Maximizing product adoption in social networks

Smriti Bhagat, Amit Goyal, Laks V. S. LakshmananWSDM 2012 · DOI

TL;DR

This paper tackles the problem of maximizing product adoption in social networks by introducing a new model, LT-C, that distinguishes between influence and actual adoption. Unlike previous work that conflates these, LT-C accounts for users who might spread opinions without adopting (tattlers) and the varying strength of endorsements based on user experience (ratings). This more realistic model leads to significantly more accurate predictions of product adoption.

What Problem Does This Paper Address?

Traditional viral marketing and influence maximization research in social networks often assumes a direct link between being influenced and adopting a product. This means that if a user is exposed to a product through their network, they are assumed to adopt it and then influence their friends. However, this is a simplification. In reality, a user might be influenced but choose not to adopt due to personal factors like budget or interest. Furthermore, even if a user adopts a product, their endorsement might be weak if they didn't like it, or strong if they loved it. This paper argues that these nuances are crucial for accurate prediction and effective marketing. They highlight that:

  • Influence vs. Adoption: Being influenced doesn't automatically mean adoption.
  • Tattlers: Users can spread information (opinions) about a product without adopting it themselves. These "tattlers" can either promote or inhibit adoption.
  • Varying Endorsements: The strength of a user's endorsement (influence on others) depends on their actual experience with the product, which can be abstracted by user ratings.

The paper aims to bridge this gap by proposing a model that explicitly separates influence from adoption and incorporates user opinions (ratings) and the behavior of non-adopters.

Methods

The core contribution is the LT-C (Linear Threshold with Colors) model. This model extends the classical Linear Threshold (LT) model for influence propagation. Here's how it works:

  • States: Instead of just "active" or "inactive," nodes (users) can be in several states:
  • Inactive: Not yet influenced.
  • Active: Influenced by neighbors.
  • Adopt (Green): The user adopts the product.
  • Tattle (Blue/Red): The user is influenced but does not adopt. They can either promote (blue) or inhibit (red) adoption by others.
  • Activation: Similar to LT, a user becomes active if the sum of incoming influence from active neighbors exceeds their threshold.
  • Opinion Incorporation: The influence from an active neighbor is modulated by the neighbor's opinion (rating) of the product. This is captured by a weighted influence that considers the difference between a user's rating and the min/max possible ratings.
  • Tattle Behavior: After becoming active, a user adopts with probability $\lambda_v$ or becomes a "tattler" with probability $1 - \lambda_v$. Tattlers then decide to promote with probability $\mu_v$ or inhibit with probability $1 - \mu_v$.
  • Adoption Maximization: The problem of finding a seed set of $k$ users to maximize the expected number of adopters (not just influenced users) is formulated. The authors prove this problem is NP-hard but that the objective function is monotone and submodular, allowing for a greedy approximation algorithm.
  • Parameter Learning: The model parameters ($\lambda_v$, $\mu_v$, and edge weights) are learned from real-world data using techniques like collaborative filtering for ratings and maximum likelihood estimation for $\lambda$ and $\mu$.

Datasets

The authors evaluate their model on three real-world datasets:

  • Flixster: A social network for movies, providing user ratings and friendship relations. A subgraph was used for experiments.
  • Movielens: A movie recommender system dataset. Since it lacks an explicit social graph, an implicit social graph was constructed based on user similarity derived from collaborative filtering.
  • Last.fm: A music social network dataset, including user listening history and tagging information.

These datasets are used for both training model parameters and evaluating the accuracy of adoption spread predictions.

Key Insights and Ideas

  • Adoption is Not Just Influence: The most significant insight is the clear distinction between influence and adoption. Conflating them leads to gross overestimation of adoption, as seen in the classical LT model's predictions.
  • Tattlers Matter: Users who spread opinions without adopting can significantly impact the overall adoption process, either positively or negatively. This behavior is common and needs to be modeled.
  • Ratings as Opinion Proxies: User ratings provide a valuable signal about the perceived quality of a product, which directly influences how strongly a user will endorse it. Incorporating these ratings makes predictions more realistic.
  • Seed Selection is Product-Specific: The optimal set of users to target for marketing can vary significantly depending on the product being promoted, suggesting that a one-size-fits-all approach to seed selection is suboptimal.
  • Network Structure is Key: The paper reinforces that network structure plays a critical role in how information and adoption spread, and that understanding this structure is vital for effective viral marketing.

Key Results

  • Improved Prediction Accuracy: The LT-C model significantly outperforms the classical LT model in predicting actual product adoption. For instance, on Flixster, the classical LT model predicted a coverage of 4200 for a seed set of 50, while LT-C predicted only 701, which was much closer to the actual coverage.
  • Reduced RMSE: The Root Mean Square Error (RMSE) for coverage estimation using LT-C was over 10 times lower than that of the classical LT model on Flixster.
  • Model Component Importance: Experiments showed that both incorporating "tattle" behavior and user ratings improve prediction accuracy. LT-C, which includes both, performs best.
  • Seed Consistency: While some highly influential users are consistently good seeds across different products, the optimal seed set often varies, highlighting the need for product-specific targeting.
  • Parameter Distribution: The parameters $\lambda$ and $\mu$ often follow a Beta distribution, suggesting that users have strong tendencies towards either adopting/tattling or promoting/inhibiting.

Limitations and Future Work

  • Data Availability: The paper notes that obtaining $\lambda$ and $\mu$ values directly from user logs can be challenging. They propose using distributions (like Beta) learned from available data as a viable alternative when direct data is limited.
  • Observable States: The model uses intermediate states (active, tattle) for modeling, but only INACTIVE, ADOPT, PROMOTE, and INHIBIT states are directly observable in real data.
  • Future Directions:
    • Splitting "Adopt" State: Further refining the model by splitting the "adopt" state into "adopt and rate" vs. "adopt and not rate" could improve accuracy.
  • Negative Opinions: Incorporating negative opinions more explicitly, similar to prior work, could lead to a more expressive model.
  • Scalability: Developing scalable heuristic algorithms for LT-C is crucial for handling very large networks and seed sets.
  • Diverse Domains: Validating the LT-C model on a wider range of real-world datasets from different domains is important.

Summary updated

BibTeX
@inproceedings{bhagat2012maximizing,
  author = {Smriti Bhagat and Amit Goyal and Laks V. S. Lakshmanan},
  title = {{Maximizing product adoption in social networks}},
  booktitle = {WSDM},
  year = {2012},
  doi = {10.1145/2124295.2124368},
}

Loading related content...

Figure