In addition to contexts and reviews, this thesis also studies how to improve prediction accuracy with multi-criteria assessments. The experiments showed that the proposed method outperforms both single and multi-criteria rating conversion techniques with higher accuracy and prediction coverage.
Motivation
On the other hand, a review-based recommendation approach has been proposed with the primary goal of alleviating rating sparsity issues encountered in standard RS by using user-generated reviews. Moreover, since contexts are information that influence users' decisions about topics, they can be useful in constructing high-quality user and topic representations for review-based recommendations.
Problem Definition
Extracting contexts from reviews
Constructing user and item representations
Utilizing multi-criteria ratings
Objective
Contribution
Unsupervised Context Extraction via Region Embedding for Context
Context-Aware User and Item Representations Based on Unsuper-
Unlike most deep learning-based methods that learn a single static user or item representation for all views, CARE-AI dynamically creates unique user and item representations for each individual view that are better suited to capture the specific contextual information embedded into this review. Experiments demonstrate that CARE-AI outperforms state-of-the-art grade prediction methods, including review-based and context-aware recommendation techniques.
Multi-Criteria Rating Conversion Without Relation Loss For Recom-
In addition, the effectiveness of using the proposed interaction and attention modules, the impact of the model parameters, the impact of the rating quality, as well as the performance on situations where ratings are scarce are analyzed in detail.
Thesis Organization
Recommendation Strategies
The memory-based CF approach relies on the ratings of users who have similar interests to predict the rating. Therefore, the scores of useru3on elements that have not yet been scored by u1 (eg, v3) can be useful for predicting the score for u1.
Challenges in Recommender Systems
The accuracy of a memory-based CF approach relies heavily on the quality of the neighbors that have evaluated the target items. However, the prediction made by the model-based CF approach may not be as accurate as the memory-based CF approach when the user item evaluation matrix is dense (when there are sufficient amounts of evaluations to identify good-quality neighbors up).
Context-Aware Recommender Systems
- What is Context?
- Improving Accuracy with Contexts
- Incorporating Contexts for Making Recommendations
- Challenges for Contextual Recommendations
Brown [16] defines context as elements of the user's environment known to the user's computer. Most context-aware recommendation techniques adopt a representational approach, where contexts are defined by predefined sets of context variables and their corresponding static values, such as those presented in Table 2.1.
Review-Based Recommender Systems
- Alleviating Rating Sparsity with Reviews
- Leveraging Reviews for Making Recommendations
- Challenges for Review-Based Recommendations
- Extracting Contexts from User Reviews
In recent years, many deep learning techniques have been adopted to model user representations and articles from reviews due to their superior predictive performance. Additionally, user and item representations are constructed in a static manner by aggregating their respective previous revisions.
![Figure 2.7: A neural network framework for constructing user representations from reviews In an extension to the CNN, NARRE [19] applied an attention mechanism [96] to construct representations by considering different contributions from reviews based on t](https://thumb-ap.123doks.com/thumbv2/123deta/10550778.0/38.892.238.633.161.546/framework-constructing-representations-extension-attention-representations-considering-contributions.webp)
Multi-Criteria Recommender Systems and Rating Conversion
- Improving Accuracy with Multi-Criteria Ratings
- Similarity Aggregation
- Rating Conversion
- DeepCoNN
- NARRE
To measure the usability similarities, they extended the similarity methods with a single criterion for the multi-criteria scheme by applying the aggregation techniques. This method converts a user rating on an item into a probability that the item in question will be favored by the user [44]. This method learned different contributions from reviews to construct user and item representations based on the usage of reviews.
Context Extraction Techniques from User Reviews
Rich-Context
3.2.2) The weighted frequency wr(nk) is calculated as the ratio of the number of all context-rich scans containing nk compared to all context-rich scans. For each noun nk, calculate the ratio of its weighted frequencies in context-rich and context-free scans: ratio(nk) = wr(nk). For each topic k, calculate the ratio of its weighted frequencies on context-rich and context-free scans: ratio(tk) = wr(tk).
CARL
After the process of topic identification, the remaining topics are considered as the contextual topics that represent contexts extracted from reviews. The different contributions of the weighted contextual feature vectors can be considered as the different influences of contexts on the user's individual preferences and item-specific features, which consequently affect the corresponding rating of each review.
