Unsupervised Item-Related Recommendation Method Combining BERT and Collaborative Filtering

Abstract

Item-related recommendation is widely used in e-commerce, news, video, and other business scenarios, but there are problems such as sparse data, a large amount of implicit feedback data, limited sample annotation, cold start of items, poor serendipity, insufficient real-time performance, and the recommendation effect needs to be continuously improved. An unsupervised recommendation method is proposed. The method included four recall strategies. The first was to use the search engine and BM25 for real-time text matching recommendation about multi fields, and the second was to combine pre-trained language model BERT and ANN algorithm for real-time semantic matching recommendation about multi fields, and the third was to calculate the similarity by reducing the influence of popular items and active users to optimize the item-based collaborative filtering recommendation algorithm, and the fourth was to introduce the heat index based on Wilson Confidence Intervals to assist the recommendation ranking. Finally, the four recall results were merged and sorted to generate the final recommendation result. Through multiple sets of comparative experiments by the AlB test in the online recommendation system, it is shown that the proposed unsupervised recommendation method is superior to the baseline method in multiple indicators and can effectively improve the recommendation effect and user satisfaction.