TL;DR
This paper proposes an interactive recommender agent for mobile commerce that uses a user's schedule and Support Vector Data Description (SVDD) to detect their "instant interest." By comparing real-time location and speed with a pre-set schedule, the agent can identify if a user is deviating from their plan, prompting them to confirm their adherence or deviation, and thus inferring their current interests.
What Problem Does This Paper Address?
In the dynamic world of mobile commerce, understanding a user's instantaneous interests is crucial for providing timely and relevant recommendations. Traditional recommender systems often rely on historical data, which may not capture the immediate needs or context of a user on the go. Factors like location, current activity, and even ambient conditions can significantly influence what a user might be interested in at any given moment. This paper tackles the challenge of inferring these "instant interests" by bridging the gap between a user's planned activities (their schedule) and their actual behavior, recognizing that deviations from a schedule often signal a shift in immediate focus or interest.
Methods
The core of the proposed system is an interactive recommender agent residing on a smartphone. It employs two main techniques:
- Schedule-Based Behavior Analysis: The agent compares the user's current geographic position (latitude, longitude) and speed against their pre-defined schedule.
- Support Vector Data Description (SVDD): SVDD is used as a one-class classification algorithm. It learns a hypersphere that encloses "adherent" behaviors (those that align with the schedule). When a new data point (user's current state) falls outside this hypersphere, it's flagged as potentially "deviated."
The process involves:
- User State Determination: The agent classifies the user's state into "normal," "emergency," or "leisure" based on their adherence to the schedule and their response to prompts.
- SVDD Classification: The agent uses SVDD to determine if the user's current behavior is adherent to their schedule.
- Interactive Confirmation: If a deviation is detected, the agent prompts the user to confirm whether they are indeed deviating or if they are still adherent. This human-computer interaction is key to refining the interest inference and gathering more data.
- Interest Inference: Based on the confirmed behavior state (adherent or deviated) and the user's context, the agent infers the user's instant interest and provides relevant information or recommendations.
The paper also discusses human-computer interaction techniques, including vibration, ringtones, and graphical interfaces, to effectively communicate with the user.
Datasets
The paper does not explicitly mention any pre-existing datasets used for training or evaluation. Instead, it describes an experiment where data was collected over 20 days from one of the authors, covering various activities like walking to parking lots, driving, parking, attending classes, and commuting home. This suggests a custom data collection approach for the experiment.
Key Insights and Ideas
- Schedule as a Baseline for Interest: The paper highlights that a user's schedule can serve as a strong baseline for understanding their expected behavior and, by extension, their likely interests. Deviations from this baseline are strong indicators of a shift in immediate focus.
- SVDD for Anomaly Detection in Behavior: SVDD is effectively used to identify deviations from "normal" or "scheduled" behavior. This is a novel application of SVDD beyond typical outlier detection, framing behavioral adherence as the "normal" class.
- The Importance of Interactive Confirmation: The reliance on user confirmation for detected deviations is crucial. It acknowledges the complexity of inferring intent and allows the system to learn from user feedback, improving accuracy and personalization over time.
- Contextualizing Interests: The paper emphasizes that instant interests are highly contextual, influenced by location, time, and planned activities. The proposed agent aims to capture this dynamic context.
- Multi-modal User Interaction: The exploration of vibration, ringtones, and graphical interfaces for user notification and interaction demonstrates a holistic approach to designing a user-friendly mobile agent.
Key Results
The conducted experiment, which involved detecting deviations in seven specific activities, yielded a total detecting accuracy of 91.43%. The error rate was reported to be lower than 10%, which the authors consider better than their expected goal. The paper also includes figures showing the cost analysis of the agent in terms of CPU usage, VSS, and RSS, indicating its feasibility on mobile platforms.
Limitations and Future Work
- Data Collection: The experiment relied on data collected over 20 days from a single author, which might limit the generalizability of the findings.
- Generalizability of SVDD Parameters: The paper mentions that the choice of parameters for SVDD, particularly the Gaussian kernel's $\sigma$ value, influences the hypersphere boundaries. While specific values were used in the experiment, further research might be needed to optimize these parameters across diverse user behaviors and environments.
- Future Research: The authors plan to combine geographic location with user interest to determine different interests in changing occasions and locations. This suggests a future direction focused on more sophisticated spatio-temporal interest modeling.
Summary updated