TL;DR
Researchers at Walmart Global Tech have developed a new recommender system that automatically generates event-based product carousels (like "Grills" or "Patio Furniture" for Labor Day). By applying iterative clustering to a query-click bipartite graph, the system identifies distinct "aspects" of an event and populates them with relevant products, significantly reducing the need for manual human curation.
What Problem Does This Paper Address?
In e-commerce, shopping behavior shifts dramatically during events like Valentine's Day, the Super Bowl, or seasonal transitions. Standard recommender systems typically focus on Product-to-Product (P2P) or User-to-Product (U2P) relationships. However, these fail to capture the multi-faceted nature of an event.
For instance, a "Labor Day" event isn't just one category; it encompasses diverse shopping demands ranging from outdoor grills to electronics. Currently, these "aspects" and their associated product carousels are manually curated by humans—a process that is slow, expensive, and difficult to scale as catalogs and trends evolve. The challenge in automating this lies in the lack of "ground truth" data: there is no predefined list of what constitutes an "aspect" for every possible event.
Methods
The authors propose an effective recommender system based on a query-click bipartite graph.
- Graph Construction: They build a graph where one set of nodes represents search queries containing event-related keywords (e.g., "valentines day cards") and the other set represents product-types (e.g., "Greeting Cards"). Using product-types instead of individual items helps reduce data sparsity and increases the model's generality.
- Iterative Clustering (IC): Because the query-click data is sparse, a single round of clustering often fails to group similar queries that don't share the exact same clicks. The authors use an iterative agglomerative hierarchical clustering approach. In each round, they:
- Create feature vectors for query and product-type clusters based on aggregated clicks and impressions.
- Merge the most similar clusters.
- Recompute feature vectors for the next round.
- Carousel Generation: Once the query clusters (representing event aspects) are formed, the system identifies the most relevant product-types for each cluster by ranking them based on their Click-Through Rate (CTR).
Datasets
The model was evaluated using real-world traffic data from Walmart.com:
- Valentine’s Day 2020: ~693k queries and 2,528 unique impression product-types.
- Summer Water Activity 2021: ~34k queries and 821 unique impression product-types.
Human experts provided ground truth labels by identifying specific product-types that should belong to these events to measure the model's precision.
Key Insights and Ideas
- Product-Types over Products: The decision to cluster at the product-type level is a crucial engineering choice. It bypasses the "cold start" problem for individual new items and makes the graph dense enough for meaningful clustering.
- Iterative Refinement: The "aha moment" is that iterative clustering acts as a bridge. By merging nodes and re-calculating similarities, the model discovers latent relationships between queries that a standard one-pass algorithm (like DBSCAN or basic Hierarchical Clustering) would miss due to the "long tail" of search queries.
- Query Dominance: The ablation study (comparing IC-1, query-only clustering, vs. IC-2, alternating query and product clustering) revealed that clustering queries is the most impactful step. This is because e-commerce platforms have a massive number of unique queries but a relatively limited, stable set of product-types.
Key Results
The proposed Iterative Clustering (IC) models significantly outperformed baselines like standard Hierarchical Clustering (HC) and DBSCAN:
- Precision: For Valentine’s Day, the IC models achieved a precision of 71.43%, compared to just 35.71% for standard HC and 64.29% for DBSCAN.
- Robustness: The IC models remained stable across different distance thresholds ($\tau$), whereas the performance of baseline models fluctuated wildly.
- Cohesion and Heterogeneity: The system successfully balanced two opposing goals: making sure products within a carousel were similar (cohesion) while ensuring different carousels covered distinct aspects of the event (heterogeneity). For example, for the Super Bowl, it automatically generated distinct carousels for "Party Supplies," "Apparel," "Game Watching (TVs)," "Snacks," and "Decorations."
Limitations and Future Work
The authors note that while the system is highly effective at mining aspects from existing search behavior, it still relies on the initial selection of queries containing event-related keywords.
Future work could involve:
- Automated Keyword Expansion: Moving beyond simple regex-based query selection to identify event-related queries that don't explicitly contain the event name.
- Dynamic Weighting: Incorporating temporal trends more aggressively to account for how event aspects change in the weeks leading up to a holiday (e.g., moving from "decorations" to "last-minute gifts").
Summary updated