Research And Application Of Telecom Package Recommendation Algorithm Based On Feature Interaction

As a pillar industry in China,the telecom industry has a pivotal role in both the national digital transformation level and the national digital economy level.As the core of traditional mobile communication business,telecom packages integrate the sales of different voice traffic and other services.For different user groups,they need different telecom packages;in order to meet the needs of user groups at all stages,operators have launched a variety of packages.But in the face of a variety of packages,users are often unable to select the most suitable for their choice.Traditional manual recommendation includes telephone recommendation,business hall on-site recommendation,which will be labor intensive and costly.And as more and more users,there may be unpredictable situations in the process of consultation,which over time makes the user’s experience greatly reduced,and thus lost.Therefore,it is important to implement an automatic package recommendation system,which apply the recommendation algorithm to the telecom package recommendation scenario and recommend suitable packages to users.Since there are a large number of features in the package recommendation scenario that cannot be well recommended to users if only relying on human judgment,the recommendation algorithm studied in this paper can make up for this deficiency by using this paper.By interacting features such as user tags,package information and user’s previous consumption habits,thus improving the recommendation algorithm,we obtain experimental results that can be drawn from.However,there are some shortcomings in the existing recommendation models with feature interactions,such as many models cannot automatically and efficiently learn finite order feature interactions,and many models treat all features equally and consider them to contribute equally to the final prediction results;in addition,the existing models are basically combined based on features,but the performance of model prediction can be enhanced if we take the perspective of feature domain.In this paper,we propose two click-through prediction models based on the above issues and apply them to telecom package recommendation prediction.The main work of the paper is as follows:(1)A feature importance based recommendation model Fi DCN(Feature Importance and Deep Cross Network)is proposed.This model uses SENET network to give different weights to distinct features,thus filtering some unimportant features before feature interaction.And the Cross Network and Deep Neural Network are used in parallel in the feature interaction layer to extract explicit feature interaction and implicit feature interaction,respectively.Since the cross network is composed of multiple cross layers,it can extract feature interactions of arbitrary order.To enable the model to reduce the training cost while ensuring performance,the Fi DCN model utilizes a low-rank technique in the cross network,which approximates a dense matrix and can approximate feature interactions in the subspace.(2)A recommedation model FBi NET(Field-aware and Bilinear Feature Interaction Network)based on feature domain and bilinear feature intersection is proposed.The model uses SENET network to filter unimportant features before feature interactions.Bilinear feature interaction is performed by combining inner product and Hadamard product,thus taking into account the important information of each dimension on the interacting feature vector.The features are also subjected to domain-based interaction,which is computed at the vector level crossover,and higher-order feature interaction is achieved by overlaying multiple layers.Finally,the two parts of the feature interaction information are connected together and fed into a deep neural network to extract the implicit feature interactions of the features and output the final prediction results.(3)In this paper,the classical public dataset Criteo in recommender systems and the dataset in a real production environment of a telecommunication company are selected.And the two models proposed above are subjected to a large number of comparison experiments,ablation experiments and parameter experiments.The experimental results prove that the Fi DCN model and the FBi NET model can achieve better performance.Based on them,a telecom package recommendation system is designed and implemented,which can display the model recommendation results to users more intuitively through the visualized system interface and improve the success rate of package recommendation.