By: Jessie Epstein
Q: Your research spans multimodal retrieval, AI-driven market segmentation, and attention-based sales forecasting. What connects these three directions?
Han: All three studies focus on how intelligent algorithms can improve decision-making in e-commerce systems. Although the applications differ, the core motivation remains the same: user behavior has become more complex, product information is increasingly multimodal, and market competition demands more accurate predictions. These papers examine how to use advanced machine learning to understand users, interpret content, and forecast demand.
Q: Let us start with your paper on multimodal retrieval. What challenge were you addressing in “Research on Multimodal Retrieval System of E-commerce Platform Based on Pre-training Model”?
Han: E-commerce platforms often rely on keyword-based retrieval, which does not fully capture the richness of product information. Images, text descriptions, user reviews, and attributes all carry meaning. My study looked at how pre-trained models can integrate these inputs into a unified embedding space. The goal was to improve retrieval accuracy by allowing the system to understand the alignment between visual content and language. This is especially important when users search with incomplete or ambiguous terms.
Q: What were the key innovations in your multimodal approach?
Han: The study introduced a retrieval architecture that uses pre-trained vision language models to extract semantic features across modalities. It also explored feature alignment strategies to ensure that text and image embeddings remain consistent. Compared with single modality baselines, the multimodal system produced more accurate matching, especially on long tail or visually complex products. This showed the value of using pretrained knowledge to enhance retrieval robustness.
Q: Moving to your second paper on AI-driven market segmentation and sequential recommendation, what inspired this research direction?
Han: E-commerce user behavior is no longer linear. A consumer may browse repeatedly, compare prices, add items to the cart, abandon them, and return days later. Traditional recommendation methods often treat these behaviors separately. The study aimed to unify market segmentation and sequential recommendation by modeling how behaviors evolve. Understanding behavioral transitions is key to predicting future actions and designing more personalized user experiences.
Q: How does your research connect segmentation with recommendations?
Han: Segmentation provides the “who,” while recommendation provides the “what” and “when.”
In the paper, I explored how integrating these two processes allows platforms to treat each consumer not as a static profile but as someone whose intentions evolve over time. When segmentation and recommendation inform each other, marketing becomes more adaptive. For example, a user showing early signs of switching interest may receive discovery-based recommendations, while a high-intent buyer may see complementary or urgency-based suggestions.
Q: How did you validate the performance of this recommendation framework?
Han: The experiments used real-world datasets containing multiple user actions. We compared the system with strong baselines on metrics such as accuracy and hit rate. The results consistently showed improvements. These gains indicate that integrating segmentation with sequential modeling can create more realistic user representations.
Q: Your third paper examines sales volume and price forecasting. How does forecasting support marketing strategy?
Han: Forecasting helps marketers make informed decisions about campaigns, inventory planning, and pricing windows. If a platform can anticipate demand shifts, it can time promotions more effectively, avoid overstock or stockouts, and respond to seasonality with greater precision.
My research looked at how patterns in historical sales can help marketers identify rising items, price-sensitive products, or categories where demand is unstable. This supports both tactical decisions and long-term planning.
Q: What practical marketing insights can e-commerce brands draw from your findings?
Han: There are several. First, multi-step behavior should be monitored holistically. A repeated “add to cart then remove” pattern signals a very different intent from a casual browser. Second, personalized messaging can be triggered by behavioral transitions rather than static labels. Third, understanding behavior sequences allows brands to design smarter funnels, such as knowing when to introduce incentives, reminders, or alternative products.
These strategies improve customer experience while reducing marketing waste.
Q: Across your three studies, what best represents your research capability in the marketing domain?
Han: I focus on identifying real marketing challenges and designing analytical methods that make those challenges visible and solvable. Whether the topic is product discovery, personalization, or forecasting, the goal is to link data-driven intelligence with consumer psychology, business goals, and operational constraints. I also emphasize rigorous validation with real-world datasets, which helps ensure that the insights can scale in commercial environments.
Disclaimer: The information provided in this article is for general knowledge purposes only. It does not constitute legal, business, or professional advice. While efforts have been made to ensure the accuracy of the information, the details shared are based on publicly available sources as of the date of publication. For specific legal or business matters, consulting a qualified professional is recommended.
