Modeling Of Telecommunication Products For Personalized Recommendation

With the development of telecommunication technology and mergence of telecom industry chain, telecom product becomes more and more complex in both content and providing channel. "Information Overload" of telecom product will be the main obstacle of consumption in telecom market. This paper including the following parts:1. Framework of telecom product which including content layer, behavior layer and function layer was built. Meanwhile, description method for tree structure was designed.Firstly, based on concept research of telecom product and SBF method, telecom product structure with three layers was built; Secondly, content attributes, behavior attributes and function attributes were distilled from telecom product instruction; thirdly, according to criterion of "Customer Perspective", "Comprehensive" and "Diversity", telecom product model was built; Finally, by using Protege, telecom product model was described in OWL. Recent research of telecom product always focuses on information sharing problem and was built as data model, this research will provide theoretical basis and piratical instruction for recommendation system research in telecom industry.2. Semantic similarity algorithm for telecom product (TPM-RS) was built, and performed well in empirical research.Based on research of Dekang Lin, semantic similarity algorithm for telecom product was designed by introducing node level information for weight of node distance. In empirical research of 1450 telecom customers, precision of recommendation by TPM-SA is 35%, recall is 70%, which are higher than related research based on Cosine-Based Similarity algorithm applied in Movielens and Jester data set (27% and 60%).3. Recommendation algorithm considering diversity and related evaluation index were built Based on results of empirical research, TPM-RS performed excellently in diversity index while keeping high precision and recall index.Firstly, diversity recommendation described by self-information was introduced; Recommendation mechanism of "recommendation based on similarity of content and behavior and diversity of function (Recom 1)", "recommendation based on similarity of function and behavior and diversity of content (Recom 2)" and "recommendation based on similarity of content and function and diversity of behavior (Recom 3)" were designed; Finally, evaluation index including precision, recall and diversity was proposed. Results of empirical study of 1450 customer samples indicated: (1) In summary, Recom 2 performed better than the other two algorithm (Recom 1 and Recom 3); (2) Recommendation results generated by TPM-RS performed much better than improved CF algorithm (K-Means-CF) in diversity index (6 to 3), while keeping great performance in precision and recall.