Source-linked AI summary
Secure Artificial Intelligence of Things for Implicit Group Recommendations
Keping Yu, Zhiwei Guo, Yu Shen, Wei Wang, Jerry Chun-Wei Lin, Takuro Sato
TL;DR
Existing group recommender systems lack secure data management workflows and often ignore implicit feedback. SAIoT-GR combines a secure IoT platform with collaborative Bayesian inference and a non-cooperative game, and experiments report performance gains over several baselines.
Problem
Group recommender systems commonly rely on explicit feedback and lack secure online data management for implicit-feedback settings.
Method
SAIoT-GR combines a secure IoT structure with a collaborative Bayesian network and non-cooperative game for group recommendation.
Results
SAIoT-GR is reported as about 4% better than RandGroup, 6% than FreGroup, 7% than ContextMF, 12% than ConfiMF, and 15% than Frequency.
Takeaways & Limitations
The proposed architecture integrates secure data management with algorithms for inferring implicit preferences and generating group recommendations.
Abstract
from arXiv · showhide
The emergence of Artificial Intelligence of Things (AIoT) has provided novel insights for many social computing applications such as group recommender systems. As distance among people has been greatly shortened, it has been a more general demand to provide personalized services to groups instead of individuals. In order to capture group-level preference features from individuals, existing methods were mostly established via aggregation and face two aspects of challenges: secure data management workflow is absent, and implicit preference feedbacks is ignored. To tackle current difficulties, this paper proposes secure Artificial Intelligence of Things for implicit Group Recommendations (SAIoT-GR). As for hardware module, a secure IoT structure is developed as the bottom support platform. As for software module, collaborative Bayesian network model and non-cooperative game are can be introduced as algorithms. Such a secure AIoT architecture is able to maximize the advantages of the two modules. In addition, a large number of experiments are carried out to evaluate the performance of the SAIoT-GR in terms of efficiency and robustness.
I. INTRODUCTION
Group recommender systems face missing secure, real-time data management and limited handling of implicit feedback. SAIoT-GR combines a secure IoT architecture with AI algorithms and is evaluated for efficiency.
- Challenges: Existing group recommender systems lack real-time secure data management and largely ignore implicit feedback such as clicks and purchases.Explicit feedback directly reflects preference degree, whereas implicit feedback expresses it indirectly through interactions.
- Proposed approach: The hardware module provides an IoT support platform for data management and security, while the software module uses a collaborative Bayesian network and non-cooperative game.The collaborative Bayesian network infers hidden preference features from implicit feedback.
- Proposed approach: SAIoT-GR addresses these challenges through coordinated hardware and software modules.The proposed architecture fuses a secure IoT framework with AI algorithms.
- Evaluation: Experiments on real-world datasets evaluate the efficiency of SAIoT-GR.The paper describes experiments as evidence for the proposed system's efficiency.
II. SYSTEM MODEL
SAIoT-GR combines a secure four-layer IoT platform with an embedded recommendation model. Its data, security, and processing layers support real-time group recommendation from heterogeneous information.
- Architecture: The SAIoT-GR architecture combines a secure IoT platform with an embedded collaborative Bayesian network algorithm.The platform has hardware and software modules, with the algorithm embedded in the processing layer.
- Architecture: The four hardware layers are data, security, processing, and application, jointly supporting group recommendations.The processing layer directly implements recommendations by carrying the embedded AI algorithm.
- Hardware Architecture: The data layer collects user profiles, item attributes, business records, and group information for real-time acquisition and updates.
- Hardware Architecture: The security layer transmits data, authenticates heterogeneous sources, and provides distributed storage for source data.
- Software Architecture: The processing layer uses collaborative inference for unknown preference feedbacks and a non-cooperative game to produce group recommendations.
B. Workflow of SAIoT-GR
The SAIoT-GR workflow first represents item content with topic indicators, then infers preference features and computes group recommendations. Its two main components are collaborative inference and recommendation.
- Workflow: The model represents users, items, and interaction records, assigning each item a latent topic indicator.User demand is modeled through inherent interest and social influence.
- Workflow: The workflow contains collaborative inference and recommendation modules, preceded by Twitter-LDA topic assignment for items.Collaborative inference uses a CBN model and Gibbs sampling, while recommendation uses a non-cooperative game.
A. Collaborative Bayesian Network
The collaborative Bayesian network models binary user-item interactions through topic contribution, inherent interest, and social influence. Its latent preference variables are learned by minimizing a probabilistic objective with stochastic gradient descent.
