Restaurant recommender systems are designed to support restaurant selection by assisting consumers with the information overload problem. However, despite their promises, they have been criticized of insuficient performance. Recent research in recommender systems has acknowledged the importance of personality in improving recommendation; however, limited work exploited this aspect in the restaurant domain. Similarly, the importance of user preferences in food has been known to improve recommendation but most systems explicitly ask the users for this information. In this paper, we explore the influence of personality and user preference by utilizing text in consumers' electronic word of mouth (eWOM) to predict the probability of a user enjoying a restaurant he/she had not visited before. Food preferences are extracted though a trained named-entity recognizer learned from a labelled dataset of foods, generated using a rule-based approach. The prediction of user personality is achieved through a bi-directional transformer approach with a feed-forward classification layer, due to its improved performance in similar problems over other machine learning models. The personality classification model utilizes the textual information of reviews and predicts the personality of the author. Topic modelling is used to identify additional features that characterize users' preferences and restaurants properties. All aforementioned features are used collectively to train an extreme gradient boosting tree model, which outputs the predicted user rating of restaurants. The trained model is compared against popular recommendation techniques such as nonnegative matrix factorization and single value decomposition.
leisure trip that plays a central role in travel experience.
Recently, interest on food experience has been growing [1], with businesses in the hospitality sector seeking insights regarding the dining behaviors and preferences of customers to improve decision making in areas such as marketing [2] and recommendation [3]. Past research that utilizes food in recommender systems such as [4] employ simple techniques such as frequencies of food vocabularies in Bag of Words. However, such techniques require a lexicon of complete list of foods that usually is not available for diferent cousins and countries. In this paper, we utilize implicit and explicit information of consumers’ eWOM to improve restaurant recommendation. Implicit information refers to textual comments in reviews that can be used to estimate consumers’ pernEvelop-O ences and needs [6]. The application of users’ personality has enhanced the performance of recommender systems (H. Herodotou); 0000-0001-5481-1560 (M. Pampaka) © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License in the tourism domain with results improving point of interest, destination recommendations, utilizing either
questionnaires or automated personality recognition.
This paper illustrates the utilization of consumers’ personality and user preferences extracted from the textual This section provides a review of recommendation techpart of electronic word of mouth (eWOM) to improve niques, the concept of personality, and elaborates on how recommendation. EWOM represent consumer opinions it has been used in recommender systems so far. about products and services and has been used extensively in identifying consumers’ preferences. Recom- 2.1. Restaurant Recommender Systems mendations are made by training an Extreme Gradient Boosting (XGBoost) prediction model using as features, Recommender systems aim to predict the satisfaction of a the users’ personality and the users’ preferences (e.g., consumer with an item (product/service) he/she has not food). XGBoost is used due to its good performance bought yet [11]. This is part of one-to-one marketing that in similar recommendation problems [7]. The research seeks to match items to consumers’ preferences in conquestion addressed in this paper focuses on whether the trast to mass marketing aiming to satisfy a target market integration of personality with other features inferred segment [12]. Popular approaches focus on consumers’ from structured and unstructured parts of online reviews past experiences (ratings) for the creation of a user-item improves restaurant recommendation, in contrast to pop- matrix and based on that predict what is more approular model-based collaborative filtering (CF) techniques priate to a user depending on either similarity between such as nonnegative matrix factorization (NMF) and sin- users or items (products, services) [11]. The relationship gle value decomposition (SVD). between consumers or between products can be found
