Introduction
While most recommender systems focus on predicting what new item a user might like, this paper tackles a more fundamental "whether" question: will a user repeat a recent behavior right now? Short-term reconsumption (STREC) behaviors—such as looping a favorite song, revisiting a local cafe, or repeatedly opening the same mobile app—constitute a massive portion of our daily digital and physical footprints.
Predicting whether the next action will be a reconsumption of the "near past" or a brand-new interaction is crucial for system optimization. Solving this "switch" problem allows systems to decide whether to cache a webpage, offer a loyalty coupon, or pre-load a mobile app to save battery and latency. As shown in Figure 1, the prevalence of these behaviors varies significantly across domains but remains a substantial factor in user interaction patterns.
The STREC Framework and Features
The authors propose a framework based on a sliding window ($W_k$) that tracks a user's $k$ most recent transactions. They identify four "generic," domain-independent features that influence STREC behaviors:
- Item Popularity (IP): High-quality or popular items are more likely to be repeated.
- Item Reconsumption Ratio (IRR): Measures the inherent "repeatability" of specific items (e.g., a hit song vs. a news article).
- User Reconsumption Ratio (URR): Captures individual differences between "creatures of habit" and constant explorers.
- Window Repeat Ratio (WRR): Measures the density of repeats within the current window, capturing the "rich get richer" effect where recent repeats trigger more repeats.
Efficient Prediction Algorithms
To turn these features into a prediction, the authors developed two algorithms designed for high-traffic, real-world deployment:
- ST-L (Linear Method): Uses a hyperplane to separate "repeat" points from "novelty" points in a 4D feature space.
- ST-Q (Quadratic Method): Uses a hypersphere (ellipsoid), motivated by the observation that if any single feature is high enough (e.g., a very repetitive user), a reconsumption becomes highly likely regardless of other factors.
Crucially, because these features can be updated incrementally as the sliding window moves, the online prediction complexity is $O(1)$.
Experimental Evaluation
The study evaluated its methods on four diverse datasets:
- Lastfm: Music listening records (~17M transactions).
- BrightKite: Location check-ins (~4.7M transactions).
- Gowalla: A sparse location-based social network (~6.4M transactions).
- ManicTime: A new dataset of desktop application usage logs from 44 college students.
Key Results
The proposed ST-L and ST-Q methods consistently outperformed standard baselines like Support Vector Machines (SVM). As illustrated in Figure 4, the models maintain high accuracy across various window lengths.
An interesting observation is the relationship between class balance and accuracy. Accuracy tends to be highest when the STREC ratio is very high or very low, and dips slightly as the problem becomes more balanced (closer to a 50/50 split of repeats vs. new items). For instance, on Lastfm, ST-L reached 87.99% accuracy, while on the ManicTime app dataset, it achieved 80.90%.
Conclusion
By focusing on the "whether" as a gatekeeper, this research provides a fast, domain-independent method for filtering user intent. If a system predicts a repeat, it only needs to search the small set of items in the recent window to find the "what," drastically narrowing the search space for recommendation engines. Future work will look toward integrating these "whether" predictions with specific item recommendation models and exploring how user habits evolve over longer time horizons.
Summary updated