Multiple User Context Inference by Fusing Data Sources

Shangguang WangarXiv 2017 · arXiv

TL;DR

This paper tackles the challenge of inferring user context (like age and gender) by combining data from multiple sources (call records and data traffic) and accounting for the interdependence between different context attributes. Their novel approach, MulAProM, fuses features using tensor outer products and then models these attributes jointly, outperforming existing methods.

What Problem Does This Paper Address?

Accurately inferring user context information, such as gender, age, and location, is crucial for building personalized services and recommender systems. However, existing methods often suffer from two main limitations:

  1. Reliance on Single Data Sources: Most studies focus on just one type of data, like call detail records (CDR) or internet browsing logs. This overlooks the complementary information that different data sources can provide, limiting inference accuracy.
  2. Ignoring Context Interdependence: User attributes are not independent. For example, a person's age might influence their communication patterns, which in turn might be related to their gender. Previous work often infers attributes in isolation, missing these crucial interdependencies and leading to suboptimal performance.

Furthermore, collecting explicit user context is difficult due to privacy concerns and technical limitations (e.g., browser restrictions). Therefore, robust inference methods are essential.

Methods

The authors propose a novel framework with several key components:

  • Feature Extraction: They extract features from two distinct data sources:
  • CDR (Call Detail Records): Features capture communication patterns, including call frequency, duration, and communication partners' demographics and age groups.
  • DTR (Data Traffic Records): Features capture data usage patterns, such as frequency and duration of data usage at different times of the day.
  • Data Source Fusion (Tensor Outer Product): To leverage complementary information from CDR and DTR, they fuse the extracted features. Instead of simple concatenation, they use a tensor outer product in the feature space. This method captures interactions between features from different sources more effectively than concatenation.
  • Dimensionality Reduction (PCA): The tensor outer product can lead to a high-dimensional feature space. To combat the "curse of dimensionality" and remove redundancy, they apply Principal Component Analysis (PCA).
  • Multiple Attribute Probabilistic Inference Model (MulAProM): This is the core of their proposed model. MulAProM is based on the Maximum Entropy Principle and is designed to simultaneously infer multiple user context attributes (specifically gender and age). By modeling these attributes jointly, it explicitly captures their interdependence, leading to improved inference.
  • Gaussian Prior: To mitigate overfitting, especially with potentially sparse training data, they introduce a Gaussian prior on the model parameters.

Datasets

The study utilizes a real-world telecommunication dataset from a local mobile operator in Shanghai, China. This dataset comprises:

  • 4.6 million Call Detail Records (CDR)
  • 7.5 million Data Traffic Records (DTR)

These records were collected from 8,000 mobile users over a period of six months. The authors filtered out users with missing context information or insufficient activity to ensure data quality.

Key Insights and Ideas

  • Fusion is Key: Combining information from CDR and DTR significantly improves user context inference performance compared to using either source alone. This highlights the value of multi-source data fusion.
  • Interdependence Matters: Explicitly modeling the relationship between user attributes (like gender and age) is crucial for accurate inference. Jointly inferring these attributes captures nuances missed by independent inference.
  • Tensor Outer Product for Fusion: This technique offers a more sophisticated way to combine features from different data sources than simple concatenation, effectively capturing cross-source interactions.
  • PCA for Efficiency: While fusion creates rich features, PCA is essential for managing dimensionality and preventing overfitting, ultimately improving model performance.
  • Age Inference is Harder: The experiments suggest that inferring user age is generally more challenging than inferring gender, as indicated by the turning point in performance curves occurring earlier for age.

Key Results

  • Superior Performance: The proposed MulAProM model, leveraging fused data and interdependence modeling, achieved state-of-the-art results, outperforming several baseline models (kNN, LRC, NB, GBDT, ANN, L-SVM, N-SVM, SM2L) in terms of precision, recall, and F1-measure for both gender and age inference.
  • Data Fusion Gains: Fusing CDR and DTR data sources led to significantly better performance and lower variance in F1-Measure for both gender and age inference compared to using individual sources.
  • PCA Effectiveness: The study found an optimal dimension reduction setting (d=80) for PCA, demonstrating that proper dimension reduction can alleviate the curse of dimensionality and even improve inference performance.
  • Training Data Impact: Performance generally increases with more training data, but plateaus after a certain point, suggesting that excessive data may not always yield proportional gains and can be computationally wasteful.

Limitations and Future Work

  • Dataset Specificity: The study is based on a single telecommunication dataset. The authors acknowledge that experiments on more diverse datasets are needed to validate the model's generalizability.
  • Extensibility: While the MulAProM framework is designed to be extensible to other user context attributes, the current experiments focus only on gender and age.
  • Fake Context Detection: The authors suggest that their model could be adapted to identify fake context information, as inconsistencies might arise when users provide inaccurate details for accounts.
  • Time-Dependent Patterns: Future work could explore modeling time-dependent patterns in mobile usage, which can vary significantly across different age groups and social circles.
  • Privacy: The scarcity of public datasets due to privacy concerns is a general challenge in this field, and the authors express a desire to release parts of their dataset for public research.

Summary updated

BibTeX
@misc{wang2017multiple,
  author = {Shangguang Wang},
  title = {{Multiple User Context Inference by Fusing Data Sources}},
  year = {2017},
  eprint = {1703.04215},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1703.04215},
}
Figure