Better Capturing Interactions Between Products in~Retail: Revisited Negative Sampling for~Basket Choice Modeling

AI summary

Overview

Research area: Machine learning for retail decision-making, specifically basket choice modeling, product embeddings, and assortment/promotion analytics.

Technical level: Intermediate. The work assumes familiarity with embedding-based models and negative sampling for training on large-scale transaction data.

Scope: The paper proposes a shopping-basket model that learns embeddings of products, prices, and stores and uses assortment-aware negative sampling to train efficiently on very large retail transaction datasets.

What This Paper Is About

Retailers need to know how products relate to one another — which items steal sales from each other and which are bought together — in order to choose assortments and plan promotions well. This paper builds a model of the shopping basket that represents products, prices, and stores as learned embeddings, so that these interactions can be recovered from sales data. The central problem it addresses is computational: with a very large catalog, training such a model is expensive, so the authors design how training examples are contrasted against one another in a way that reflects the product assortment actually available.

Key Contributions

1. A basket choice model with joint embeddings. The model learns embeddings representing interactions among products, prices, and stores rather than treating each of these in isolation.

2. A training scheme designed for very large transaction datasets. The optimization loss is computed using random negative samples, a choice motivated by the computational bottlenecks that arise when the number of items is large.

3. Assortment-aware negative sampling. Instead of sampling negatives arbitrarily, the paper draws them based on the actual assortment of products available, which the authors present as the key methodological distinction from prior work.

4. Validation in two settings. Experiments on synthetic data compare the approach against methods from the literature, and the model is trained and evaluated on millions of transactions from Fortenova Group, a leading retailer in Southeast Europe.

Main Findings

• Assortment-based negative sampling performs better. On synthetic data, the authors report that drawing negative samples according to the actual available assortment yields better results than approaches from the literature. The abstract does not state the size of the gains or the specific baselines used.

• The approach scales to industrial-scale data. The model was trained and evaluated on a dataset of millions of transactions from Fortenova Group, demonstrating feasibility on real retail data rather than only synthetic benchmarks.

• Product, price, and store interactions are learnable jointly. The abstract claims the embedding approach captures interactions between products, prices, and stores in a single model, which is what enables the downstream retail applications.

• Practical tooling for category managers. The paper presents the model as having promising applications in retail, including enriched interfaces intended to efficiently support category managers. The abstract does not describe those interfaces in detail.

Methodology in Plain English

The researchers build a model of the shopping basket that, in effect, converts every product, price level, and store into a compact numerical representation (an embedding). Relationships between these representations are what allow the model to express that two products compete with each other, or that certain products tend to be bought together.

Training such a model requires contrasting real observed purchases against purchases that did not happen. Checking every possible alternative product would be far too slow with a large catalog, so the model learns from a random sample of alternatives instead — a standard trick for scaling up this type of learning. The specific twist here is which random alternatives are chosen: the authors sample them from the assortment of products actually available rather than from the full catalog, so the comparisons the model learns from reflect realistic retail conditions.

They first test the idea on synthetic data, where they can compare against existing methods, then train and evaluate on real transaction data from Fortenova Group to check that it holds up in practice. The abstract notes that the method is designed to uncover sales patterns from very large transaction datasets but does not give architectural details, hyperparameters, or evaluation metrics.

Why This Matters

Impact on research: Negative sampling is a routine component of large-scale embedding models, and this paper argues that how negatives are drawn matters — specifically, that grounding them in the available assortment is better than generic sampling. That is a transferable idea for anyone training contrastive or embedding-based models over very large item catalogs, and it connects machine learning methodology to the practical constraints of retail data.

Real-world applications:

• Assortment optimization — choosing the mix of products a store should carry, using the model's representation of which products cannibalize one another.
• Promotion planning — identifying complementary products that are frequently purchased together so that promotions can be designed around them.
• Inventory and availability decisions — ensuring complementary items are stocked conjointly across stores rather than one of the pair going out of stock.
• Category manager decision support — the paper explicitly frames the model as feeding "enriched interfaces" that help category managers work more efficiently.

Industry relevance: The work is validated on data from Fortenova Group, one of the leading retail companies in Southeast Europe, which signals that the method was built with brick-and-mortar retail deployment in mind rather than as a purely academic exercise. The problems it targets — cannibalization, complementary product availability, regional and seasonal variation — are core to retailer strategy.

Future Directions

• Quantifying the advantage over prior methods. The abstract claims better results than approaches from the literature on synthetic data, but it does not report how large that advantage is or which baselines were used — a natural follow-up is a fuller empirical comparison.

• Evaluating downstream business outcomes. The model is described as supporting assortment optimization and promotion planning, but the abstract stops at training and evaluating the model itself. Testing whether it actually improves assortment or promotion decisions is the next step.

• Better understanding the role of negative sampling choices. The paper's main lever is how negatives are drawn; open questions remain about how sensitive the results are to that design and whether other assortment-aware sampling strategies could do better.

• Extending beyond products, prices, and stores. The abstract names these three as the embedded entities. Whether additional retail signals — such as seasonality, regional variation, or customer segments, which are mentioned as motivation — could be incorporated is left open.

• Deployment and usability. The "enriched interfaces" for category managers are mentioned as a promising application but not described in the abstract; how these models are surfaced to non-technical users remains to be detailed.

Target Audience

This paper is most useful to machine learning researchers and practitioners working on recommendation, ranking, or embedding models over large item catalogs, and to data scientists in retail who deal with transaction-level data. It also speaks to retail analysts, category managers, and merchandising strategists interested in how product interaction modeling can inform assortment and promotion decisions. Readers should have some familiarity with embeddings and negative sampling to get the most from it; the retail framing itself is accessible to a business-oriented audience.

Authors’ abstract

Brick-and-mortar retailers face many different challenges that involve understanding thoroughly its products catalog and customer preferences. In particular, assortment optimization - proposing the ideal mix of products - and promotion planning hold a pivotal role in their strategy. By leveraging sales data, retailers can make informed decisions on which products to sell and how to manage inventory, based on customer preferences as well as regional and seasonal trends. It is especially crucial to capture interactions between products, in order to minimize the number of items that cannibalize each other's sales and to ensure that complementary products, which are often purchased together, are conjointly available and sold across all stores. In this paper, we propose a model of shopping basket that learns embeddings to represent interactions between products, prices and stores. Our model is built to uncover sales patterns from very large transaction datasets. In particular, the optimization loss is computed with random negative samples in order to overcome the computational bottlenecks that arise with large number of items. Our experiments on synthetic data show the efficiency of drawing such negative samples based on the actual assortment of available products, with better results than approaches from the literature. We also validate our approach by training and evaluating our model on a dataset composed of millions of transactions from Fortenova Group, one of the leading retail company in Southeast Europe. Our model showcases promising applications in the sector of retail, with enriched interfaces to efficiently support category managers.