MerRec, A Large-scale Dataset For The Development Of A Consumer-to-consumer (C2C) Recommendation System, A Challenge For Mercari

3 main points
✔️ Development of the large-scale dataset "MerRec": Based on data collected from Mercari, we propose a new large-scale dataset for research and development of recommendation systems for consumer-to-consumer (C2C) transactions. Facilitates the development of recommendation systems that include diverse user behaviors and product characteristics and that can respond to the unique environment of C2C marketplaces.
✔️ Developed Mercatran, a recommendation system for C2C: Developed a new model, Mercatran, designed to address the unique challenges of C2C, using the MerRec dataset for CTR prediction, session-based recommendations, and user action prediction. Evaluated model performance and utility through multi-task learning.
✔️ Contributions to Recommender Systems in eCommerce: research results bridge academic research and practical applications, and present new possibilities for future recommendation systems in eCommerce marketplaces.

MerRec: A Large-scale Multipurpose Mercari Dataset for Consumer-to-Consumer Recommendation Systems
written by Lichi Li, Zainul Abi Din, Zhen Tan, Sam London, Tianlong Chen, Ajay Daptardar
(Submitted on 22 Feb 2024)
Comments: Published on arxiv.
Subjects: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI)

code：  

The images used in this article are from the paper, the introductory slides, or were created based on them.

Summary

In the emerging era of e-commerce, recommendation systems play an important role in improving user experience and customer engagement. These systems, which provide users with what they are looking for from the vast array of products and services offered online, are the foundation of e-commerce.

And while the business-to-consumer (B2C) model has been dominant in the past, the consumer-to-consumer (C2C) model has recently been expanding and offers new possibilities. in C2C, a single user sometimes becomes a seller and sometimes a buyer, creating a dynamic C2C is a dynamic marketplace that differs from B2C in that a single user is sometimes a seller and sometimes a buyer. To adapt to this unique environment, new forms of recommendation systems are required that cannot be accommodated by the traditional B2C model.

However, no study has yet fully modeled the novelty and complexity of this C2C model. There is a large gap between actual service needs and academic research. To bridge this gap, this paper develops a new large-scale dataset, MerRec, which is useful for building C2C recommendation systems.

MerRec builds on data collected by Mercari, the largest C2C e-commerce platform, to capture the diversity of consumer behavior and preferences in detail. In addition, this dataset includes basic attributes such as user ID, item ID, and session ID, as well as detailed features such as time-stamped behavior types, product categories, and text-based product attributes, making it a valuable It is a valuable data source for deep understanding of C2C user and service characteristics.

MerRec is unique in its design, which allows it to flexibly adapt to the high fluidity of the C2C marketplace. As product listings are constantly updated, this data set enables the development of recommendation systems that continue to adapt to a changing environment. We also present a prototype model, Mercatran, which provides an initial performance benchmark using MerRec.

MerRec is a versatile dataset for a variety of tasks related to recommendation systems, measuring the performance of recommendation models through different tasks such as CTR prediction, session-based recommendations, and multi-task learning. This provides a broad validation of MerRec's utility.

MerRec provides an important data source and insight into the study of C2C recommendation systems, and is expected to further encourage the development of effective recommendation systems for e-commerce.

MERREC Dataset Overview

Mercari is an online marketplace where consumers buy and sell products to each other and where users can be both sellers and buyers. This study focuses specifically on analyzing buyer preferences in order to improve the accuracy of the recommendation system within Mercari.

When listing an item on Mercari, sellers must update detailed information about the item, including title, brand, category, image, who pays shipping, price, and condition. The seller can also change the display of the item and the information is dynamic.

Because many of Mercari's sellers are not retailers, but rather ordinary users with no specialized knowledge, there are unique challenges that are not encountered on B2C platforms. Because they are self-reported, registration information can include misrepresentation of brands and categories, incomplete product descriptions, and lack of important information such as size and color. There are also no standardized identifiers such as SKUs, making it difficult to identify products. Each product on Mercari is unique and once sold cannot be repurchased.

In addition, Mercari offers a variety of UI options, allowing users to take a variety of user actions from product discovery to purchase, including clicking, liking, adding to cart, submitting offers, and initiating and completing transactions. These interactions are indicators of user interest and provide useful information for underlying analysis in the MerRec dataset.

The MerRec dataset has been meticulously designed to capture user behavior and product characteristics on Mercari. By combining user behavior and product characteristics, we aim to gain a deep understanding of the relationship between users and products, address the challenges inherent in the C2C business model, particularly the variability of product descriptions and lack of standard identifiers, and improve the performance of recommendation systems in the unique marketplace environment.

MERREC Dataset Features

The MerRec dataset is designed to be a resource that captures the various user behaviors and product characteristics in Mercari that have been described so far, with the following perspectives

Product diversity: incorporating the wide range of product categories handled on the Mercari platform provides a wealth of data for understanding user interests and behavior.

User Behavior: This covers a wide range of user behavior, including not only final purchases, but also product views, likes, cart additions, offer applications, and more. We provide a wealth of data for detailed analysis of the entire user purchasing process.

Product details: includes detailed information on product title, category, price, condition, size, color, etc. It provides a wealth of data to detail user interests and the factors that influence the decision process.

