Deep Learning Modeling and Personalized Recommendation System Based on CNN and Attention Mechanism

随着信息时代的到来，互联网成为学习和共享的重要平台，但庞大的电影数据量和信息复杂性使用户查询变得困难。为帮助用户快速找到感兴趣的电影，我们提出了一个结合深度学习和注意力机制的个性化电影推荐系统。通过分析用户基本信息和评分数据，该系统利用特征嵌入、卷积、自注意力机制、LSTM层以及残差模块，训练出个性化推荐模型，以提升推荐效果和准确性。在特征处理方面，用户特征通过嵌入层将年龄、性别(one-hot编码)及职业映射为稠密向量；电影特征借助multi-hot编码与卷积层捕捉类型组合的特征共现模式，再结合电影ID嵌入向量保留独特性。特征交互模块通过卷积提取用户与电影的偏好关联，自注意力机制动态调整特征权重以适应兴趣变化，LSTM层捕捉观影行为的长期依赖与短期变化，残差模块则有效融合兴趣信息并避免梯度消失，确保模型的稳定性与高效训练。With the advent of the information age, the Internet has become an important platform for learning and sharing. However, the vast amount of movie data and the complexity of information make it difficult for users to search. To help users quickly find movies they are interested in, we propose a personalized movie recommendation system that combines deep learning and attention mechanisms. By analyzing users’ basic information and rating data, the system uses feature embedding, convolution, self-attention mechanisms, LSTM layers, and residual modules to train a personalized recommendation model to improve the recommendation effect and accuracy. In terms of feature processing, user features are mapped to dense vectors through an embedding layer, including age, gender (one-hot encoded), and occupation; movie features capture the co-occurrence patterns of genre combinations through multi-hot encoding and convolutional layers and combine movie ID embedding vectors to retain uniqueness. The feature interaction module extracts the preference correlation between users and movies through convolution, the self-attention mechanism dynamically adjusts feature weights to adapt to changes in interest, the LSTM layer captures the long-term dependencies and short-term changes in viewing behavior, and the residual module effectively fuses interest information and avoids gradient vanishing, ensuring model stability and efficient training.