Self-Adversarial and Importance) evaluated on both CPU and GPU executors using MovieLens-1M and MovieLens-25M datasets. The research aims to structure four research questions regarding scalability, hardware-acceleration gaps, performance, and portability, which will guide the investigation of sampling–efficiency trade-offs and hardware–algorithm synergy. The results clearly reveal a scalability wall for matrix-based methods, with ALS going from a throughput collapse of -98.3 percent from MovieLens-1M to MovieLens-25M on GPU, and Async-SGD showing a scalability gain of +6880 percent over the same transition. Showing that how well an optimization method depends on the specific setting, not solely on the method itself. A synchronization paradox is also introduced, showing that in small datasets, asynchronous GPU training (Async-SGD) performs at 0.5 times the CPU speed—indicating that GPU acceleration is not always beneficial. The opposite result is obtained in sampling: hard negative sampling on GPU is 26 percent slower in NDCG@10 on MovieLens-1M, and self- adversarial sampling is 60 percent slower but doesn't provide any significant boost in accuracy. Importance sampling is the most practical approach evaluated to be performed on the GPU, where it offers accuracy and throughput within 5 percent of uniform and has principled bias correction. To tackle the trade-off between training speed and recommendation quality, the Accuracy–Efficiency Score (AES) is introduced, a single score that captures the accuracy and efficiency of the model. AES turns this trade-off into a single score that helps decide which setup to use, rather than measuring accuracy alone. The results offer a decision support model for choosing parallelism and sampling strategies given scale and hardware availability and inspire adaptive models that optimize over these parameters in online recommender systems in production. Overall, no single optimization strategy works best in every case; the right choice depends on aligning the hardware and algorithm with the data scale.
Disclaimer: Curated by HT Syndication.
