Deep Learning for Personalized Ecommerce

Sharing key learnings and insights from our Real-Time (In-Session) Personalization journey at CARS24 — a capability that has transformed how we personalize the car buying experience at scale. Leveraging advanced sequence-based neural networks and real-time Kafka streaming infrastructure, we've developed a dynamic machine learning pipeline that processes more than a million user interactions daily. Our deep learning models rapidly adapt to user behaviour, delivering personalized car recommendations with sub-200ms latency. Highlights: ✅ Advanced sequence-based neural network architecture  ✅ Real-time streaming and processing of user behaviour signals with Kafka  ✅ Rapid feature engineering and inference using optimized real-time databases  ✅ High scalability for continuous model retraining and deployment Performance Impact: 📈 Across all discovery widget we achieved a highest Impression-to-View (I2V) rate and on the 'Best Matches' recommendation rail on our car detail page and buyer home page. 📈 Delivered a strong Impression-to-Booking Initiation (I2BI) conversion rate across different discovery widgets, underscoring high user relevance and engagement. Business Outcomes: 🚀 Significant uplift in user engagement  🚀 Marked reduction in user drop-offs  🚀 Enhanced personalization and superior user experience The attached flow chart outlines the architecture behind this AI-powered personalization pipeline — from real-time clickstream ingestion to ML inference and personalized recommendations. #RealTimePersonalization #AI #MachineLearning #DeepLearning #Kafka #DataScience #RecommendationEngine #TechInnovation #AI  #Personalization #pubsub #CARS24 #transformers #llm #genai

🚀 Can AI understand what you want before you even know it? 🛍️ Most recommendation systems struggle to accurately predict user behavior because they rely on one-dimensional data. But what if we could fuse visual, textual, and graph data into a single, smarter AI model? The new Triple Modality Fusion (TMF) framework does just that — combining images, text, and user interaction patterns through Large Language Models (LLMs). Here’s why this is a game-changer: 🔍 Traditional models use only one type of data, missing out on critical insights. TMF integrates: - Visual data (item aesthetics) - Textual data (descriptions, brand details) - Graph data (user interactions) By aligning these with LLMs, TMF offers a 38% increase in recommendation accuracy on complex datasets like electronics! 💡Unlike typical recommenders, TMF uses natural language prompts to interpret user intent, modeling shopping behavior as if it were a dialogue. This context-aware approach means that even subtle cues — like a user browsing colorful bike helmets — can lead to spot-on suggestions (e.g., matching bike shorts for kids). This model isn't just theory; it's already deployed in production, showing 20% higher user satisfaction in tests. Imagine the implications for e-commerce, where understanding diverse data could personalize user experiences at scale. But here’s the big question: How far can we take multi-modality fusion in AI? Will we soon see AI predict our needs with even more accuracy? Read more about the research in the comments. 👇 #AI #MachineLearning #RecommendationSystems #ECommerce #DataScience

In e-commerce and digital apps, user representation drives personalization, segmentation, and marketing. Current approaches rely on implicit embeddings (from clicks/purchases) that are powerful but opaque, or explicit features (like category preferences) that are interpretable but shallow. The paper 𝗬𝗼𝘂 𝗔𝗿𝗲 𝗪𝗵𝗮𝘁 𝗬𝗼𝘂 𝗕𝗼𝘂𝗴𝗵𝘁: 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗻𝗴 𝗖𝘂𝘀𝘁𝗼𝗺𝗲𝗿 𝗣𝗲𝗿𝘀𝗼𝗻𝗮𝘀 𝗳𝗼𝗿 𝗘-𝗰𝗼𝗺𝗺𝗲𝗿𝗰𝗲 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 (𝗦𝗜𝗚𝗜𝗥’𝟮𝟱) introduces GPLR, a framework using LLMs to turn transaction histories into rich, human-readable personas (e.g., “Busy Parent,” “Bargain Hunter”)—making user profiles both interpretable and actionable. GPLR generates interpretable persona labels and descriptions (e.g. “DIY Enthusiast – purchases tools, building supplies, home improvement gear.”) , uses LLMs to annotate a small set of users with these personas, and then scales to millions of users cost-effectively by propagating labels through a random walk method (RevAff). 𝗛𝘆𝗯𝗿𝗶𝗱 𝗮𝗽𝗽𝗿𝗼𝗮𝗰𝗵: Personas and embeddings can be used together in a hybrid approach to make personalization actionable, explainable and scalable. Embeddings capture 𝗳𝗶𝗻𝗲-𝗴𝗿𝗮𝗶𝗻𝗲𝗱, 𝗲𝗽𝗵𝗲𝗺𝗲𝗿𝗮𝗹 𝘀𝗶𝗴𝗻𝗮𝗹𝘀 that change quickly with context and are great for micro-personalization. Personas are more stable and capture 𝗿𝗲𝗰𝘂𝗿𝗿𝗶𝗻𝗴 𝗹𝗼𝗻𝗴-𝘁𝗲𝗿𝗺 𝗵𝗮𝗯𝗶𝘁𝘀 or need. They are useful for segmentation and macro-personalization. 1. 𝗘-𝗰𝗼𝗺𝗺𝗲𝗿𝗰𝗲 𝗰𝘂𝘀𝘁𝗼𝗺𝗲𝗿 𝗿𝗲𝘁𝗲𝗻𝘁𝗶𝗼𝗻 𝗲𝘅𝗮𝗺𝗽𝗹𝗲: Customers churn if recommendations feel irrelevant or impersonal. Member: Dave Personas: Outdoor Adventurer (0.9), DIY Enthusiast (0.7) Embedding signals: Recently browsed winter jackets and power drills. Hybrid retention action: Embedding detects browsing of power drills and personas add context that Dave is DIY-focused. Personalized seasonal campaign: “Adventure-ready camping gear for fall”. Send cross-sell offer: “Outdoor toolkit bundle for DIY explorers.” Recommender System: Suggests camping-ready multi-tools + DIY accessory kits. Result: Cross-sell aligns with both immediate interest (drill) and stable persona (DIY). Dave feels understood and returns for repeat purchases. 2. 𝗛𝗲𝗮𝗹𝘁𝗵𝗰𝗮𝗿𝗲 𝗺𝗲𝗺𝗯𝗲𝗿 𝗮𝗰𝗾𝘂𝗶𝘀𝗶𝘁𝗶𝗼𝗻 𝗲𝘅𝗮𝗺𝗽𝗹𝗲: Members disengage when care feels generic. Member: Carol Personas: Family Caregiver (0.9), Preventive Health Seeker (0.7) Embedding signals: Recently searched “flu shot near me” and read “stress management for parents.” Hybrid retention action: Embedding flags interest in stress relief and preventive care and personas emphasize stable Family Caregiver role. Recommender System: Personalized suggestions for “Flu vaccine appointment for your family” + “Mindfulness sessions for caregivers.”    Result: Carol gets timely, relevant and personalized suggestions via proactive outreach which makes her feel supported. Link: https://lnkd.in/eu6k_bGd