The proposed approach utilizes consumers’ food pref- using similarity metrics, and this method is known as erences and personalities along with perceptions about Collaborative Filtering (CF) [13]. This has been successvenues from eWOM to recommend most suitable restau- fully applied in tourism recommendation problems such rants to tourists. Labelled personality data is utilized as hotels or points of interests, and is considered as one of to train a BERT (Bidirectional Encoder Representations the most popular techniques [14]. Another popular techfrom Transformers) classifier using the personality model nique is content-based filtering, that attempts to guess of Myers-Briggs Type Indicator (MBTI) due to its good what a user may like based on items’ features rather than results in previous studies [8]. User preferences are ex- their rating [13]. A hybrid approach takes the advantage tracted through topic modelling and a trained food named of both content-based filtering and collaborative filtering entity recognizer. An XGBoost model is generated to pre- [11]. dict the probability of a user liking an unvisited restau- CF techniques, however, sufer from the cold start probrant based on its personality, preference, and themes that lem that occurs when very little or no data is available characterize the venue. about a user and thus inability to identify similar con
The research question addressed in this work is how sumers [15]. In addition, data sparsity exacerbates the to best combine user preference and personality mod- problem when there are a lot of unrated items in the userels with topic features inferred from eWOM to produce item matrix. This occurs when there is not enough data the best recommendation, in contrast to popular model- to populate the user-item matrix based on which to make based collaborative filtering (CF) techniques. This is a reliable inferences [16]. In tourism, the collection of data continuation of our previous work in [9, 10] that exam- is dificult and time-consuming due to the limited time ine the use of personality and emotion in recommender that tourists spent at a destination. The cold start probsystems. The contribution of this work lies in the auto- lem appears with first-time users (tourists) since there are mated detection of food preferences from eWOM and its no records of their purchasing activity at a specific descombination with user personality and topic modeling tination.To address these CF problems, recent methods for restaurant recommendation. utilise machine learning techniques such as matrix fac
The paper is organized as follows. The next section torisation to approximate the user-item matrix content introduces background knowledge on restaurant recom- using latent variables that emerge from the initial data. mender systems and techniques for extracting food pref- The singular value decomposition (SVD), optimized SVD erences and personality from text. The next section de- (SVD++), and non-negative matrix factorization (NMF) scribes techniques for identifying topics discussed in con- models factorize the user-item matrix and predict the sumers’ eWOM and personality prediction using deep satisfaction of users for products that are unknown [17]. neural networks. Subsequent sections elaborate on the Alternatively, content-based approaches utilize metadata methodology followed and the results obtained. The pa- about new products to address the cold start problem. per concludes with the discussion and future directions. A useful source for obtaining these metadata is textual information from eWOM and its analysis using text analytics [18]. An example application includes work by Sun et al. [19] that improved CF performance by analysing restaurants eWOM to define numerical features corre- mechanism for defining the most essential aspects of persponding to consumers satisfaction through sentiment sonality that describes people characteristics that creates analysis. In the same vein, topic modelling techniques and reflects their behaviour [ 28]. have been used with CF to assist in estimating the simi- Personality prediction is an important phase of larity between consumers or items [20]. Finally, work by personality-aware recommender systems, and the two Zhang et al. [21] used consumers or items characteristics main methods for doing so is through questionnaires and to cluster them into groups, and then find correlations automated means. Generally, questionnaires are more between clusters to address the data sparsity problem. accurate in assessing personality; however, the process
Recently, a strong interest emerged in using the per- is tedious while the automated approach is easier to consonalities of consumers in an efort to better understand duct, by utilising user’s existing data that can be either and match their needs, as “personality” relates to the text, images, videos, likes (behavioural data) etc. [18] perceptions, feelings, motivations, and preferences of in- Predicting personality from text is a popular automated dividuals [22]. The application of user personality has approach that is based on personality theory claiming improved the performance of recommendations in the that words can reveal some psychological states and pertourism domain for points of interest compared to tradi- sonality of the author of the text. There are two main tional methods [23]. Personality-based recommendations categories of techniques, the feature-based and the deep have also been shown to greatly reduce the cold start and learning: the former uses unigrams/n-grams (open vodata sparsity problems, and improved the performance cabulary approach) or lexicons (closed vocabulary) of of recommendations in areas such as online advertising, features relevant to personality, and the latter text emsocial media, books, and music [24]. However, these beddings learned from large corpus of text in an unsuperapproaches do not take advantage of eWOM data from vised manner (language models). Popular feature-based users on the web to extract their preferences and their methods utilize the Mairesse [29] and linguistic inquiry personalities. They focus mainly on the extraction of and word count techniques [30]. Features from these are user data from specialized questionnaires to collect con- fed into diferent machine learning classifiers (e.g., Naïve sumers’ behaviours and personalities. Such approaches Bayes, support vector machines) to make predictions. fail to continuously update the system because of the Obtaining such features however is a costly process and time-consuming use of questionnaires that leads to the cannot efectively represent the original text semantics. loss of automation and update limitations. To avoid feature engineering, deep neural models and language models are employed to learn text representa2.2. Personality Extraction from Text tions that currently result in improved accuracy. Deep models focus on the context of the text and not just a Personality is a set of characteristics and behaviours of an static representation for a word or a sentence. Those kind individual that influence many areas of his/her life such of deep learning techniques are using an attention mechas motivations, preferences, as well as consumer pref- anism that focuses on giving weights to words based on erences and behaviour [23]. Applications of automated how they are used in a text giving the ability of capturpersonality predictions have been applied by researchers ing the semantic content [31]. A popular architecture is on data from various social networks such as Facebook, the BERT (Bidirectional Encoder Representations from Twitter, to explore correlations between personalities and Transformers) that utilizes transformers neural network the diferent user activities, purchasing behaviors and architecture. Attention-based transformers have shown liking of foods from specific cuisines [ 25]. that collecting the semantics of a text improves the per
The two most popular text-based personality classifi- formance level and the predication accuracy of ML percation methods are based on the Myers-Briggs Type Indi- sonality models [32]. Given this, the method proposed in cator (MBTI) [26] and the Big Five [27] personality traits this paper utilizes attention-based personality prediction. due to the availability of labelled data on these models. Most approaches use a binary classifier for each of The classifiers with best performance are usually employ- the personality traits (MBTI) such as a classifier for ing the MBTI personality model that focuses on 8 key extraversion-introversion etc. Such methods require pre types of characteristics that people have, Extraversion or labeled data with the personality class. The first step in Introversion, Sensing or Intuition, Thinking or Feeling, the process is the vectorization of the text into a form and Judging or Perceiving, behaviours. The combina- that can be processed by ML algorithms [33]. This can be tion of characteristics can shape 16 diferent personality done using open/closed lexicons or sentence embeddings types and classify people to the proper personality clus- in the case of deep learning methods (BERT). The vectorter [13]. The Big 5 Personality model express personality ized data is used to train a classifier using the data label in the following 5 dimensions: Agreeableness, Extraver- or fine tune a pretrained model in the case of BERT. The sion, Openness to Experience, Conscientiousness, and trained and validated model can be used to predict unseen Neuroticism. Such taxonomies are recognized as a valid data. Recent personality classification techniques that The proposed method utilises a named entity recognition to extract food preferences, an automated eWOM topic modelling for the identification of themes discussed in review’s text, a BERT-based personality classification, and an ensemble tree-based regression for the prediction of consumer restaurant ratings.
A named entity recognition (NER) is utilized to extract the food preferences of customers. NER is a major component in NLP systems to extract information from unstructured text. An entity can be any word or sentence that refers to the concept of question. There are two main approaches for the creation of a NER, the modelbased, and rule-based approach. The latter focuses on the grammatical rules and linguistic terms to extract entities.
The model-based approach generates machine learning models using a text with pre-labelled entities. Most food NER models as reported in [35] are trained on data that did not include Cyprus dishes thus their predictions were insuficient for our case. There is a wide range of generic libraries suitable for NER, such as NLTK, spaCy, Stanford NER, Stanza, and Flair, but none was able to provide appropriate food recognition in text.
