Personalized recommendation algorithm based on electricity users’ usage behavior portrait

This study addresses the challenges associated with precisely targeting marketing SMS for power users of the State Grid. Existing collaborative filtering recommendation algorithms, while capable of predicting user preferences through representation learning and matching functions, fall short in accurately capturing user preferences for various activities and str uggle to effectively mine the correlations between embedded dimensions during interactions between users and activities. To overcome these limitations, we propose a Neural Collaborative Filtering (NCF) recommendation algorithm grounded in the electricity usage behavior portrait of users. Our approach begins by extracting feature vectors for both users and activities via a self-encoder, subsequently employing a multilayer perceptron enhanced by the Convolutional Block Attention Module (CBAM) to learn higher-order interactions among the feature vector dimensions. This methodology is integrated with traditional neural collaborative filtering algorithms to enhance the model’s generalization capabilities and recommendation accuracy. Additionally, recognizing the challenges associated with negative sampling in implicit feedback datasets, which can significantly affect model performance, we implement a probabilistic negative sampling method derived from traditional matrix decomposition techniques to bolster the robustness of the model. Experimental results demonstrate that our method achieves an HR (Hit Ratio) of 0.9784 and an NDCG (Normalized Discounted Cumulative Gain) of 0.9749, substantiating that the design of the recommendation model, based on the neural collaborative filtering algorithm and user electricity behavior portraits, significantly enhances both recommendation accuracy and the effectiveness of marketing pushes.