Build an e-commerce real-time personalized recommender using the 4-stage architecture. (It's simpler than you think!) The problem with real-time recommenders is that you must narrow from millions to dozens of item candidates in less than a second while the items are personalized to the user! The 4-stage recommender architecture solves that! Let's understand how to implement it for e-commerce products using an AI lakehouse, such as Hopsworks The data flows in 2 ways: An 𝗼𝗳𝗳𝗹𝗶𝗻𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 that computes the candidate embeddings and loads them to a vector index in Hopsworks (offline batch mode). It leverages the Items Candidate Encoder Model to compute embeddings for all the items in our database. . An 𝗼𝗻𝗹𝗶𝗻𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 that computes the actual recommendations for a customer (batch, async, real-time or streaming mode). The online pipeline is split into 4-stages (as the name suggests), starting with the user's requests and ending with the recommendations: 𝘚𝘵𝘢𝘨𝘦 1/ Take the customer_id and other input features, such as the current date, compute the customer embedding using the Customer Query Model and query the Hopsworks vector DB for similar candidate items. 𝘙𝘦𝘥𝘶𝘤𝘦 𝘢 𝘤𝘰𝘳𝘱𝘶𝘴 𝘰𝘧 𝘮𝘪𝘭𝘭𝘪𝘰𝘯𝘴 𝘰𝘧 𝘪𝘵𝘦𝘮𝘴 𝘵𝘰 ~𝘩𝘶𝘯𝘥𝘳𝘦𝘥𝘴. 𝘚𝘵𝘢𝘨𝘦 2/ Takes the candidate items and applies various filters, such as removing items already seen or purchased using a Bloom filter. 𝘚𝘵𝘢𝘨𝘦 3/ During ranking, we load more features from Hopsworks' feature store describing the item and the user's relationship: "(item candidate, customer)." This is feasible as only a few hundred items are being scored, compared to the millions scored in candidate generation. The ranking model can use a boosting tree, such as XGBoost or CatBoost, a neural network or even an LLM. 𝘚𝘵𝘢𝘨𝘦 4/ We order the items based on the ranking score plus other optional business logic. The highest-scoring items are presented to the user and ranked by their score. 𝘙𝘦𝘥𝘶𝘤𝘦 𝘵𝘩𝘦 ~𝘩𝘶𝘯𝘥𝘳𝘦𝘥𝘴 𝘰𝘧 𝘤𝘢𝘯𝘥𝘪𝘥𝘢𝘵𝘦𝘴 𝘰𝘧 𝘪𝘵𝘦𝘮𝘴 𝘵𝘰 ~𝘥𝘰𝘻𝘦𝘯𝘴. . All these recommendations are computed in near real-time (in milliseconds). As you interact with the platform, you create new features that modify the customer embedding, creating a new list of candidates based on your latest preferences. . Consider reading the first lesson in Decoding ML . For free: 🔗 https://lnkd.in/ds3rJBe5 #machinelearning #artificialintelligence #mlops #datascience

For years, true personalization in ecommerce felt out of reach, too complex, too reliant on massive data infrastructure But in 2025, it’s not just possible, it’s expected * Customer Data Platforms (CDPs) can now unify behavioral, transactional, and anonymous data to recognize visitors in real-time and dynamically segment audiences. * Generative AI builds on that foundation, automating hyper-personalized product recommendations, emails, and even entire storefronts tailored to browsing habits, purchase history, and preferences * Today’s ecommerce personalization means: individualized landing pages, AI chat that understands customer intent, and product suggestions that evolve with each click Brands are no longer optimizing for demographics, they’re creating a “segment of one” The results? Higher conversion rates, deeper customer retention, and a distinct competitive advantage But unlocking this requires more than tech; it demands a strategic approach to data, tools, and team readiness Are you leveraging personalization as a growth engine?