3.1. Dataset and Data Preprocessing

Two datasets are used in the proposed POI recommendation system using SA: the election dataset for SA and the Foursquare dataset for POI recommendation.

3.2. Sentiment Analysis Model

Fig. 2 presents the complete data preprocessing and model-building pipeline for a SA model using the Bag of Words (BOW) approach. The steps involved in the design and implementation of the proposed SA model are as follows.

Fig. 2

Schematic diagram of sentiment analysis model. BiLSTM, bidirectional long short-term memory.

Loading and Cleaning the Election Dataset: The first step is to import all necessary libraries, such as pandas, NumPy, Keras, sci-kit-learn, matplotlib, nltk, etc., and load the election dataset, which contains text and sentiment columns. Any null or missing numbers, duplicate data, and fields that are not needed are taken out of the data to clean it up.

Preprocessing: Some classes of sentiment columns have significantly fewer samples than others. Equalizing the class distribution of the sentiment score column ensures that the model is not biased toward any particular class. It under-samples each sentiment label to the specified number of samples using the resample method of sklearn.utils library. The next step is to tokenize the text data, which involves breaking down the text into a list of individual words or tokens. This process also involves removing special characters, numbers, punctuation, and non-letter characters using regular expressions, and then removing the stop words using the nltk library, which involves eliminating common terms with little or no meaning, such as ‘a,’ ‘the,’ ‘and,’ and so on, and in the last step lemmatization to convert words to their base form is performed.

Feature Extraction: Feature extraction is performed using the BOW approach in the proposed model. It transforms the text data into a token count matrix. The CountVectorizer method of the sklearn.feature_extraction.text library is used to perform the BOW feature extraction.

Splitting the Dataset: The cleaned and preprocessed dataset is divided into 70% train and 30% test sets. The training set is used to train the BiLSTM model, while the test set evaluates the model’s performance.

BiLSTM Model for SA: Input text is first transformed into a fixed-length vector representation by an embedding layer; next, a BiLSTM layer is added to the model to learn about the text’s context and meaning. In conclusion, a dense layer with a softmax activation function is utilized to predict the sentiment of the text. The architecture of the BiLSTM model is shown in Table 2.

The BiLSTM performs both forward and backward evaluations of the input sequence, allowing gathering contextual information from past and future aspects of the input sequence, making the method well-suited for NLP applications such as SA. The BiLSTM model takes preprocessed text data as input and returns a sentiment score that might be positive, negative, or neutral. The model is built with cross-entropy loss and optimized with the Adam optimizer. Various metrics are used to determine the effectiveness of the proposed SA model using the testing set. After testing, the SA model is saved to predict the sentiments of a Foursquare text review as positive, negative, or neutral.

3.3. POI Recommendations Model

In the “POI Recommendation System using sentiment analysis,” the Foursquare dataset contains reviews for different POIs and their corresponding user IDs, POI IDs, and coordinates. Fig. 3 depicts a schematic diagram of the POI Recommendation Model, which includes the system processes. The Foursquare dataset review column is fed into the SA model to predict sentiment scores, which are then used to recommend POIs to users based on their choices.

Fig. 3

Schematic diagram of POI recommendation model. POI, point of interest; LSTM, long short-term memory.

Rows containing null or missing values, duplicate rows, and unnecessary columns are dropped as a part of data preprocessing. The model uses tokenization, removing stop words and other unnecessary words, and lemmatization to preprocess the text review data of the Foursquare dataset. In tokenizing and padding the review data, the function tokenize_pad_sequences() takes in the text data and tokenizes it using the Tokenizer() method from the Keras library. It then converts the tokenized text into sequences of integers using the texts_to_sequences() method of the Tokenizer object. Finally, it uses the pad_sequences() method from TensorFlow.Keras library for padding the sequences with zeros to ensure sequences of the same length. The preprocessed text is then fed into the trained SA model, which predicts the sentiments. The predict_class () function is used to predict the sentiment class of each text entry, which can be negative, neutral, or positive.

The predict_class() function takes in a text as input, tokenizes and pads it using the same Tokenizer and padding method as before, and uses the loaded model to predict the sentiment class of the text. The model takes in a sequence of words, and for each word, it produces an output that is used to predict the sentiment score for the entire sequence. The function returns the predicted sentiment class. The predicted sentiment scores are -1, 0, and 1. Here, 1 represents positive sentiment, 0 represents neutral sentiment, and -1 represents negative sentiment.

Once the sentiment scores are predicted, they are added as a new column in the Foursquare dataset, called “score.” These scores are then used to filter and recommend top-k POIs with positive or neutral sentiment to the user.

To recommend POIs, the user and POI embedding are generated using the cleaned and preprocessed Foursquare dataset. User embeddings are a vector representation of each user based on their interactions with POIs. In contrast, POI embeddings are a vector representation of each POI based on the sentiment scores of the reviews. The embeddings are generated using a Word2Vec technique in the gensim library.

