Source-linked AI summary
Solving the Cold-Start Problem in Recommender Systems with Social Tags
Zi-Ke Zhang, Chuang Liu, Yi-Cheng Zhang, Tao Zhou
TL;DR
The paper addresses how to improve personalized recommendation when conventional object information is limited, especially for unpopular or new objects. It proposes a diffusion algorithm on user–tag–object relations and finds improved accuracy and diversity, particularly for small-degree objects, while noting tag polysemy and synonymy as a limitation.
Problem
Object attributes are limited by vocabulary and describe global properties, while recommender systems lack sufficient information to recommend unpopular or new objects effectively.
Method
The proposed diffusion algorithm uses tag frequencies as users’ topic preferences and tag–object links as semantic relations connecting users and objects.
Results
The algorithm outperforms two baselines in accuracy and diversity, especially for objects with small degrees, particularly when k_o ≤10.
Takeaways & Limitations
Social tags can help recommender systems find less popular yet interesting objects and broaden recommendation diversity within social tagging systems.
Takeaways & Limitations
Tag polysemy and synonymy may produce coarse and inaccurate performance.
Abstract
from arXiv · showhide
In this paper, based on the user-tag-object tripartite graphs, we propose a recommendation algorithm, which considers social tags as an important role for information retrieval. Besides its low cost of computational time, the experiment results of two real-world data sets, \emph{Del.icio.us} and \emph{MovieLens}, show it can enhance the algorithmic accuracy and diversity. Especially, it can obtain more personalized recommendation results when users have diverse topics of tags. In addition, the numerical results on the dependence of algorithmic accuracy indicates that the proposed algorithm is particularly effective for small degree objects, which reminds us of the well-known \emph{cold-start} problem in recommender systems. Further empirical study shows that the proposed algorithm can significantly solve this problem in social tagging systems with heterogeneous object degree distributions.
Introduction. –
Personalized recommendation must estimate user–object relevance while filtering irrelevant objects, but conventional object attributes are limited by vocabulary and describe global rather than personalized properties. Social tagging supplies user-generated semantic links that the proposed diffusion algorithm uses to connect users and objects.
- Motivation: Conventional recommendation combines similarity estimation from user activity histories with auxiliary information for filtering irrelevant objects.Computing and storing all user-pair similarities is costly, so methods often retain only the top-k similar users.
- Motivation: Object attributes are constrained by their vocabulary and primarily describe global properties, limiting their usefulness for personalized recommendations.
- Social tagging: Social tagging lets users freely annotate collections, broadening semantic relations among users and objects without requiring specialized participation skills.
- Proposed approach: The proposed diffusion algorithm treats social tags as a bridge between users and objects, using tag frequencies as personal preferences and tag–object relations as global information.
- Evaluation: The study evaluates recommendation accuracy and diversity on Del.icio.us and MovieLens, including object-degree-dependent accuracy and tag-usage diversity analyses.
Data. –
The study uses Del.icio.us and MovieLens tagging data represented as user–object–tag entries, with each dataset split into training and testing portions. The experiments examine recommendation-list diversity across algorithms.
- Datasets: The empirical datasets are Del.icio.us social-bookmarking data and MovieLens movie ratings with tagging functionality.Isolated nodes are removed, and each user must have collected at least one object.
- Evaluation measures: Recommendation diversity is evaluated using ⟨InterD⟩ as a function of recommendation-list length for the three algorithms in both datasets.
- Data representation: Each record contains a user, an object, and the tags that user assigned to that object.The entry form is F={user, object, tag1, tag2, · · · , tagt}.
- Evaluation split: The datasets are randomly divided into training and testing sets, with 90% of entries used for training and 10% for testing.
Algorithms. –
The paper compares three diffusion algorithms on a user–object–tag system, culminating in a user-tag-object method that uses tag preferences and tag–object relations. It ranks uncollected objects for recommendation and is presented as more personalized and computationally efficient than the two alternatives.
- The three methods are user-object diffusion, user-object-tag diffusion, and user-tag-object diffusion.
- Algorithm II: Algorithm II passes resources from objects to neighboring tags and then back to neighboring objects.
- Algorithm III: Algorithm III initializes resources on tags according to the target user’s tag frequencies, then sends them directly to neighboring objects.
- After scoring, uncollected objects are ranked in descending order and the top L objects are recommended.
- Algorithm III: Algorithm III treats tag usage frequency as personal preference and tag–object relations as global information.
Metrics. –
The evaluation measures recommendation accuracy, inter-user diversity, and within-list topic diversity. These metrics capture ranking quality, differences between users’ lists, and the breadth of topics within each user’s recommendations.
- The study uses three metrics to characterize recommendation accuracy and diversity.
- Ranking Score: Ranking Score divides each testing-set object’s rank by the user’s uncollected objects; lower average RS indicates higher accuracy.
- Inter Diversity: InterD measures differences between users’ recommendation lists, with higher average InterD indicating greater personalization.
- Inner Diversity: InnerD measures differences among objects within one user’s recommendation list, with higher average InnerD indicating greater topic diversification.
Results. –
The proposed algorithm improves recommendation accuracy, particularly for low-degree objects, and generally increases recommendation diversity. Its advantage is strongest for cold objects, which comprise most objects in both evaluated datasets.
- The evaluation examines accuracy and diversity using ranking score, inter-diversity, inner-diversity, tag-overlap analysis, and entropy-based tag-usage measures.Table 2 reports Del.icio.us accuracy values averaged over 50 independent training/testing splits.
- The degree-dependent ranking score decreases as object degree increases, with algorithm III outperforming algorithms I and II for k_o ≤ 10.The three algorithms’ curves intersect around k_o=10; algorithm III performs worse when k_o > 10.
- 90.04% of Del.icio.us objects and 69.35% of MovieLens objects have k_o ≤ 10, making low-degree objects a substantial evaluation regime.The paper identifies these unpopular or new objects as the cold-start problem and reports that social tags help users find them.
- Social-tag diversity conditions recommendation diversity: tags generate more diverse recommendations only when users employ the tags in diverse ways.The overlapping-ratio analysis links greater tag overlap with smaller diversity.
- Algorithm III generally improves ⟨InnerD⟩, indicating broader recommendation horizons, except for MovieLens with very small L.The exception is attributed to MovieLens’s narrow tag choice.
Conclusions and Discussion. –
The proposed social-tag recommendation algorithm treats tag frequencies as user preferences and tag-object links as semantic relations. It outperforms two baselines in accuracy and diversity, especially for small-degree objects, while tag ambiguity remains a limitation.
- The algorithm uses tag frequencies as user preferences and tag-object links as semantic relations between tags and objects.
- The proposed method outperforms two baseline algorithms in both recommendation accuracy and diversity.
- The method performs especially well for objects with small degrees (k_o ≤10), which constitute the majority of objects.
- Polysemy and synonymy among tags might produce coarse and inaccurate performance; tag clustering is suggested as a possible remedy.