US 12,306,866 B2
Multimedia resource classification and recommendation
Lingzi Zhu, Shenzhen (CN); and Lianyang Ma, Shenzhen (CN)
Assigned to TENCENT TECHNOLOGY (SHENZHEN) COMPANY LIMITED, Shenzhen (CN)
Filed by Tencent Technology (Shenzhen) Company Limited, Guangdong (CN)
Filed on Nov. 22, 2022, as Appl. No. 17/992,541.
Application 17/992,541 is a continuation of application No. PCT/CN2022/072341, filed on Jan. 17, 2022.
Claims priority of application No. 202110113770.7 (CN), filed on Jan. 27, 2021.
Prior Publication US 2023/0084466 A1, Mar. 16, 2023
Int. Cl. G06F 16/40 (2019.01); G06F 16/43 (2019.01); G06F 16/45 (2019.01)
CPC G06F 16/45 (2019.01) [G06F 16/43 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A multimedia resource classification model training method, comprising:
acquiring an attribute information set and a training label set of training multimedia resources, the attribute information set comprising attribute information corresponding to a plurality of dimensions, and the training label set comprising training labels corresponding to a plurality of tasks, the training labels indicating a quality of corresponding training multimedia resources;
inputting the attribute information set of the training multimedia resources into a multimedia resource classification model comprising a plurality of feature sub-networks corresponding to the attribute information and a plurality of task sub-networks corresponding to the plurality of tasks;
vectorizing, using the plurality of feature sub-networks, the attribute information to obtain attribute feature vectors outputted by the plurality of feature sub-networks;
inputting the obtained attribute feature vectors into the plurality of task sub-networks to obtain prediction labels corresponding to the plurality of tasks; and
obtaining a trained multimedia resource classification model by adjusting model parameters of one of the task sub-networks based on a training label from the training label set and based on a prediction label that correspond to a task associated with the one of the task sub-networks, and by adjusting model parameters of one of the feature sub-networks based on the training label and the prediction label that correspond to the task until a convergence condition is satisfied.