Language model based interactive estimation of distribution algorithm
Graphical abstract
Introduction
Objective functions of personalized optimization problems, e.g. product design, personalized search, and information retrieval, are impossible to precisely define using mathematical expressions since they are highly dependent on user preferences. Although evolutionary computation (EC) has been proven to be a powerful tool for solving complex optimization [1], [2], [3] problems, the requirement of precise mathematical definitions cannot be fulfilled in personalized search. In such scenarios, interactive evolutionary computation (IEC) [4], [5] is more feasible and efficient as it involves a human user in the evaluation process; they have been developed and successfully applied to various practical problems, such as product design, web page layout design, and anti-collision design of vehicles [5], [6], [7], [8]. This paper considers the optimization problems that occur in, for example, the following scenario: A person is searching the web for a particular movie, beginning the search with a query of a few words. The search engine presents a few movie candidates. The user clicks on some of the candidates and saves some of them. Based on the user’s actions, the search engine presents some new candidates. This process continues until the user is satisfied with the result. The purpose of this proposed novel method is to speed up the search process and reduce the need for user interference.
IEC requires human evaluations, which can inevitably cause user fatigue given a complex problem. The restriction on population size and evolutionary generation prevents the use of IEC in tackling a wide range of problems [5]. Accordingly, much more attention has been paid to alleviating user fatigue and improving explorations from the following three aspects [9]: (1) the design of friendly human–computer interfaces or novel evaluation modes to reduce the user burden, e.g. evaluating individuals with discrete, fuzzy, or interval numbers [9], [10], [11]; (2) the use of a preference surrogate, with a small number of evaluated individuals and then apply it to help with the assessment. With surrogates, IEC can upscale the population size and generation as conventional EC approaches [12], [13], which greatly improves the explorations of IEC. (3) the use of knowledge from evolution to modify evolutionary operators to accelerate search and reduce fatigue [14], [15]. These studies have greatly enhanced IEC. Although the above-mentioned studies have greatly enhanced IEC , the application of those methods in resolving preference-related complex problems remains a challenging task. Particularly in word-described ones such as online personalized search in E-commerce.
The main reason is that the information covered by the numerical representations for these word-described problems is insufficient to model the preference-related objectives, leading to inefficient or even wrong searches. Sun et al. [9], [16] used a limited number of numerical values to encode these word-described items in the framework of an interactive genetic algorithm for the personalized search, which made the traditional evolutionary operators easier to implement. Wang et al. [17] modeled TV programs with five attributes in their experiments when studying preference recommendations for personalized search.
These studies are easier to understand and implement; however, such numerical encoding loses a considerable amount of implied semantic information contained in the words. In addition, IEC depends on the evaluations assigned by the users, who have got used to thinking and evaluating with words instead of numbers. The gap between the users’ evaluations and numerical encoding results in a need for additional human–computer interactions and inevitably causes more user fatigue. Accordingly, designing a non-numerical encoding method which minimizes loss of semantic information and developing corresponding evolutionary operators becomes essential as this will help to enhance the performance of IEC in solving more practical and complex problems. Furthermore, user preferences or decisions will be influenced greatly by other user comments, and social or group comments should be integrated with IEC so that the current user is able to precisely evaluate the searched solutions. Motivated by the above, we focus on developing an enhanced IEC by integrating social comments, designing a novel encoding and corresponding evolutionary operators for solving problems described with words in the personalized search.
Applying IEC, short for Interactive Evolutionary Computation, to the word/document-described optimization problems relies on establishing a bridge between the textual phenotype evaluated by the user and the numerical genotype operated by EC. The language model Doc2Vec [18], [19] is a good choice to convert textual phenotypes into numerical vectors by preserving most of the semantic relationships among the words. Therefore, we employ this model to represent the genotype of a document, i.e., a searched item, including the word description and social comments. Clearly, both are naturally combined in the Doc2Vec based representation. Given this, a new IEC must be developed to gain the advantages both from itself and the model, i.e., Doc2Vec-assisted initialization and interactive evolutionary operators.