Rating Conversion Techniques
Linear Mapping
Letrub,max and rub,min indicate the maximum and minimum rating given by the target user. Although the linear mapping approach is able to solve some normalization problems, some still occur. One is that if there are users who have different rating patterns but have the same highest and lowest rating, the prediction results of the same neighbors are still the same as what happened in the normalizations.
Lathia’s Rating Conversion function
Another issue is that if the target user has only rated items with one rating value, all ratings from all neighbors will be assigned that value.
Warat’s Rating Conversion Function
By incorporating contexts, many context-aware methods have been able to achieve improved prediction accuracy, compared to standard CF-based approach. In practice, most context-aware methods [61, 70] identify the relevance of contexts by applying statistical tests such as the paired test to each contextual variable. In this chapter, the context-aware region embedding (CARE), a new unsupervised method to define, extract and represent context from overview data, is proposed [91].
Model Overview
Identifying Candidate Context Words
Extracting Contextual Regions
We would definitely come back here!”. which is very convenient because the room is comfortable and clean, comfortable and clean and. and the staff never fails region size = 5 Candidate context words. Choose the combination that contributes the most variance to the rating distribution as context for region(cn,d)m. Store the rating distribution for this combination, dist(cn,d)m∈R|Rating|in the list of rating distributions Dist for indexm.
Learning the Region Embeddings
After obtaining all predicted word embeddings, the embedding of the region γγγcn,m∈Rh of the contextual region region(cn,d) is calculated. pwt+d, where max max is the pooling operation over all predicted word embeddings used to extract the most predictive features in the region [77]. This indicates that the meaning of region(cn,d)m is now semantically defined by the meaning of neighboring wordswt given the candidate context wordcn. For example, the two context regions "very clean room" and "not clean room" would give completely different region embeddings for samecn= "clean".
Experiment Settings
Therefore, in order to correctly analyze the actual influence of a word on the rating distribution, a data standardization technique is used, as expressed by. The frequency of reviews with the word "rude" is now distributed towards low review scores, which is appropriate for its negative meaning. After standardization, the variance of the rating distribution for each word was calculated, and the words with variances above minvar=1 were selected as a set of candidate context words for each data set.
Results and Discussion
- Context Analysis
- Influences of Candidate Context Words and Their Neighboring Words 53
- Contexts as Regions
- Applications for CARE
As discussed in Section 4.4, their neighboring words can change the impact of candidate context words on their rating distributions. For these reasons, we see that local context units can effectively capture the influences of candidate context words, together with their neighboring words, on the distribution of ratings. For a single model, I used word embeddings of all candidate context words learned by CARE for region size = 1.
Conclusion
The attention module then generates the user and item representations based on different levels of relevance between contexts in each review. Finally, the user and item representations are used to predict ratings using a latent factor model [50]. CARE-AI's user and item representations are dynamically constructed for each specific review to effectively capture the contextual information embedded in that review.
Model Overview
Workflow
Model Architecture
Therefore, those contextual regions containing words related to room price should be more relevant to these user preferences than those containing words related to room size. The same assumption can be applied to the importance of contextual regions for unique features of an object. I believe that the more relevant a context region is to individual user preferences or object functionality, the more it should contribute to the evaluation of a particular review compared to other regions in the same review.
Interaction Module
Attention Module
Prediction Layer
Experimental Evaluation
- Data Preparation
- Baselines
- Experimental Settings
- Experimental Results
For PMF and NMF, the number of latent dimensions was set to 15, the learning rate to 0.005, and the regularization parameter to 0.001. For all these models, the number of latent dimensions was set to 32 and the embedding size to 300. 74 from Reviews Table 5.3: The NDCG values for the compared methods on the Amazon Software dataset.
Discussion
- Predictive Performance
- Attention and Interaction Modules
- Parameter Sensitivity
- Impact of Review Quality
- Performance on Sparse Data
Furthermore, the impact of the review quality on the performance of context-aware methods is discussed. This subsection first studies the impact of model parameters on the predictive performance of CARE-AI. From Figure 5.13, all methods using review data for constructing user and item representations (DeepCoNN, NARRE, CARL and CARE-AI) produced more accurate predictions than those constructing such representations using only rating data (PMF and NMF).
