Personalized Recommendation on Discount Coupons

Teemu Mutanen, Sami Nousiainen, Liang He2010 · DOI

TL;DR

This paper tackles the challenge of personalizing discount coupon recommendations in a real-world, high-volume coupon dispenser system in China. By leveraging customer and dispenser metadata alongside past printing behavior, they propose and evaluate various recommendation approaches to help users find relevant coupons faster, thereby reducing dispenser usage time and increasing customer satisfaction.

What Problem Does This Paper Address?

Imagine standing in front of a coupon dispenser with dozens, even hundreds, of coupons available. Finding the one you actually want can be a time-consuming and frustrating experience. This is the problem the authors address: how to make this process more efficient and user-friendly through personalization. In China, a company called Velo operates a large-scale coupon dispensing service where millions of users print coupons from physical machines. The core challenge is to help these users quickly identify and print the most relevant coupons from a vast selection, reducing their time spent at the dispenser and improving their overall experience. This is particularly important for the service provider, as faster customer throughput at dispensers can increase overall system utilization and customer satisfaction.

Methods

The authors explore several recommendation approaches, ranging from simple baselines to more sophisticated techniques:

  • Aggregate Popularity (P): A baseline method that recommends the most popular coupons across all users and dispensers.
  • Dispenser Level Popularity (P_eq): Recommends popular coupons specific to a particular dispenser location.
  • Customer's Previous Prints Filtered (P_f): Filters out coupons a customer has already printed from the aggregate popularity list.
  • Customer Segments (U and U_o): Customers are clustered into segments based on demographic and background information. Recommendations are then made based on the popularity of coupons within these segments.
  • Customer Interest Based (I and I_U): This approach uses a binary interest vector derived from customer registration data. It calculates coupon interest vectors and then recommends coupons based on the similarity between the customer's interest vector and the coupon's interest vector. An updated version (I_U) also incorporates the customer's printing history to refine these interest vectors.
  • Item-based Collaborative Filtering (item CF): A standard collaborative filtering approach that calculates item-similarity matrices and recommends coupons based on what similar items have been printed.

Datasets

The study utilizes data collected from a commercially operational Velo coupon service in China, spanning from August 1, 2009, to February 21, 2010. The dataset includes customer IDs, coupon IDs, dispenser IDs, and print records with timestamps.

Key data characteristics and challenges noted include:

  • Language Barrier: Significant portions of the data were in Chinese, requiring filtering and leading to a loss of information about coupon content and dispenser locations.
  • Customer Identification Ambiguity: While RFID tags are used, the possibility of tag sharing or one customer printing for others means the recorded print might not always reflect the actual user's preference.
  • Dispenser Variability: Dispensers have different numbers of buttons and available coupons, making it difficult to precisely know the set of coupons available at any given moment.

For their experiments, the authors filtered the transaction records and selected a subset of 10,000 customers.

Key Insights and Ideas

  • Implicit vs. Explicit Data: The study highlights that while explicit customer interests (provided during registration) are useful, implicit behavior (actual coupon printing) often reflects customer preferences more accurately, especially when sufficient historical data is available.
  • Popularity as a Strong Baseline: Simple popularity-based recommendations, especially at the dispenser level, proved to be surprisingly effective and competitive, suggesting that what's popular in a given context is a good indicator of what a user might want.
  • The Value of Context: Recommendations that consider the specific dispenser location (P_eq) performed better than global popularity, indicating that local context matters.
  • Trial Users vs. Loyal Customers: The authors explored segmenting customers, including a distinction between trial and loyal users, but found that filtering out trial users didn't significantly improve recommendation accuracy for segmentation-based methods.
  • The Challenge of "Discovery": The current evaluation metric focuses on whether the correct coupon was in the top-N. This inherently favors popular coupons. The paper acknowledges that truly novel recommendations (lesser-known coupons) are harder to evaluate with this setup and require different approaches.

Key Results

The paper evaluates recommendation methods based on the percentage of correctly predicted coupons within a top-12 list.

  • Dispenser Level Popularity (P_eq) achieved the highest accuracy at 54.6%.
  • Aggregate Popularity (P) followed closely at 48.9%.
  • Customer Interest Based (I_U), which incorporates updated interest vectors, reached 24.7%.
  • Item-based Collaborative Filtering (item CF) achieved 13.0%.
  • The study found that more training data generally improved the performance of interest-based and item-based CF methods.
  • Customer segmentation methods (U and U_o) performed similarly to each other and were generally less effective than popularity-based methods.

Limitations and Future Work

  • Popularity Bias: The authors acknowledge that their results are heavily influenced by coupon popularity. This is partly due to the evaluation setup (predicting the correct coupon from a set) and the fact that recommendations were not actually implemented in the live system, meaning customers likely printed what was readily available and popular.
  • Lack of Real-time Implementation: The study uses historical data and does not reflect a live deployment where recommendations are actively influencing user choices in real-time.
  • No Usage Feedback: The system cannot determine if a printed coupon was actually used, limiting the ability to assess the true value or "goodness" of a recommendation.
  • Future Research:
    • Focus on "Discovery": Investigating how to recommend less-known but relevant coupons and evaluating the "discovery" value for customers.
  • Actual Implementation and A/B Testing: Deploying and testing these recommendation systems in a live environment to gather empirical data and user feedback.
  • Mobile Solutions: Extending personalization to mobile coupon services, which could provide richer contextual information and direct usage feedback.
  • Contextual Factors: Exploring how factors like weekday, time of day, and specific coupon types (e.g., coffee vs. clothing) influence printing behavior.

Summary updated

Abstract

We present several recommendation approaches how discount coupon service could be personalized in order to cut down customer’s usage time. The Velo system is already implemented coupon service system in China with millions of users. The proposed approach makes use of customer and dispenser meta data together with previous user coupon prints in order to predict upcoming prints for a customer. The implementation of the service, data collection, and how the implementation affects the customers’ behavior all have an effect on the accuracy of the printed coupons. The analysis of customer characteristics help the service to present relevant coupons to a customer in order both to raise customer’s satisfaction and to aid customers on finding relevant coupons faster. The results show that item recommendation systems provide valuable benefits for coupon service. Keywords-Personalization; Discount Coupon; Recommendation Systems; User Modeling;
BibTeX
@misc{mutanen2010personalized,
  author = {Teemu Mutanen and Sami Nousiainen and Liang He},
  title = {{Personalized Recommendation on Discount Coupons}},
  year = {2010},
  doi = {10.5176/978-981-08-7466-7\_kd-18},
}
Figure