MUMBAI, India, Sept. 14 -- Intellectual Property India has published a patent application (202641106627 A) filed by Dr. Suma Srinath; Dr. Nagaraju Baydeti; and Nitte Meenakshi Institute Of Technology, Nitte Deemed To Be University, Bengaluru on September 04, 2026, for Ai-Driven Optimization Of Video Sharing In Social Networks Eliminating Redundant Traffic And Enhancing User Engagement.
Inventors include Dr. Suma Srinath; Dr. Nagaraju Baydeti; and Nitte Meenakshi Institute Of Technology, Nitte Deemed To Be University.
The application for the patent was published on September 11, 2026, under issue no. 37/2026.
Abstract: CLAIMS I / We Claim: 1. A self-optimizing deep learning based framework (200) for intelligent management of video traffic in a social networking platform through content-aware redundancy elimination and user-interest driven prioritization, the said framework (200) comprising: a video upload interface (202) coupled, through a communication network (222), to a plurality of sender end-user devices (224), said interface (202) being configured to receive a video uploaded by a sender end-user device; a Convolutional Neural Network (CNN) based video content classification engine (204) configured to receive the said uploaded video and to produce a content embedding and at least one content category label therefor; a perceptual hash generation and matching engine (208) coupled to a video content repository (220) and configured to generate at least one perceptual hash signature from one or more key frames of the said uploaded video and to compare the said perceptual hash signature against perceptual hash signatures stored in the said repository (220) using a similarity threshold; a Long Short-Term Memory (LSTM) based sequential user-interest prediction engine (206) coupled to a user interaction profile database (218) and configured to process time-ordered interaction sequences of one or more candidate recipient users to produce respective user-interest vectors; a Transformer based attention priority scoring engine (212) configured to compute, using a multi-head attention mechanism, a scalar priority score for each candidate pair of (the said uploaded video, one of the said candidate recipient users) based on at least the said content embedding, the said user- interest vector and at least one real-time network metric; a Deep Q-Learning based redundancy policy controller (210) operatively coupled to the said perceptual hash generation and matching engine (208), to the said LSTM engine (206) and to the said Transformer engine (212), and configured to select, by reinforcement learning, an action from an action space comprising at least blocking of re-transmission, throttled transmission and full transmission, so as to maximize a reward function defined over bandwidth saving and predicted user engagement; a content distribution scheduler (214) configured to schedule transmission of the said uploaded video, as an optimized video delivery stream (228), to one or more recipient end-user devices (226) in accordance with the said priority score and the said selected action; and a real-time traffic and engagement monitor (216) cooperatively coupled with an AI orchestration and retraining engine (230), characterised in that, the said real-time traffic and engagement monitor (216) and the said AI orchestration and retraining engine (230) together implement a closed-loop retraining feedback path that continuously measures live bandwidth, latency and engagement metrics and uses the said metrics to periodically retrain the said CNN engine (204), the said LSTM engine (206), the said Transformer engine (212) and the said Deep Q-Learning controller (210), whereby the framework (200) continuously self-optimizes its decisions to suppress redundant video transmissions and to prioritize delivery of relevant videos to interested users in the said social networking platform. 2. The framework (200) as claimed in claim 1, wherein the said perceptual hash generation and matching engine (208) is further configured to (i) extract a plurality of key frames at temporally spaced intervals from the said uploaded video, (ii) generate a respective perceptual hash signature for each of the said key frames, (iii) partition each said perceptual hash signature into a coarse segment and a fine segment, (iv) use the said coarse segment to retrieve a candidate group of stored signatures from the said video content repository (220), and (v) declare the said uploaded video as a near-duplicate when a Hamming distance between the said fine segment and a fine segment of any stored signature in the said candidate group is below a configurable similarity threshold. 3. The framework (200) as claimed in claim 1, wherein the said Long Short-Term Memory based sequential user-interest prediction engine (206) comprises one or more stacked LSTM layers configured to process time-ordered sequences of interaction events including at least one of view-duration, replay, like, share, comment and skip events, and to output a user-interest vector whose components represent predicted relevance, at the present time instant, of each of a plurality of content categories to the said user, the said user-interest vector being supplied directly as an input embedding to the said multi-head attention mechanism of the Transformer based attention priority scoring engine (212). 4. The framework (200) as claimed in claim 1, wherein the said Deep Q-Learning based redundancy policy controller (210) is trained, by interaction with the said framework (200), to maximize a reward function R defined as a weighted sum of (i) bandwidth saved by suppression of redundant transmissions, (ii) a predicted engagement score derived from the said priority score, and (iii) a content freshness score derived from the said content category label and the said user interaction profile database (218), the weights of the said reward function R being configurable parameters of the said framework (200). 5. A method, implemented in the framework (200) as claimed in claim 1, for intelligent management of video traffic in a social networking platform, the said method comprising the steps of: receiving (304), at the said video upload interface (202), a video uploaded by a sender end-user device (224); classifying (306), at the said CNN based video content classification engine (204), the content of the said video and generating a content embedding therefor; generating (308), at the said perceptual hash engine (208), at least one perceptual hash signature from one or more key frames of the said video and matching (310) the said perceptual hash signature against signatures stored in the said video content repository (220); determining (312) whether a near-duplicate of the said video has been found and, if so, blocking (314) re- transmission of the said video and referencing the previously stored copy; if no near-duplicate is found, predicting (316), at the said LSTM engine (206), user-interest vectors of candidate recipient users, computing (318), at the said Transformer engine (212), attention-based priority scores for each candidate (video, user) pair, selecting (320), at the said Deep Q-Learning controller (210), a distribution action, and transmitting (322) the said video to targeted recipient end-user devices (226); and monitoring (324), at the said real-time traffic and engagement monitor (216), live bandwidth, latency and engagement metrics, and retraining (326), at the said AI orchestration and retraining engine (230), the said CNN engine (204), LSTM engine (206), Transformer engine (212) and Deep Q-Learning controller (210) using the said metrics in a closed-loop self-optimizing manner.
Disclaimer: Curated by HT Syndication.