To extract food preferences, the spacy library was utilized, and several rules have been specified that enabled the extraction of sentences that mention food consumption such as “I ate ”,“I had for dinner” etc. To generate a suficient training set, a local and international food recipe dataset was used. The returned sentences were annotated automatically based on the position in the sentence where the food entity occurred. This was necessary in order to create a training dataset labelled with food names, and their start/end position in the sentence, based on which the food NER training was performed.
The trained NER achieved an overall accuracy of 81% (70/30 train-test split) and was applied on the restaurant reviews to extract foods associated with each review. Due to the large number of food entities that were generated, there were a lot of repetitions due to diferent spellings.
Thus, to reduce the dimensionality of the dataset, a feature selection process was performed using a random forest machine learning model to identify the most important food names using the review ratings as the target variable. The process yielded the optimum number of feautilize deep learning for Big Five personality prediction, tures (220) that resulted in the best model performance. such as the DeepPerson [34] demonstrate classification The selected food features where then one-hot encoded performance (AUC score) of around 70% per personal- for each consumer review. To identify the food preferity dimension, using diferent training datasets, which is ences of each user, reviews were grouped by user and the much lower compared to classifiers that use MBTI data. most frequent food entities in each user’s reviews were considered as food preferences. This process considers that, when customers visit diferent restaurants and write 3. Technical Background comments about the food they ordered, irrespective of the food’s quality and the review rating, it constitutes food preference of the user.
Topic modelling is a popular tool for extracting information from unstructured data and is used in this work to identify themes discussed by consumers in eWOM. Topic models generally involve a statistical model aiming at ifnding topics that occur in a collection of documents [36]. Two of the most popular techniques for topic analysis are the Latent Dirichlet Allocation and the Structural Topic Model (STM). In this study, the STM approach [37] is used to develop a topic model due to its ability to incorporate reviews’ metadata such as sentiment (rating>3) that help with interpreting and naming the identified topics. Each topic in STM represents a set of words that occur frequently together in a corpus and each document is associated with a probability distribution of topics per document. The process for learning the topic model initiates with data preprocessing that includes removal of common and custom stop-words and irrelevant information (punctuation), followed by tokenization (breaking sentences into word tokens), and stemming (converting words to their root form). Initially, common stop-words were considered and gradually with the refinement of the model, additional stop-words that were irrelevant to our goal were added to the list of custom stop-words such as names of people, restaurants, cities, etc. The optimum number of topics that best fits the dataset is identified through an iterative process examining diferent values for the number of topics (K) and inspecting the semantic coherence, held out likelihood, and exclusivity of the model at each iteration until a satisfactory model is produced [37]. Coherence measures the degree of semantic similarity between high scoring words in the topic. Held out likelihood tests a trained topic model using a test set that contains previously unseen documents. Exclusivity measures the extent to which top words in one topic are not top words in other topics. The naming of the topics was performed manually based on domain knowledge and the most prevalent words that characterize each topic.
BERT uses a multi-layer bidirectional transformer encoder and is inspired by the concept of knowledge transfer, since in many problems it is dificult to access suficiently large volume of labelled data to train deep models. In transfer learning, a pre-train a model is learned from massive unlabeled datasets not representing the target problem, but allows the learning of general knowledge. BERT-like approaches provide pretrained models and their embedded knowledge can be transferred to a target domain where labelled data is limited. Fine-tuning such models is performed using a labelled dataset representing the actual problem; these tune the model to the task at hand. Fine-tuning adds a feedforward layer on top of the pre-trained BERT. Previous work has demonstrated that this pre-training and fine-tuning approach outperforms existing text classification approaches. In our case, fine-tuning the BERT model was performed using publicly available personality labelled data. BERT has been used for personality prediction using the Personality Cafe MBTI dataset in [38] achieving an accuracy of around 0.75. In contrast, other deep learning methods that use the Big 5 model as well as the popular streamof-consciousness essay dataset such as the one reported in [39] using CNN, achieve inferior classification performance.