This work develops a language model based interactive estimation of distribution algorithm (LMIEDA) to perform the evolutionary optimization in personalized search. In LMIEDA, the Doc2Vec is applied to convert the word/document-described problem into a dynamic document matching one by encoding the word frequency as individuals. A preference function is approximately constructed based on user interactions and is used to estimate the individual’s fitness. Based on the fitness, the Dirichlet-Multinomial compound distribution and a Bayesian inference involving the Dirichlet-Multinomial compound distribution is designed to track the user’s preference on the variables. With these, the probabilistic model of estimation of distribution algorithm (EDA) and the corresponding sampling are presented. The user’s burden here can be greatly reduced since our algorithm is able to estimate the fitness of all individuals without the user.
The main contributions of this study are as follows. (1) To the best of our knowledge, language models have not been used in EC/IEC. As an encoding method, it helps to reduce information loss and naturally introduces social intelligence. (2) Since the encoded variables are discrete (with finite states), the Dirichlet-Multinomial compound distribution is adopted as the probabilistic model of IEDA to be compatible with encoded candidates. (3) The probabilistic model is updated with the help of Bayesian learning to directly track variable distribution, and most conventional EDAs employ Bayesian networks to depict the dependencies between variables. (4) The proposed algorithm is applied to some personalized search for books and movies to prove its effectiveness and efficiency.
The remainder of this paper is organized as follows. Section 2 describes the related work on the personalized search assisted with evolutionary algorithms, the estimation of distribution algorithms (EDAs), and the basic concept of Doc2Vec. The details of the proposed algorithm, including the definition of the word-described personalized search, the framework, the critical encoding, preference expression, and the IEDA, are presented in Section 3. Section 4 addresses the application of the proposed algorithm together with the experimental results and analyses. Conclusions are drawn in Section 5.
Section snippets
Personalized search assisted with evolutionary algorithms
The task of the personalized search is to find the items that give the user the most satisfaction; therefore, it is an optimization problem in nature. However, what distinguishes personalized search from typical optimization problems is that users, rather than mathematical functions, play the role of the fitness functions. Although it is hardly possible to solve this problem involving cognitive processes only with tractable mathematical calculations, researchers can still describe some
Definitions of word/document-described optimization problems
For the word-described personalized search, by naturally combining social intelligence from comments, an item can be expressed as , in which the first part comes from its seller, and with specific meaning on the item come from users. Supposing a user’s preference on a is , the search can be formulated as Eq. (1), where is the feasible space of searched items. Evolutionary algorithms will be powerful for solving
Experimental setting
Comparisons on the personalized search of two different fields, i.e., movies and books, among the proposed algorithm and other IECs are conducted to prove its effectiveness and efficiency. Movies and books, which are commonly described with text, are chosen as the search target because they cannot be well modeled with the key–value pattern. The data for movies and books (updated in March 2018) are from imdb.com and Douban.com. IMDb is an online database of information related to films, TV
Conclusions
For solving word/document-described problems that cannot be well encoded with the structural numerical methods, LMIEDA is proposed by integrating the mixture of unigrams, LDA, and Doc2Vec into the EDA framework. From the viewpoint of optimization, language model based encoding, a novel preference expression by use of Dirichlet-Multinomial compound distribution, and a Bayesian inference-enhanced interactive version of EDA (estimation of the distribution algorithm) have been studied to
CRediT authorship contribution statement
Yang Chen: Conceptualization, Methodology, Software, Writing - original draft. Yaochu Jin: Conceptualization, Supervision, Validation, Writing - review & editing. Xiaoyan Sun: Conceptualization, Supervision, Software, Validation, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
This work is supported by the National Natural Science Foundation of China with Grant No. 61876184 and 61473298.
References (63)
- et al.
A novel hybrid multi-objective artificial bee colony algorithm for blocking lot-streaming flow shop scheduling problems
Knowl.-Based Syst.
(2018) - et al.
Preference recommendation for personalized search
Knowl.-Based Syst.
(2016) - et al.
Sequential funding the venture project or not? a prospect consensus process with probabilistic hesitant fuzzy preference information
Knowl.-Based Syst.
(2018) Evaluating customer aid functions of online stores with agent-based models of customer behavior and evolution strategy
Inform. Sci.
(2010)- et al.
Incorporating sentiment into tag-based user profiles and resource profiles for personalized search in folksonomy
Inf. Process. Manage.
