- A new session-based recommendation system has been developed and implemented, utilizing item graphs, metadata, and a LightGBM model for personalized suggestions.
- The system can generate up to 20 recommendations from 300 candidates, significantly improving the relevance of suggestions for anonymous visitors.
- The evaluation approach was revised, acknowledging that the absence of a purchase record does not always imply no purchase, leading to more accurate quality measurement.
PROSTO24: Session-Based Recommendation System Driven by Item Graphs and LightGBM
PROSTO24 reports on the development and successful deployment of an innovative session-based recommendation system designed to provide personalized product suggestions to anonymous website visitors. This system, built on Coveo data, employs advanced machine learning techniques, including item relationship graphs, product descriptions, and images, along with a powerful LightGBM model for ranking.
How the System Works: From Data Collection to Recommendations
The primary goal of the system is to offer relevant items to a visitor who has just viewed a product page, without any prior information about their past visits. The recommendation generation process involves several key stages:
- Candidate Generation: Based on the viewed item, the system generates up to 300 potential candidates. This is achieved by drawing from various sources: a historical graph of item relationships, analysis of similar descriptions, and image data.
- LightGBM Ranking: The generated candidates are then ranked using the LightGBM model, which selects up to 20 of the most relevant recommendations. This stage accounts for factors such as source quotas, missing features, and pairwise training.
- Availability: The trained model, with its measured quality, is already operational, delivering recommendations via a local HTTP API. Launch instructions and reproduction guidelines are available.
Rethinking Metrics: Why No Purchase Record Doesn’t Mean No Purchase
A crucial aspect of the development involved a change in the approach to evaluating the recommendation system’s quality. Initially, 60,000 sessions were prepared for evaluation, but the Hit@20 metric was calculated for only 748 of them. This led to a significant realization: the absence of a purchase record is not always sufficient evidence to conclude that the visitor made no purchase.
This understanding necessitated a reevaluation of the objective, data preparation, and training methods. For instance, the system now considers actions the model can observe, the origins of zero labels, and the importance of retaining candidate selection misses within the metric. These revised rules are now successfully applied in the active project.
| Parameter | Implementation Details |
|---|---|
| Candidate Sources | Item relationship graph, descriptions, images, search |
| Number of Candidates | Up to 300 |
| Ranking Model | LightGBM |
| Final Recommendations | Up to 20 |
| Quality Evaluation | Hit@20 based on 748 sessions (out of 60,000) |
| Availability | Local HTTP API, instructions, and code |
For online retailers and content platforms, the implementation of such a session-based recommendation system offers the potential to significantly enhance the relevance of suggestions, even for new or anonymous users. This can lead to increased conversion rates and an improved user experience. The reevaluation of evaluation metrics underscores the importance of deep user behavior analysis, enabling the creation of more accurate and effective recommendation algorithms.
This is fascinating how PROSTO24 leverages item graphs and LightGBM for anonymous visitor recommendations! I’m particularly interested in the ‘rethinking metrics’ section. Could you elaborate more on how the system differentiates between a user truly not purchasing an item versus the system simply not observing the purchase? Also, how does the system handle cold-start items within the item relationship graph, especially for new products with limited interaction data?