NLP-Based Recommendation Approach for Diverse Service Generation

In this study, we examine the potential of language models for natural language processing (NLP)-based recommendations, with a distinct focus on predicting users' next product purchases based on their prior purchasing patterns. Our model specifically harnesses tokenized rather than complete product names for learning. This granularity allows for a refined understanding of the interrelations among different products. For instance, items like 'Chocolate Milk' and 'Coffee Milk' find linkage through the shared token 'Milk.' Additionally, we explored the impact of various n-grams (unigrams, bigrams, and trigrams) in tokenization to further refine our understanding of product relationships and recommendation efficacy. This nuanced method paves the way for generating product names that might not exist in current retail settings, exemplified by concoctions like 'Coffee Chocolate Milk.' Such potential offerings can provide retailers with fresh product brainstorming opportunities. Furthermore, scrutiny of the frequency of these generated product name tokens can reveal prospective trends in purchasing keywords. This facilitates enterprises in creative brainstorming of novel products and swiftly responding to the dynamic demands and trends of consumers. The datasets used in this study come from UK e-Commerce and Instacart Data, comprising 71,205 and 166,440 rows, respectively. This investigation juxtaposes the NLP-based recommendation model, which employs tokenization, with its non-tokenized counterpart, leveraging Hit-Rate and mean reciprocal rank (MRR) as evaluative benchmarks. The outcomes distinctly favor the tokenized NLP-based recommendation model across all evaluated metrics.