(2016) - et al.
Word sense disambiguation: A comprehensive knowledge exploitation framework
Knowl.-Based Syst.
(2020) - et al.
Hybrid query expansion using lexical resources and word embeddings for sentence retrieval in question answering
Inform. Sci.
(2020) - et al.
Unsupervised language identification based on latent Dirichlet allocation
Comput. Speech Lang.
(2016) Finding structure in time* 1
Cogn. Sci.
(1990)- et al.
Evolutionary multi-objective blocking lot-streaming flow shop scheduling with machine breakdowns
IEEE Trans. Cybern.
(2019)
Indicator-based set evolution particle swarm optimization for many-objective problems
Soft Comput.
Federated learning assisted interactive eda with dual probabilistic models for personalized search
Interactive evolutionary computation: fusion of the capabilities of EC optimization and human evaluation
Proc. IEEE
DPM-IEDA: Dual probabilistic model assisted interactive estimation of distribution algorithm for personalized search
IEEE Access
A proposal for user’s intervention in interactive evolutionary computation for optimizing fragrance composition
Commun. Comput. Inf. Sci.
Interactive genetic algorithm with CP-nets preference surrogate and application in personalized search
Control Decis.
Neural network surrogate models of interactive genetic algorithms with individual’s interval fitness
Control Decis.
Interactive genetic algorithms with individual’s uncertain fitness
Chinese J. Electron.
An interactive simple indicator-based evolutionary algorithm (I-SIBEA) for multiobjective optimization problems
Lecture Notes in Comput. Sci.
An interactive evolutionary multiobjective optimization method based on the WASF-GA algorithm
Interactive genetic algorithm with group intelligence articulated possibilistic condition preference model
Efficient estimation of word representations in vector space
Distributed representations of sentences and documents
Research development of user interest modeling in China
J. Intell.
Survey of personalized recommendation systems
Comput. Eng. Appl.
Fuzzy rankings for preferences modeling in group decision making
Int. J. Intell. Syst.
Hesitant fuzzy linguistic preference utility set and its application in selection of fire rescue plans
Int. J. Environ. Res. Public Health
User preference modeling by global and individual weights for personalized recommendation
Acta Polytech. Hung.
Cited by (8)
Identification of emerging business areas for business opportunity analysis: An approach based on language model and local outlier factor
2022, Computers in IndustryCitation Excerpt :This pioneering method introduced Negative sampling, which learns more accurate vectors for frequent words—instead of the hierarchical SoftMax—to facilitate both faster training and better representation of uncommon words. The success of Word2Vec ignited the continuous development of semantic language models (Chen et al., 2020) that has led to the evolution of such models based on the attention mechanism (Bahdanau et al., 2014) or transformer (Vaswani et al., 2017). Recently, universal language models such as Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al., 2018), a robustly optimised BERT training approach (RoBERTa) (Liu et al., 2019), and text-to-text transfer transformer (T5) (Raffel et al., 2019) have been developed.
A two-stage approach for multicast-oriented virtual network function placement
2021, Applied Soft ComputingCitation Excerpt :Running a stochastic search algorithm 20 times is meaningful from the point of view of statistics, i.e., the results collected to some extent reflect that algorithm’s optimization performance [63,64]. All results are collected by running each algorithm 20 times [65,66] on a machine with Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10 GHz and 16 GB RAM. Fig. 12 illustrates the best fitness curves obtained by the six algorithms in 18 test instances.
Multi-scale Self-Organizing Map assisted Deep Autoencoding Gaussian Mixture Model for unsupervised intrusion detection
2021, Knowledge-Based SystemsCitation Excerpt :For instance, support vector machines and random forests are common options. Recently, deep models become competitive because more complex algorithms coupled with greater computational capacity can be obtained [2,3]. In industries and academia, deep anomaly detection has received a wide range of applications [4], e.g., deep models based on self-taught learning [5] and DAGMM [6].
Review on personalized search and recommendation algorithms for multi-source heterogeneous data
2024, Kongzhi Lilun Yu Yingyong/Control Theory and ApplicationsMarine Goal Optimizer Tuned Deep BiLSTM-Based Self-Configuring Intrusion Detection in Cloud
2024, Journal of Grid Computing