Intelligent Marketing And Its Applications Based On User Preference Modeling

Marketing has experienced significant changes due to shifts in society.The rise of big data has brought forth transformative technologies like artificial intelligence and the Internet of Things,reshaping people’s lifestyles.This has given rise to a new era of intelligent marketing amidst societal changes.Unlike traditional marketing,intelligent marketing relies on leveraging big data to empower businesses to prioritize user preferences and transfer market power to users.It also encourages users to actively participate in value creation.Consequently,big data plays a crucial role in modeling user preferences and helps companies place the user at the core of their marketing strategy.Intelligent marketing utilizes artificial intelligence algorithms,like machine learning and deep learning,to analyze vast user consumption behavior and feature data for preference modeling.This technology presents significant opportunities and challenges for researchers to explore scientific applications and commercialize artificial intelligence.Consequently,intelligent marketing is rapidly evolving in academia and industry,with various theories,technologies,and applications emerging.It can be broadly categorized into two aspects:the first focuses on the user-product consumption relationship,including news and video recommendations aligned with user interests.The second aspect centers on users,products,and advertisers,involving internet display advertising to target users most likely to click or consume the product.Understanding personalized needs through modeling user preferences is a fundamental challenge in intelligent marketing due to the complexity of user behavior.Firstly,users’ interests and preferences are often expressed implicitly,requiring the analysis of extensive consumption behaviors to offer personalized recommendations.Secondly,users’ attributes exhibit high dimensionality and sparsity,necessitating adaptive models for representing user features.Lastly,the diversity of marketing scenarios calls for reliable algorithms in user preference modeling that are suitable for real-world situations to guide product marketing strategies.To tackle these challenges,this dissertation conducts comprehensive research and applications in intelligent marketing systems centered around user preference modeling.It explores user consumption behavior modeling,user attribute feature modeling,and intelligent marketing system applications.The dissertation’s primary contributions and focus are summarized as follows:Firstly,in terms of user consumption behavior modeling,this dissertation proposes a Multiple Pairwise Ranking algorithm,a Collaborative-List-and-Pairwise Filtering algorithm,and a Multi-Objective Pairwise Ranking algorithm.The Multiple Pairwise Ranking(MPR)algorithm is introduced to overcome the limitations of traditional collaborative filtering algorithms in capturing user interests in cold-start scenarios.MPR relaxes the strict assumptions of user interest preferences by considering two reasons for unconsumed products:"unobserved" or "not interested".Experimental results demonstrate that MPR outperforms other advanced recommendation algorithms in cold-start scenarios.The Collaborative-List-and-Pairwise Filtering(CLAPF)algorithm is proposed to optimize the effectiveness of recommended item rankings.Traditional collaborative filtering algorithms prioritize Area Under the Curve(AUC),neglecting users’sensitivity to rankings.Listwise methods,such as maximizing the Mean Reciprocal Rank(MRR),attempt to address this issue but may be inefficient and less effective in general implicit feedback scenarios.CLAPF smooths the well-known rank-biased measure called Mean Average Precision(MAP)and combines rank-biased metrics with the pairwise objective function to capture the performance of top-k recommendation.This enables CLAPF to recommend a ranking-sensitive item list that effectively captures potential user interests.Experimental results on various datasets show that CLAPF outperforms state-of-the-art recommendation algorithms on ranking-sensitive evaluation metrics.A sampling scheme is also discussed to enhance CLAPF’s convergence speed.To cater to the diverse demands of real-world scenarios,personalized ranking algorithms like MPR and CLAPF have been developed for specific needs,such as cold-start and ranking-sensitive situations.However,these demands are diverse,and tradeoffs are in evitable.Retaining the unique advantages of a single objective function and giving up others may not be sufficient for the requirements of a recommender system.To overcome this challenge,the dissertation proposes the Multi-Objective Pairwise Ranking(MOPR)algorithm.MOPR employs a gradient-based optimizer within matrix factorization to identify Pareto optimal solutions for two pairwise objective functions.By applying MOPR to various pairwise objective functions,it generates multiple algorithms,each maintaining the advantages of its respective objective function without introducing external hyperparameters or extensive computations.Extensive experiments conducted on four public datasets validate the effectiveness and efficiency of MOPR in diverse scenarios.Regarding user attribute feature modeling,this dissertation introduces three proposals:Extreme Cross Network(XCrossNet),Cognitive Evolutionary Learning,and Cognitive Evolutionary Search.In Internet display advertising,user attribute features can be classified into two types:numerical(dense)parameters and categorical(sparse)parameters.Dense features are represented by numerical field values,while sparse features are represented by one-hot encoded vectors associated with categorical fields.However,existing feature representation models lack specific methods for modeling different feature types,with most models focusing on cross sparse feature modeling while overlooking cross dense feature modeling.To address this limitation,XCrossNet is introduced,which explicitly learns both cross dense and cross sparse features.XCrossNet comprises three components:feature crossing,feature concatenation,and feature selection.The feature crossing component includes a cross layer for dense feature crossing and a product layer for sparse feature crossing.By considering feature structures,XCrossNet offers a more expressive representation and enhances ClickThrough Rate(CTR)prediction accuracy.Experimental results on the Criteo Kaggle dataset demonstrate that XCrossNet outperforms state-of-the-art baselines.Existing approaches to feature and interaction enumeration often use pre-defined operations with expert guidance.However,these methods may lack adaptability and introduce unnecessary noise and complexity during training.To overcome these issues,the dissertation introduces the Cognitive Evolutionary Learning(CELL)framework,which leverages cognitive ability for adaptive performance in diverse environments.The CELL framework comprises three stages:DNA search,genome search,and model functioning.Feature interactions are analogous to double-stranded DNA,and relevant features and interactions resemble genomes.Survival rates of organisms under natural selection simulate the fitness of the model on operations,features,and interactions.CELL evolves into different models for various tasks and data,outperforming state-ofthe-art baselines.Real-world dataset experiments demonstrate CELL’s effectiveness,and synthetic experiments show its consistent discovery of pre-defined interaction patterns for feature pairs.Traditional automated machine learning(AutoML)and CELL often operate under an unrealistic assumption-that the parameter space is continuous,differentiable,and convex.However,fulfilling this assumption is often impractical in many real-world situations.As a solution,this dissertation introduces the Cognitive Evolutionary Search(CELS)framework.CELS leverages discrete selection to pinpoint optimal operations,eliminating the need for continuity and differentiability.It generates new solutions,or "offspring",via individual mutation and population crossover processes.The framework incorporates four variants:two that rely on individual-based search algorithms and two that are population-based.Evaluations conducted on three distinct advertising datasets demonstrate that CELS implementations efficiently and reliably address the issue of user feature selection.In terms of intelligent marketing system applications,this dissertation partners with a leading fintech bank in China to overcome the constraints of traditional manual financial advisory methods,which are often slow and inadequate for handling large volumes of data.As a response to these limitations,Intelligent Financial Advisors(IFAs)have been introduced in online financial apps.IFAs deliver customized,high-quality portfolio suggestions by accurately predicting a user’s propensity to invest in suggested portfolios through the analysis of vast,high-dimensional user data.Specifically,financial scenarios pose two critical challenges:sample selection bias and data sparsity.To overcome these challenges,user behaviors are analyzed to identify a potential conversion relationship,wherein the user first converts to an activated user and subsequently becomes a client.Multitask learning is employed to predict both conversion rates simultaneously,and a multi-stage feature selection algorithm is designed to identify the most relevant user features from a large pool of feature fields.Extensive experiments conducted in a real bank scenario validate the effectiveness of the proposed method in client identification.