After applying the sentiment model, user and POI embedding are generated to represent each user and POI in a vector space. These embeddings are used as input features (X), and the predicted sentiment (score) is used as the target variable (y) for splitting and training the LSTM model. The dataset was then split 70:30 into train and test sets. Next, the system trains the proposed LSTM model as the architecture shown in Table 3 displays using the embedding and the user’s ratings for the POIs. The trained model predicts the sentiment of POIs based on user and POI embeddings.

Then, the test dataset is used and preprocessed by tokenization, removal of unnecessary words, and lemmatization to filter the dataset based on the user’s interests. Further, a dataset is created using the list of keywords related to the user’s interest (a few users’ interests are defined, such as restaurant, museum, mall, library, hotel, educational, and emergency services). The process filters the data by keeping only the rows where the preprocessed “text” column contains keywords related to the particular user’s interest. It then shortens the resulting data. For example, if a user is interested in “restaurants,” we will filter the Foursquare dataset to include data related only to restaurants, cafes, and bars. Similarly, if a user is interested in “museums,” we will filter the dataset only to include data related to museums such as museums, galleries, and historic sites.

This step ensures that the recommended POIs correspond to the user’s interests, making it more likely that the user will visit and appreciate the places. To create the list of keywords related to user interest, we can combine manual curation and automated techniques, such as keyword extraction from users’ previous check-ins or search queries. Applying a filtering mechanism to the dataset according to the user’s interests guarantees that the suggested POIs are directly related to the user’s preferences and desires. This enhances the user’s overall experience and level of satisfaction.

The next step in the recommendation model is to take user input for their ID, location, interests, and the number of top recommendations they want. Based on this input, the system filters the dataset to show POIs that match the user’s interests and are within a certain radius of the user’s location.

After the dataset has been refined based on user interests, it is further filtered based on the user’s location. It ensures that the recommended POIs are within a reasonable distance of the user’s current location. The filtering is done by calculating the distance between the user and the POI’s location using the coordinate information available in the dataset, increasing the likelihood that the user will visit and enjoy the places suggested. The top POIs with positive or neutral sentiments are recommended to the user.

This study aims to give users tailored suggestions based on their interests, geography, and sentiment score. These recommendations can be positive or neutral, considering the sentiment of previous evaluations. Overall, this study blends SA with geographically specific recommendations.

4.1. Sentiment Analysis Model

Table 4 displays the performance metrics of the SA model after 50 training epochs. These evaluation metrics showed that the model is good at categorizing the sentiment of text data (reviews) as positive, negative, or neutral. The model has an accuracy of 0.9952, which means that 99.52% of all reviews in the dataset were correctly classified as positive, negative, or neutral sentiment, a precision of 0.9953, which means that 99.53% of all reviews predicted as positive sentiment were positive, and a recall of 0.9951, which means that 99.51% of all actual positive reviews in the dataset were correctly predicted as positive sentiment by the model, and F1 score of 0.9952, showing that it predicts the sentiment of reviews in the election dataset fairly accurately. The comparative results of the proposed SA model and other state-of-art problems (Alemayehu et al., 2023; Dang et al., 2020; Khruahong et al., 2022) are also presented in Table 4, which clearly shows that the proposed models outperform the others in performance. Further, it is observed that the proposed models outperform others when applied to the other datasets, such as Tweets Airline and Amharic Political.

A high F1 score suggests that the model has an appropriate balance of precision and recall, which are important for effective SA. Furthermore, the model consistently obtained good accuracy, precision, recall, and F1 score across 50 epochs, implying that the model is stable and does not overfit the data. The model’s improved performance, as shown in Table 4, can be due to implementing a BiLSTM with feature extraction using BOW, a deep learning model capable of properly capturing the sequential information in the review. Furthermore, the elimination of stop words and tokenization of the text data improved the model’s performance by lowering noise and boosting the quality of the input data. Overall, these findings illustrate that the SA model is highly accurate in predicting the sentiment of reviews and may be applied to analyze reviews in other datasets, as shown in Table 4.

Fig. 4 illustrates the accuracy and loss of the proposed SA model. It shows that our model is working accurately. Validation and training accuracy increase as the epochs increase. Similarly, validation loss and training loss decrease as the epochs increase.

Fig. 4

Graphs of accuracy and loss of sentiment analysis model.

The SA model’s confusion matrix is displayed in Fig. 5. This method may be used to gauge how well the SA model is doing. Results show that 24,372 negative feelings, 24,373 neutral sentiments, and 24,469 good sentiments are all accurately predicted. The dataset also shows that the number of erroneous predictions is relatively low.

Fig. 5

Confusion matrix of sentiment analysis model.

4.2. POI Recommendation Model Using SA

The proposed POI recommendation model is subsequently evaluated with the help of a test set that consists of thirty percent of the dataset. The model is trained using the training dataset, while the testing dataset determines the system’s performance.

The POI recommendation model’s output is the recommended POIs based on the user’s input. The recommended POIs are ranked based on the SA of the reviews linked with each POI and are filtered depending on the user’s location and interest. When filtering data based on user interest, the model’s performance may vary according to user input. If the user input is correct or complete, the predictions and accuracy of the recommendations may be improved.

