With the Internet technology developing, international trade becomes convenient and fast, while cross-border e-commerce trade has become more common when people's demand of domestic consumers for overseas goods increases. In order to quickly and accurately recommend cross-border e-commerce commodities that consumers want, the keywords are extracted using TF-IDF (Term Frequency-Inverse Document Frequency) algorithm, and the commodity data are sentiment classified using the deep learning LSTM (Long Short-Term Memory) model, and finally the association rule algorithm is used to classify the extracted keywords. Finally, the extracted keywords are analyzed using an association rule method, so as to study the consumer's online shopping preference. For the cross-border e-commerce product selection method on the foundation of TF-IDF, the results show that TF-IDF in the cosmetics category has the highest accuracy rate of 90.8%. TF-IDF in the health care category has achieved the highest recall rate of 82.6%. The model achieved the highest classification accuracy of 90.20%, the average accuracy of 86.47% and the lowest accuracy of 83.20%. In summary, the researched cross-border e-commerce product selection method on the foundation of TF-IDF is able to recommend suitable products for consumers in a targeted manner.