{"id":60458,"date":"2025-04-18T17:26:32","date_gmt":"2025-04-18T17:26:32","guid":{"rendered":"https:\/\/www.bloomreach.com\/?post_type=library&p=60458"},"modified":"2025-10-21T17:00:32","modified_gmt":"2025-10-21T17:00:32","slug":"how-sequential-two-tower-neural-networks-revolutionize-product-recommendations","status":"publish","type":"library","link":"https:\/\/www.bloomreach.com\/en\/blog\/how-sequential-two-tower-neural-networks-revolutionize-product-recommendations","title":{"rendered":"Beyond Simple Similarity: How Sequential Two-Tower Neural Networks Revolutionize Product Recommendations"},"content":{"rendered":"\n

In the dynamic world of ecommerce, shopper behavior is rarely static. Customers\u2019 needs and interests evolve over time, and you need to meet those preferences if you hope to keep them coming back to your brand. <\/p>\n\n\n\n

We\u2019ve previously covered the benefits of Two-Tower Neural Networks<\/a> (2TNN) and how they offer significant improvements over traditional Alternating Least Squares (ALS) for product recommendations. In particular, neural networks can understand user preferences to show more relevant products.

But the story doesn\u2019t end there. In the past few years, we\u2019ve witnessed amazing advances in AI science, which currently powers agentic products like ChatGPT or Gemini. We use the same neural networks and deep learning techniques to power Recommendations+<\/a> \u2014 our next-generation recommendation engine \u2014 enabling Loomi AI to understand each customer\u2019s journey and recommend products just for that individual customer. By leveraging a sequential understanding<\/strong> of each customer, Recommendations+ ultimately drives higher engagement and conversion. <\/p>\n\n\n\n

Read on to understand how this technology will truly revolutionize personalization. <\/p>\n\n\n\n

The Power of Sequence: Moving Beyond Static Recommendations<\/h2>\n\n\n\n

Imagine a shopper is browsing for a new coffee maker. With ALS, recommendations might be based on their general purchase history and users with similar overall profiles. However, with a sequential model, we understand the entire journey. Did they just view several pour-over coffee makers? Did they recently search for coffee beans? Are they now looking at kettles? A sequential 2TNN, unlike ALS, recognizes these patterns and adjusts recommendations in real time, suggesting complementary items like filters, grinders, or even coffee mugs that align with their current shopping sequence.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Here\u2019s how sequential 2TNNs, particularly those leveraging transformer architectures, outperform ALS in capturing this dynamic user behavior:<\/p>\n\n\n\n

Contextual Understanding, Not Just Similarity <\/h3>\n\n\n\n

Unlike ALS, which recommends items based on overall similarity without considering interaction order, sequential models leverage the sequence of user actions to predict what the user is most likely to click or purchase next. They don\u2019t just react to the latest interaction but understand the broader context, such as the user’s browsing path, transitions between items, and evolving intent. <\/p>\n\n\n\n

This allows the recommendation engines to guide users toward conversions by surfacing recommendations that align with their current journey, rather than just suggesting similar products. This journey-aware approach is critical for effective personalization.<\/p>\n\n\n\n

Capturing High-Order Dependencies With Transformers <\/h3>\n\n\n\n

The sequential models used in 2TNN can learn high-order dependencies across a sequence of interactions. This means they\u2019re not limited to immediate past actions but can incorporate long-term dependencies and complex behavioral patterns. <\/p>\n\n\n\n

Additionally, transformers feature a self-attention mechanism that weighs interactions differently based on their relevance, which helps capture both local and global dependencies effectively. This flexibility leads to more nuanced recommendations that account for subtle shifts in user behavior.   <\/p>\n\n\n\n

\"Overview<\/figure>\n\n\n\n

Data Efficiency and Granular Personalization<\/h3>\n\n\n\n

2TNN benefit from increased data as they can extract more patterns with a non-linear capacity of neural layers. With larger datasets, they become better at distinguishing subtle differences in user preferences (i.e., more granular). They also learn to recommend items based on the specific context of a user’s recent sequence of interactions, such as browsing history or purchase patterns. <\/p>\n\n\n\n

The non-linear nature of neural networks, especially in the transformer architecture, allows 2TNNs to extract richer insights from user data. The more data, the more refined and personalized the recommendations become, moving beyond broad categories to truly individual preferences.<\/p>\n\n\n\n

Visualizing the Difference: Retrieval Processes Compared<\/h2>\n\n\n\n
\"Comparison<\/figure>\n\n\n\n

The image above visually illustrates the key difference in how ALS and our Sequential Two-Tower Neural Network retrieve recommendations.<\/p>\n\n\n\n

Imagine a customer leaving a trail of breadcrumbs as they browse. ALS picks up all those breadcrumbs and tries to find other customers who left similar trails overall. It then suggests products based on what those similar customers interacted with in the past. Think of it as a broad, general approach, like recommending a coffee maker to someone who once bought coffee beans, regardless of whether they’re currently looking at espresso machines or tea kettles. It’s useful for understanding general preferences, but it misses the nuances of a specific shopping journey.<\/p>\n\n\n\n

Now imagine that same customer leaving those breadcrumbs, but this time, you’re following their trail step by step. You see them first looking at French presses, then browsing different coffee bean types, and finally, they pause on a page with milk frothers. The 2TNN acts like that attentive observer. It understands the sequence of actions and recognizes the evolving intent.  Instead of just recommending any coffee-related product, it understands the context of the current session and suggests something highly relevant, like a specific type of milk or a cleaning kit for their potential French press. It’s all about understanding the journey and providing personalized recommendations in real time that align with the customer’s immediate needs.<\/p>\n\n\n\n

How Bloomreach Drives Impact for Ecommerce Businesses<\/h2>\n\n\n\n

Bloomreach Engagement is the only autonomous marketing platform that incorporates the latest AI techniques, such as sequential modeling via transformers, to power personalized product recommendations across all channels. What\u2019s more, we make this advanced AI easily accessible for non-technical marketers so they can still deliver campaigns that improve business metrics. <\/p>\n\n\n\n

These AI innovations have culminated in our release of Recommendations+<\/a>, a next-generation recommendation engine that incrementally improves product click-through rates by 9% or more. Recommendations+ can help marketers: <\/p>\n\n\n\n