Overall, the POI recommendation model provides a useful feature for users to discover and explore new locations of interest based on their location and SA. If the recommended POIs are relevant and have positive sentiments, they can be considered successful recommendations. Further research can explore ways to improve the model’s performance and accuracy, such as incorporating additional features, using more advanced deep learning models, or optimizing the model architecture. With further improvements and enhancements, the system has the potential to become a valuable tool for users looking to discover new and interesting places to visit.

The recall@k results for the proposed top POI recommendation model are provided in Table 5, along with a comparison to the outcomes of preceding models. The proportion that recall@k calculates out of the entire number of relevant items that might have been suggested is the percentage of relevant things that were recommended out of the total number of relevant items. Even though Zhou et al. (2019) have recall@5, recall@10, recall@15, and recall@20 values of 0.06, 0.10, 0.12, and 0.14, respectively, the recommended model recall@k climbs as k increases and reaches over 50%. Liu and Wu (2021) concluded that the values for recall@5, recall@10, recall@15, and recall@20 are, respectively, 0.19, 0.23, 0.28, and 0.32. Their research obtained this knowledge. For recall@5, recall@10, recall@15, and recall@20, respectively. Wang et al. (2019) obtained the values of 0.27, 0.32, 0.399, and 0.47. These values were recorded in their study. The researchers collected all of these values. Zhang et al. (2019) state that recall@5, recall@10, recall@15, and recall@20 values are 0.09, 0.15, 0.17, and 0.20, respectively. These values may be compared to one another. This set of values was obtained in a related manner. Yu et al. (2019) established that the values for recall@5, recall@10, recall@15, and recall@20 are, respectively, 0.12, 0.16, 0.19, and 0.22. The values Dai et al. (2019) found for recall@5, recall@10, recall@15, and recall@20 are 0.14, 0.17, 0.22, and 0.24, respectively. The researchers collected all of these values. It has been determined by Lu et al. (2019) that the values of recall@5, recall@10, recall@15, and recall@20 are 0.17, 0.24, 0.25, and 0.28, respectively. It was possible to acquire these values. The proposed model performed significantly better in terms of recall@5, recall@10, recall@15, and recall@20 when compared to the research works published by Zhou et al. (2019), Liu and Wu (2021), Wang et al. (2019), Zhang et al. (2019), Yu et al. (2019), Dai et al. (2019), and Lu et al. (2019). This was determined by comparing the suggested model to the research works performed by the authors above. Fig. 6 compares the proposed model and studies for recall@5, recall@10, recall@15, and recall@20.

Fig. 6

Performance of proposed POI recommendation model for recall@k versus top k POI recommendations. POI, point of interest.

Table 6 shows how the proposed model stacks up against other state-of-the-art recommendations for top POIs in terms of precision@k. Precision@k is a measure that determines the percentage of the top k suggestions that are relevant or wanted by the user. It is evident from Table 6 that the precision@k of the suggested model drops with increasing k and approaches 38% approx. Still, the results for precision@5, precision@10, precision@15, and precision@20 were 0.07, 0.057, 0.049, and 0.043, respectively, according to Zhou et al. (2019). The precision@k values were not observed by Liu and Wu (2021). According to Wang et al. (2019), the various precision values were 0.38 for precision@5, 0.345 for precision@10, 0.326 for precision@15, and 0.288 for precision@20. According to the study by Zhang et al. (2019), the accuracy values for precision@5, precision@10, precision@15, and precision@20 were 0.12, 0.011, 0.01, and 0.005, respectively. Yu et al. (2019) documented precision values of 0.018, 0.02, 0.011, and 0.001 for precision@5, precision@10, and precision@15, respectively. The precise values for precision@5, precision@10, precision@15, and precision@20 achieved by Dai et al. (2019) were 0.028, 0.026, 0.017, and 0.001, respectively. Lu et al. (2019) achieved precision values of 0.045, 0.038, 0.028, and 0.02, respectively, for precision@5, precision@10, precision@15, and precision@20. Concurrently, the suggested POI recommendation models yielded values of 0.85 for precision@5, 0.63 for precision@10, 0.478 for precision@15, and 0.384 for precision@20. Therefore, in terms of accuracy@5, accuracy@10, accuracy@15, and accuracy@20, the suggested model surpassed other state-of-the-art.

The graph of the precision@k with the k of the POI Recommendation Model is displayed in Fig. 7. It is observed that the values of precision@k decrease as k increases, and the proposed model has the best precision@k values as compared to other state-of-the-art (Fig. 8). Fig. 9 illustrates the bar graph comparing the results of recall@5, recall@10, recall@15, and recall@20 between the research presented by Zhou et al. (2019), Wang et al. (2019), Zhang et al. (2019), Yu et al. (2019), Dai et al. (2019), and Lu et al. (2019), and the model that proposed.

Fig. 7

Performance of proposed POI recommendation model for precision@k versus top k POI recommendation. POI, point of interest.

Fig. 8

Comparison of results of precision@k for top k point of interest recommendation.

Fig. 9

Comparison of results of recall@k for top k point of interest recommendations.

Fig. 10 presents recommended top k POIs with coordinates for a given user using the web application developed for the proposed model.

Fig. 10

Recommended top POIs with coordinates for a given user. POI, point of interest.