- Interaction model: The interaction probability is modeled from an item's topic, topic contribution rate, user inherent interest, and social influence.The same factorization supports modeling both positive and negative interaction outcomes.
- Latent preferences: The model defines latent vectors for each user's inherent interest and social influence across topics.These variables are represented as topic-indexed vectors I(i,d) and S(i,d).
- Interaction model: The topic contribution rate is based on a user's total interactions and interactions associated with the topic.The supplied formulation defines these quantities through m_i and m_i^(d).
- Learning: The learning objective minimizes the negative log probability of observed interactions under the generative model.The objective is derived from the joint probability and conditional interaction probabilities.
- Learning: Stochastic gradient descent infers the unknown inherent-interest and social-influence variables.
B. Recommendation
SAIoT-GR models group recommendation as a non-cooperative game in which members select topic strategies and receive utility-based allocations at equilibrium.
- Group members act as players, choose topic indicators as strategies, and receive utilities associated with their selected strategies.
- The model normalizes individual preference and social-confidence values before computing each member’s profit and utility.
- At Nash equilibrium, members cannot increase utility by changing strategies, and the resulting topic allocation is recommended to the group.
- Recommendations are represented as ratios of items assigned to each latent topic indicator rather than as specific individual items.
A. Datasets
The evaluation constructs group-recommendation datasets from Last.fm and Delicious individual-recommendation data, filtering users and enforcing a minimum social density.
- The study randomly forms user groups from benchmark datasets because publicly available datasets specifically for group recommendation are unavailable.
- Last.fm provides an online music-community dataset containing user interests, tags, and social-community participation.
- Delicious provides a bookmark-sharing and social-platform dataset with user tags and community participation.
- Inactive users are removed and each group must meet a minimum social density of 0.25 to reduce sparsity and ensure social-data richness.
B. Experimental Settings
Experiments assess recommendation efficiency and robustness by comparing predicted with real topic distributions and varying parameter combinations against implicit-feedback baselines.
- Recommendation precision is measured by the distance between real and predicted topic-proportion distributions, while robustness is assessed across varied parameter groups.
- Five implicit-feedback recommendation methods serve as baselines: Frequency, ConfiMF, ContextMF, RanGroup, and FreGroup.
- The experiments set dataset-specific Gaussian means and variances, with learning rate 0.01 and convergence threshold 0.001 for SAIoT-GR.
C. Results and Analysis
SAIoT-GR is evaluated against multiple baselines using topic-distribution distances and robustness experiments. It generally outperforms the baselines, although it is not best under every distance criterion.
- SAIoT-GR produces the smallest distance measurement among six methods in the topic-distribution comparisons.The comparisons cover Last.fm and Delicious; the figures contrast real and predicted topic distributions.
- SAIoT-GR is about 4% better than RandGroup, 6% better than FreGroup, 7% better than ContextMF, 12% better than ConfiMF, and 15% better than Frequency overall.TABLE I and TABLE II report six evaluation metrics across six experimental methods.
- SAIoT-GR does not achieve the best result for every metric, including Chebyshev distance on Last.fm.The paper attributes this variation to differences among distance criteria and uncertainty from unsupervised learning.
- Additional stability experiments vary parameters to evaluate SAIoT-GR alone rather than compare it with baseline methods.These experiments use Euclidean distance as the representative metric.
V. CONCLUSION
The paper proposes SAIoT-GR, a secure AIoT group recommender for implicit feedback, combining an IoT support platform with CBN and non-cooperative-game algorithms. It evaluates the system’s efficiency and robustness and identifies cold-start groups as a future extension target.
- SAIoT-GR combines a secure IoT framework with AI algorithms to address online data management and implicit-feedback challenges in group recommendation.Its design includes an IoT support platform, a collaborative Bayesian network, and a non-cooperative game.
- SAIoT-GR is intended to support group recommendations when group preferences are represented through implicit feedback rather than explicit ratings.The paper frames implicit feedback as common in real-world settings and insufficiently handled by existing approaches.
- The proposed system is evaluated on efficiency and robustness through experiments using real-world datasets.The conclusion describes experiments assessing both aspects.
- The authors identify explainable recommendations for cold-start groups as a future direction for SAIoT-GR.Cold-start groups lack enough explicit ratings or implicit feedback to clarify preferences.