Despite their good results, BERT-based approaches have been criticized that their best performance is reported with short texts. Long text refer to text with more than 512 tokens. Such text however are computationally expensive to process thus most transformers models limit the number of tokens they can process simultaneously. In our case, most reviews produced by consumers exceeded the 512 tokens limit and thus the prediction of personality was considered as a long text classification problem. Diferent methods exist to dealing with this issue, which include the naïve head-only, tail only or semi-naïve approaches, that either use the top number of words, bottom number of words, or combination of top/bottom/important words in the text. Such approaches lose information but have a minimum computational cost. Recent works have sought to alleviate the computational cost constraint by applying more sophisticated models to longer text instances such as dividing the long text into chunks and combining the embeddings of the chunks. However, work by Sun et al. [40] that investigated diferent long-text treatment methods for consumer reviews, showed that the best classification performance is achieved using naïve methods such as using only the head or tail tokens of the text while dropping all other content. In this work, we explore the naïve and semi naïve methods to find the one with the best personality classification performance prior to labelling users with their personality. The results, described in a subsequent section, show that the naïve approach yielded the best performance, which is in line with [40].
XGBoost regression is used in this study due to its ability of producing good results in similar problems. It is an ensemble method; hence multiple trees are constructed with the training of each tree depending on errors from previous trees’ predictions. Gradient descent is used to generate new trees based on all previous trees while optimizing for loss and regularization. XGBoost regularization component balances complexity of the learned model against predictability. XGBoost optimization is required to minimize model overfitting and treating data imbalance, by tuning multiple hyperparameters. The optimal values of hyperparameters can be determined with diferent techniques such as the exhaustive (grid search), Bayesian, or random. The grid search method combines all possible values of each parameter, to obtain the model with best performance, while the Bayesian utilizes results from previous optimization cycles to identify hyperparameters values with higher probability in improving the classifiers performance. Grid search is better but slower while Bayesian is faster but not as accurate. In this work, the grid search approach is adopted. 4. Methodology matrix is generated with rows corresponding to consumers and columns to restaurants. The cells of the matrix contain ratings when these are available since customers did not visit all restaurants; 3. Development of a topic model using as corpus the eWOM (reviews) to identify consumers’ opinions and how these are associated with each review. Restaurant’s topics are generated by averaging the topics theta values associated with each restaurant. This represents common consumer opinions per restaurant; 4. Assessment of customers’ personality from eWOM is achieved using the MBTI BERT personality classification model; 5. Food preferences of users are extracted from eWOM’s text using a custom NER model; 6. The explicit information from each restaurant is combined with implicit information that emerges from personality analysis, food preference, and topic modelling. These features are used collectively to enhance the user-item matrix and are used to train an Extreme Gradient Boosting (XGboost) regressor model using as output variable the user rating of restaurants and taking values in the range [1-5]. The XGBoost is optimized using hyperparameter tuning and validated using train/test data split (70/30). The trained model is used to predict user ratings for restaurant users have never visited; 7. The performance of the XGBoost model is compared against that of three popular model-based CF techniques, namely SVD, SVD++, and NMF. The comparison models are trained using the initial user-item matrix while the XGBoost using the enhanced user-item matrix that includes explicit and implicit information. The performance of the models is assessed using popular evaluation metrics such as mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE).
The data collected includes 105k reviews written in En1. Collection of restaurant reviews from TripAdvi- glish from tourists who visited Cyprus between 2010 sor and extraction of consumers’ eWOM and ad- to 2020 and posted reviews about their experience with ditional explicit information of restaurants such restaurants in Cyprus (publicly available). The total numas cuisine type, price range, and value for money; ber of unique users were 56800 and the number of restau2. Preprocessing of the data and preparation for sub- rants were 650. Figure 2 depicts descriptive statistics of sequent analyses (topic modelling, personality reviews ratings per year. For this study only users with classification). Preprocessing includes punctu- at least 20 reviews are considered and only restaurants ations and URLs elimination, lowering of text, with at least 50 reviews yielding 93 unique users and 410 stop words removal, tokenization, stemming, and venues.
lemmatization. During this step, the user-item
an STM topic model was developed using the estimated optimum K (30) number of topics based on the model’s performance metrics in Figure 3, with focus on high coherence, high held-out likelihood, low residuals, and high lower bound scores.