Conclusion
In a multi-criteria (MC) recommendation approach, users are able to specify their preferences on each item in multiple aspects rather than just one single rating [6]. In this chapter, a new method is proposed to simultaneously convert the multi-criteria ratings from one user to another user's aspect. The proposed method maintains the implicit relationship between the multi-criteria ratings by simultaneously converting all criteria ratings from one user to another user's aspect.
Model Overview
Also, converting each criterion rating independently may cause a loss in the implicit relationship between criterion ratings. This method can convert multi-criteria evaluations simultaneously in the same space in order to preserve the implicit relationship between criteria evaluations. Experiments showed that the proposed method outperformed the single and multi-criteria estimation conversion techniques in terms of prediction accuracy and prediction coverage.
Variance Normalization
Finding Multi-Criteria Rating Patterns
Multi-Criteria Rating Conversion
The example of assessment conversion process is presented by Figure 6.2, which transforms the assessment of ua into pattern of ubon the same plane.
Rating Prediction
Experimental Evaluation
Dataset
In addition to the overall rating, users also give ratings according to four criteria, which are acting, story, direction and visuals. In addition to an overall rating, this dataset contains ratings based on six criteria: cleanliness, location, rooms, service, sleep quality, and value. A five-fold cross-validation was performed to evaluate the performance of the models on both datasets.
Evaluation Results
Discussion
- Multi-Criteria Collaborative Filtering
- Warat’s and Lathia’s Rating Conversion
- Multi-Criteria Warat’s Rating Conversion
- Applications of Multi-Criteria Recommendations
As shown in Tables 6.2 and 6.3, the MC-Warat method achieves more accurate prediction than the single criterion Warat's method. In terms of prediction coverage, the MC-Warat method is slightly lower than Warat's single-criteria method. For the accuracy, the proposed method is better than the MC-Warat method on the Yahoo Movie dataset.
Conclusion
InProceedings of Third ACM Conference on Recommender Systems, RecSys ’09, faqet 245–248, Nju Jork, NY, SHBA. Proceedings of Third ACM Conference on Recommender Systems, RecSys ’09, faqet 265–268, Nju Jork, NY, SHBA. Proceedings of the 9th ACM Conference on Recommender Systems, RecSys ’15, faqet 383–386, Nju Jork, NY, SHBA.
A user rating system from Amazon.com
A user-item rating matrix
A sparse user-item rating matrix
A user-item rating tensor for context Season
Examples of context-rating co-occurrences for contexts (a) Companion and
Example of a user-generated review and rating
A neural network framework for constructing user representations from reviews 22
The model architecture of DeepCoNN
The model architecture of NARRE
The architecture of review-based feature learning of CARL
A workflow architecture of CARE
Example of word-rating co-occurrences and their corresponding variances
Example of rating co-occurrences with two bigrams containing the word
Extracting the contextual regions from a review
Example process for identifying context words in a contextual region
Illustration of the proposed context extraction model for a contextual region
Rating distributions for example words: (a) before standardization, (b) after
Visualization of the local context units for some chosen candidate context
Projection of sampled region embeddings for the TripAdvisor and Amazon
Example of a review with extracted contexts highlighted to show their rating
An overview of the workflow of CARE-AI
Illustration of the CARE-AI model
Illustration of CARE-I model
Illustration of CARE-A model
Illustration of CARE-0 model
The NDCG values for variants of the prediction models of CARE
Impact of different numbers of candidate context words
Impact of different region sizes
Impact of different embedding sizes
Training times with respect to four model parameters
Predictive performances of CARE-AI on three datasets based on review quality. 88
An overview of the mechanism of the proposed method
The user pattern transformation
Examples of predefined contextual variables and their values
Example of the overall and multi-criteria ratings on hotel rating data
Statistics of review datasets from multiple recommendation domains
Examples of candidate context words extracted from the TripAdvisor and
Examples of candidate context words extracted from six categories within
Criteria for categorizing a rating distribution based on correlation score
Classification results for the single word-context and the contextual-region
Statistics for the three review datasets
Comparison of the characteristics of all methods
The NDCG values for the compared methods on the Amazon Software dataset. 74
The NDCG values for the compared methods on the Amazon Movies & TV
The HR and MRR values for the compared methods on the three review
Comparison of context-rich and context-free reviews
Statistical data of Yahoo and TripAdvisor datasets
Experimental results on Yahoo Movie Dataset
Experimental results on TripAdvisor Hotel Dataset