Contextual information: Each user behavior contains contextual information such as the type and timing of the behavior. We provide a wealth of data to analyze behavior patterns over time and to understand user behavior in detail.

Currentness: we make sure to include recent data to reflect the latest trends and user interests of the platform. We provide a wealth of data to develop recommendation systems tailored to current market trends.

In addition, the MerRec dataset is built in accordance with legal and privacy regulations and with respect for ethical use and user confidentiality. Researchers and developers can use the dataset with confidence to pursue academic and practical research on recommendation systems in C2C e-commerce platforms such as Mercari.

The table below provides statistics on the features included in the MerRec dataset.

The MerRec dataset is slightly more concentrated in the Women and Toys & Collectibles categories, but overall it is balanced, covering a broad range of products and categories that are sold on Mercari. The MerRec dataset has a slight concentration in Women and Toys & Collectibles, but overall it is well balanced, covering a wide range of categories.

This and other aspects of data set composition are discussed in detail in the paper.

MERREC Dataset cleaning and processing

The MerRec dataset is subjected to the following procedures for data cleaning and processing to improve the quality of the dataset.

1. filtering users and items: eliminating suspended accounts, items that violate platform rules, etc.

2. Segmentation of sequences: Due to the long-tail distribution observed in the user's sequences, long sequences were split into shorter, fixed-length segments to standardize the data structure. Although this process facilitates analysis, accuracy is compromised, and researchers and developers are encouraged to reconstruct the original sequence as needed.

3. reduction of redundancy: Repetitions in the sequence, such as clicks on successive identical items, are eliminated. Redundancy is reduced and data is simpler.

4. privacy protection: To protect user privacy, users in certain regions are excluded in compliance with regional regulations, all ID fields are anonymized by pseudo-name, and timestamps are unified in UTC format to hide the original local time information.

5. alternative search for SKUs: As a new alternative to SKUs, we are introducing a synthetic field called "product_id" that integrates the brand and the most detailed category ID. While this is an approach that addresses product identification in situations where SKUs are not available, we must allow for the risk that in a real marketplace environment it may not fully incorporate the detail needed for effective recommendations.

Experiments and Analysis

Here we apply various machine learning and recommendation models to specific tasks using the MerRec dataset and evaluate their effectiveness and performance. Here we mention "CTR prediction" as one of the tasks addressed in this paper. This is a fundamental task in recommendation systems and involves predicting the likelihood that a user will click on an item. The prediction model is designed to predict the CTR at which an item viewing action (item_view) will occur based on user interaction and item metadata.

The study also uses a rolling window method to reconstruct the data into snapshots, allowing each model to make contextual predictions. In particular, unlike traditional CTR prediction, this experiment does not limit the types of user actions, but considers a wide variety of actions as inputs; the MerRec dataset does not include user demographic information (e.g., age, gender, ethnicity), but instead uses a rich set of item characteristics to determine user actions to understand their interests. This approach is based on the idea that it can provide more relevant signals than demographic information about how a user's interests are likely to change through actions on the platform.

The dataset is also used for multi-task learning (MTL) tasks involving a large number of unique items, users, sequences, sessions, and product IDs. Given the computational time and resource limitations, we have chosen to omit a comprehensive hyperparameter search and reduce the number of samples for the benchmark runs. The study uses the first month of MerRec's six months of data, sets the shortest input history window sequence to 7, and makes binary predictions for the eighth item; sequences equal to 8 events are a single snapshot row, and sequences longer than 8 events are treated as multiple snapshot rows using a rolling window. With this setting, sequences of less than 8 events that do not exist in MerRec are not padded to make them longer sequences.

As a result of the above condition set, this subset of CTR predictions contains 30,221,983 unique items, 2,767,956 unique users, 9,809,155 sequences, 915,453 unique product IDs, and the training, validation, and test The set is divided in an approximate 8:1:1 ratio.

The CTR prediction was performed on a Linux system on the Google Cloud Platform and the hardware used was an Nvidia T4 GPU with 8 cores and 104 GB of RAM. The performance on the test set is shown in the table below.

Tests on the MerRec dataset show that CTR prediction is challenging and that many models perform similarly under limited hyperparameter tuning. In particular, Attention FM (AFM) performs better than the other models. It is also shown that models with cross-networks may be difficult to tune or perform poorly when capturing different degrees of interaction within MerRec. These results represent how existing models can adapt to the delicate and dynamic data presented by the MerRec dataset and offer potential for future research and development of recommender systems in the C2C marketplace.

In addition to the CTR forecasting mentioned in this article, this paper also addresses session-based recommendations and multi-task learning.

Summary

In this paper, we develop MerRec, a large-scale dataset dedicated to recommendation systems for consumer-to-consumer (C2C) transactions, based on data collected from the Mercari platform. Through analysis of this dataset, we are examining the importance of recommendation systems in e-commerce and their potential in marketplaces. In particular, the Mercatran model, designed to address the unique challenges of C2C, represents a remarkable advance in the study of recommender systems.

And how the MerRec dataset and Mercatran model capture the dynamic nature of C2C transactions and potentially improve the user experience through multiple tasks, including click-through-rate prediction, session-based recommendation, and multi-task learning of user action prediction We demonstrate that the

This paper is expected to bridge academic research and practical applications, and to demonstrate the future potential of recommender systems in e-commerce.

• 
• 
• 

Categories related to this article

• Recommendation