The training of the binary classifiers was performed using the Personality Cafe MBTI dataset consisting of joint user posts on a social network labeled by personality type defined using MBTI questionnaire. The dataset is publicly available on “Kaggle” [41]. To identify the BERT long text approach with the best classification performance, two techniques were examined, namely the naïve and semi naïve approach and the one with the best perFigure 3: Topic performance measures for identifying the formance was used in the workflow. For the naïve apoptimum number of topics. The red circle indicated the K proach, we used the head-only using as sentence length number of topics selected the 256 and 512 words and for the semi-naïve we used chunking of text into 128 words and combining their embedding. The results from this process, presented in
The naming of the topics in Table 1 was based on Table 2, showed that the 512-naïve-head approach outperdomain knowledge, words with highest probability in formed the other approaches and thus it was employed in each topic and words with high Lift score. Lift gives users’ personality classification. Results from the BERT higher weight to words that appear less frequently in model outperformed personality models trained using other topics. the same dataset and thus improved our confidence in
The probability distribution of topics per review de- the personality prediction of each user. notes the probability of each topic discussed in a review The personality distributions in Figure 5 show descripand the sum of all topics’ probabilities in each review tive statistics regarding the personalities of users accordtotal 1. Reviews are associated with the distribution of ing to the detected personality from the MBTI BERT clastopics prevalence per review. The trained STM model’s sifier fine-tuned using a labelled datasets and treating theta values per review refer to the probability that a long text using the naïve-head approach with 512 tokens. topic is associated with each review. These theta val- The acronyms refer to combination of dimensions of the ues, shown in Figure 4, were used as features during the MBTI model. The trained BERT model predicts for each training of the XGBoost model along with other features. dimension of the personality model the probability that a user belongs to any of the personality traits (i.e., probability for Extraversion – Introversion (E/I), Sensing –
Words with high probability and lift scores great, really, music, live, day, atmosphere, enchiladas, music, really, live, pub nice, prices, atmosphere, reasonable, big, family, polite, quick, nice, relaxing, families, cafe time, excellent, went, night, amazing, first, every, occasions, stay, went, amaze, week eat, new, end, found, places, second, thai, always, second, none lovely, recommend, highly, enjoyed, beautiful, setting, party, setting, party, hosts, fabulous, thoroughly, absolutely well, lunch, local, attentive, wonderful, presented, chose, breaks, attentive, presented, chose evening, bar, friends, though, group, customers, quiet, whiskies, though restaurant, location, must, beach, view, right, perfect, definitely, must, far visit, will, back, really, worth, definitely, going, visit, called wedding, amazing, even, similar, impression, organize, guest, events, beyond, pleasure many, birthday, soon, booked, kitchen, also, october, good, love see, year this, flight, travel, celebration experience, nothing, special, whole, maybe, dining, perfection, fiancée, maybe restaurant, probably, also, mountains, open, available, well, best, more, owner, managers, troodos summer, use, even, range, late, cool, evenings, use, dine, cozy always, can, class, owners, restaurant, number, first, classy, number, varied, feeling, interesting, hidden two, outside, can, inside, sit, world, get, disappointed, aircon, magic, noise, trafic, heat thai, tourist, across, gem, trying, partner, avoid, duck, again, overall, always, bespoke, gimmicks, hardcore bit, little, better, average, like, quite, expensive, much, however, criticisms, average, diferent, small, cheese, also, breakfast, euros, greek, platter, platter, options, vegetarian, bacon, eggs wife, return, disappointed, restaurant, reviews, favourite, holiday, isn’t, trip, done old, cypriot, road, stop, village, along, street, waitresses, road, walk busy, get, people, table, lot, need, without, joyful, early, book years, restaurant, made, visiting, coming, several, since, ago, forward, visits staf, friendly, always, see, come, welcoming, feel, chat, truly, smile, come chips, served, priced, set, large, course, portion, adults, chips, portion, reasonably best, cyprus, don’t, ever, never, restaurants, know, traditional, meze value, money, recommended, excellent, variety, high, meals , bringing, best, cyprus, eaten couldn’t, enough, friend, eat, fresh, away, wow, take, out, basilica, excellent, lovely, rice, more, time ,highly ordered, came, table, order, waiter, asked, arrived, minutes, waited, waitress, left, seated, orders, bill just, restaurant, basic, way, like, standard, much, full, unacceptable, much, restaurant Topic Name
