[Paper Review] Family Shopping Recommendation System Using User Profile and Behavior Data

Jiacheng Xu|arXiv (Cornell University)|Aug 24, 2017

Recommender Systems and Techniques13 references3 citations

TL;DR

This paper proposes a novel family-based shopping recommendation system that integrates user profile, behavior, and family relationship data using a hybrid collaborative filtering approach. By combining user and family-level similarities across product brand, type, and category, the Hybrid Family Model outperforms both User and Hybrid User models in recall and precision, demonstrating superior performance on a real-world shopping mall dataset.

ABSTRACT

With the arrival of the big data era, recommendation system has been a hot technology for enterprises to streamline their sales. Recommendation algorithms for individual users have been extensively studied over the past decade. Most existing recommendation systems also focus on individual user recommendations, however in many daily activities, items are recommended to the groups not one person. As an effective means to solve the problem of group recommendation problem,we extend the single user recommendation to group recommendation. Specifically we propose a novel approach for family-based shopping recommendation system. We use the dataset from the real shopping mall consisting of shopping records table, client-profile table and family relationship table. Our algorithm integrates user behavior similarity and user profile similarity to build the user based collaborative filtering model. We evaluate our approach on a real-world shopping mall dataset.

Motivation & Objective

• To address the gap in group recommendation systems by focusing on family-level shopping behavior rather than individual users.
• To develop a unified recommendation model that integrates user purchase history, activity data, and shopping mall events.
• To leverage family relationship data to improve recommendation accuracy for shared accounts or family-purchased items.
• To empirically validate the proposed model on a real-world shopping mall dataset with time-based splitting and standard evaluation metrics.

Proposed method

• The system constructs user-based collaborative filtering using similarity metrics for user behavior, profile, and activity data.
• It computes five similarity matrices: $W_{brand}$, $W_{type}$, $W_{category}$, $W_{activity}$, and $W_{profile}$ for both individual and family-level recommendations.
• The Hybrid User Model combines all five similarity types to generate top-N recommendations for individual users.
• The Hybrid Family Model extends this by computing family-level similarity matrices ($W_{brand}^{F}$, $W_{type}^{F}$, etc.) using aggregated family purchase behavior.
• The dataset is split chronologically at '2016-07-15 00:00:00', with 80% for training and 20% for testing to ensure temporal validity.
• Evaluation uses precision and recall at top-k recommendations (k = 1 to 10), with results visualized across models.

Experimental results

Research questions

• RQ1 How does integrating family relationship data improve recommendation performance compared to individual user-based models?
• RQ2 To what extent does combining user profile, behavior, and activity data enhance recommendation accuracy in a family context?
• RQ3 Can a hybrid model that fuses multiple similarity types (brand, type, category, activity, profile) outperform single-source models?
• RQ4 How does the performance of the Hybrid Family Model compare to the Hybrid User Model and the basic User Model across varying top-k recommendation thresholds?

Key findings

• The Hybrid Family Model achieves higher recall and precision than both the User Model and Hybrid User Model across all top-k values from 1 to 10.
• The recall rate for all models increases with k, indicating better coverage of user purchase history as k grows.
• The precision rate for all models decreases when k exceeds 3, suggesting diminishing returns in recommendation relevance at higher k.
• The Hybrid User Model outperforms the basic User Model in both recall and precision, confirming the benefit of multi-source similarity integration.
• The Hybrid Family Model demonstrates the strongest performance, indicating that modeling family-level behavior significantly improves recommendation quality.
• The results are validated on a real-world dataset with time-based splitting, ensuring robustness and practical relevance.