Context-aware prediction of user's first click

Liang Wu, Alvin Chin, Yuanchun Zhou, Xia Wang, Kangjian Meng, Yonggang Guo, Jianhui Li2012 · DOI

TL;DR

This paper tackles the challenge of predicting a user's immediate interest in an e-commerce app by analyzing their context (location, time, and past behavior). They propose a novel multi-label classification approach that leverages location clustering and periodic behavior mining to extract rich features, significantly improving prediction accuracy.

What Problem Does This Paper Address?

In the age of location-based services (LBS) and ubiquitous smartphones, applications collect vast amounts of user context. A key challenge is effectively using this data to predict what a user actually wants at a given moment, especially in e-commerce where understanding immediate needs is crucial for personalized recommendations. Traditional recommendation systems often overlook contextual nuances, leading to irrelevant suggestions. This paper focuses on a specific, yet important, aspect: predicting the first product type a user will click on after updating their location in a mobile e-commerce app. This "first click" is considered a strong indicator of their immediate need.

Methods

The core of the paper is a multi-label classification framework. They break down the prediction problem into binary classifications for each product type. The novelty lies in the feature engineering:

  • Location Features: Raw GPS points are sparse, so they use DBSCAN (a density-based clustering algorithm) to group nearby locations. Each cluster is then tagged with the most frequently viewed product type within it. This "cluster ID" and "type by frequency" serve as location features.
  • Time Features: Standard temporal features like hour of the day and day of the week are used.
  • Behavioral Features: This is where the most innovation occurs.
  • Switching Preference: They analyze how frequently users switch between product types, quantifying this as "switching preference." Features include the previous product type, how long it was stayed on, and average stay times.
  • Periodic Preference: They employ the Periodica algorithm (originally designed for moving object trajectories) to mine periodic patterns in user behavior. By setting a 24-hour period, they generate "periodicity templates" representing the probability distribution of product type clicks at different times of the day. These templates are then used as behavioral features.

Two standard classification models, Support Vector Machine (SVM) and K-Nearest Neighbors (KNN), are used to evaluate these features.

Datasets

The study uses a real-world dataset collected from WuXian GouXiang, a context-aware product recommender application developed by Nokia Research Center, Beijing. This dataset, collected between September 2011 and March 2012, contains logs from approximately 20,000 users and 3,000 shops. The logs include user ID, timestamp, latitude, longitude, and the product type of the first click. For analysis, they specifically used logs with GPS positions, resulting in a dataset of 262 users and 3,231 first-click behavior logs.

Key Insights and Ideas

  • First Click as a Proxy for Need: The paper strongly posits that a user's first click after updating their location is a highly indicative signal of their immediate interest, making it a valuable target for prediction.
  • Context is King (Especially Location and Behavior): While time is important, the paper demonstrates that location (even when clustered) and, more significantly, user behavioral patterns (both switching habits and periodic tendencies) are crucial for accurate predictions.
  • Periodic Behavior Mining is Powerful: Adapting the Periodica algorithm for e-commerce contexts proves effective. It captures the cyclical nature of user interests, which is often overlooked by simpler models.
  • Feature Fusion is Key: No single feature set (time/location or behavioral) is sufficient on its own. Combining all three sets yields the best performance, highlighting the synergistic nature of contextual information.
  • User Activeness Matters: The performance of different feature sets varies with user activeness. For less active users, time and location features are more robust, while for more active users, behavioral features become increasingly dominant.

Key Results

  • The proposed approach, using a combination of time, location, and behavioral features, significantly improves the prediction of a user's first click product type.
  • When using all three feature sets, the best overall F-measure was achieved.
  • Behavioral features generally outperformed time and location features when used individually.
  • SVM generally performed better than KNN across the board.
  • The model struggled with predicting rare product types (Clothes, Hotel, Entertainment), achieving near 0% precision and recall for these categories, likely due to data sparsity.

Limitations and Future Work

  • Rare Product Types: The most significant limitation acknowledged is the poor performance on predicting rare product types. The authors suggest exploring transfer learning from more common types to improve predictions for these less frequent categories.
  • Dataset Specificity: While the dataset is real-world, its effectiveness for rare types might be a general issue in such applications. Future work could involve validating the method on other datasets.
  • Context-Aware Recommendation System: The authors envision using this first-click prediction model as a core component to build a more sophisticated context-aware recommender system, moving beyond just predicting clicks to actively recommending items.
  • Monetization Opportunities: The paper highlights the potential for accurate user need estimation to unlock significant monetization opportunities for businesses, such as optimizing store placement or timing marketing campaigns.

Summary updated

Abstract

Location-based services has attracted attentions from both industry and academia. The development of position tracking technologies and the increasing popularity of smart phones has collected large amounts of contexts. An important issue is how to leverage the rich contexts to predict a user's need accurately. In this paper, we propose a novel approach to predict the product type that a user will first click in an e-commerce application, after they update their location manually. Our proposed approach models the problem as a multi-label classification. We introduce three sets of features including location feature, time feature and behavioral feature. We use the Periodica algorithm [10], which was designed to mine the periodic behaviors of moving objects, to generate a series of periodicity templates. The templates are further exploited as behavioral features. Finally, we design several experiments using a real world dataset collected by an e-commerce application called WuXianGouXiang, which is developed by Nokia Research Center, Beijing and was released in September 2011. We have obtained a dataset from the service logs and the dataset contains over 3000 registered shops and 20000 users. Our experimental results demonstrate that the three sets of features contribute significantly to the classification of different users and the best result is achieved when using all of them.
BibTeX
@misc{wu2012context,
  author = {Liang Wu and Alvin Chin and Yuanchun Zhou and Xia Wang and Kangjian Meng and Yonggang Guo and Jianhui Li},
  title = {{Context-aware prediction of user's first click}},
  year = {2012},
  doi = {10.1145/2346604.2346613},
}
Figure