Entertainment Atmosphere Family Restaurant Special Occasion New Place Party Place Lunch Evening/Bar Location Worth Visiting Wedding Place Celebration parties Not Worth Out of town Summer location Fabulous Place Outside eating Asian Cuisine Average Place Breakfast Disappointment Stop during trips Busy Place Visit over years Welcoming Staf Good portions Traditional foods Value for money Fresh ingredients Bad service Nothing Special Intuition (S/N), Thinking – Feeling (T/F), and Judging – which is depicted in Figure 5. The BERT model deals Perceiving (J/P)). Combinations of letter from each cate- with 4 classifiers, one for each of the dimensions above. gory generate 16 four-letter personality types: ISFJ, INFP, The classifier’s average area under the curve (AUC) perINFJ, ISTP, ISTJ, ISFP, INTP, INTJ, ENTP, ESFP, ENFP, formance is 87%. This is an improved performance comESFJ, ESTP, ESTJ, ENFJ and ENTJ, the distribution of pared to alternative personality classification techniques personality-based approaches over these baseline models. The traditional techniques were also optimized by tuning two hyperparameters, the number of factors and the regularization value.
In the experiments conducted using the aforementioned restaurants reviews, the data was initially split into test and training sets (70/30) using stratified sampling to guarantee that all user ratings are suficiently represented in the test and training samples. The models were hyper tuned, trained, and tested using the same samples. The aforementioned metrics were computed, and the results that emerged (see Table 3) show that the MBTI XGBoost model produced the best performance among all other models. Both personality-based models outperformed traditional approaches, which indicates that the use of personality and eWOM-extracted topics improved the recommendations.
user personality restaurant recommendation approach and constitutes one of the first studies that use customer 5.3. Training and Evaluating the XGBoost preference along with personality in the restaurant recModels ommendation problem. It utilizes a popular personality model (MBTI) to enhance the restaurant recommendaThe enhanced user-item matrix that emerged from the tion process by fine tuning a BERT classification model personality model, food preferences, and the topics asso- on personality labelled dataset. Due to the length of ciations per user and venue were used to train an XGB the training data, the best long-text handling approach regressor model. (naïve-head 512) was employed during BERT model tun
The XGBoost model underwent hyperparameter tun- ing. EWOM themes are extracted through topic moding prior to training by tuning the models’ learning rate, elling from eWOM’s text and are also used as additional gamma, subsample, and regularization options using grid features of restaurants and users that refer to implicit search. Traditional recommendation models, namely preferences of users and properties of restaurants. All SVD, SVD++, and NMF were generated using the sur- aforementioned features are used collectively to train an prise python library. The models were compared based XGBoost regressor to predict consumers’ satisfaction (i.e., on the following performance metrics: the mean abso- rating) for unvisited restaurants. The results show that lute error (MAE) that represents the average of the ab- the MBTI model in combination with topics from eWOM solute diference between the real and predicted values, outperforms the model-based collaborative filtering techMean Squared Error (MSE) and Root Mean Squared Er- niques, ofering a first indication that the application of ror (RMSE) that is the square root of MSE. Comparison personality and food preferences in restaurant recomof the two models against traditional recommendation mendation can have valuable results. Future work will techniques revealed an improved performance of the focus on evaluating additional long-text handling tech