关键词: 购买意图识别, 卷积神经网络, 深层语义模型, 深度学习

Abstract: With the rapid proliferation and intelligent development of e-commerce platforms, accurate identification of user purchase intention has become a crucial influencing factor in driving users from intent toactual purchases. Therefore, identifying user purchase intention is one of the significant methods to enhance the Purchase Rate (PR) in the realm of e-commerce. Purchase intention identification aims to infer the intended purchase of potential customers or users by analyzing the similarity between user query text and product description text, ultimately increasing the PR. Due to the diversity and colloquial nature of user search queries, identifying user purchase intention becomes increasingly challenging, and even more so in vertical e-commerce where users may not even be aware of the names of the products they need.
In response to the phenomenon of unclear user purchase intention, this paper proposes a novel model aimed at identifying user purchase intention from user queries with unclear purchase intention. This model first employs the Word2Vec (WV) algorithm’s Continuous Bag-of-Words (CBOW) model to train word vectors. Subsequently, these word vectors are fed into a one-dimensional Convolutional Neural Network (CNN), followed by further feature extraction using the Deep Semantic Similarity Model (DSSM). This process calculates semantic similarity using cosine similarity, subsequently transforming semantic similarity into a posterior probability form to construct a loss function. During model training, it narrows the textual representations in a high-dimensional space between user queries and intended products while expanding the representations between user queries and non-intended products.
An empirical analysis is conducted using real search data from the U.S. building materials e-commerce website Home Depot, within the Keras framework. The results indicate that our proposed model achieves an F1-score of 80.6% on the test dataset in a five-class classification problem. To test the performance of the model proposed in this paper in more complex purchase scenarios, six, seven, and eight-class classification tasks are designed. The results also indicate that as the number of categories increases, the values of various evaluation metrics decrease. However, the F1-scores for all three classification tasks remain above 70%, demonstrating competitive performance in multi-class tasks.
Through the empirical research, this paper draws the following conclusions: (1)The proposed model leverages Word2Vec and CNN for text feature extraction and employs the DSSM model to further extract feature representations of user queries and product descriptions in a high-dimensional space. This maximizes the utilization of semantic similarity between user queries and the correct product descriptions while avoiding subjective interference during feature extraction, ultimately enhancing the identification of purchase intention for products. (2)Deep learning models are often too large to be practical in real-world scenarios. In contrast to typical deep learning models, the model proposed in this paper converges at a faster rate. (3)The model’s F1-score is significantly higher than the baseline model, and as the number of categories increases, the model’s evaluation scores still maintain a high level. (4)Real training data often exhibit class imbalance issues. The model proposed in this paper constructs negative examples based on positive data to balance the data quantity across different categories, enabling the model to consider all categories during the training process. The method proposed in this paper can only identify users’ intended products within a small number of product descriptions. How to identify users’ intended products within a massive volume of product descriptions is a further research direction.

Key words: purchase intention identification, convolutional neural networks(CNN), deep structured semantic model(DSSM), deep learning