Academia.eduAcademia.edu

Figure 11 – uploaded by Giovanni Semeraro

See full PDFdownloadDownload figure
Table 7: Comparison with state of the art collabo- rative approaches on Movielens.

Table 7 Comparison with state of the art collabo- rative approaches on Movielens.

Related Figures (30)

Figure 1: Recommender Dashboard
Figure 1: Recommender Dashboard
Figure 2: Analysis of the user’s click behavior
Figure 2: Analysis of the user’s click behavior
Figure 3: Daily clicks on remove (figures in per cent)
Figure 3: Daily clicks on remove (figures in per cent)
Figure 2: Prediction accuracy for SVM and Ran- dom forest in 4 time frame scenarios (2, 4, 8 and 12 months) using different article content analysis techniques
Figure 2: Prediction accuracy for SVM and Ran- dom forest in 4 time frame scenarios (2, 4, 8 and 12 months) using different article content analysis techniques
Figure 3: Prediction accuracy based on the three content analysis techniques for the 4 time frame sce- narios
Figure 3: Prediction accuracy based on the three content analysis techniques for the 4 time frame sce- narios
Table 2: Performance comparison of various recommendation techniques
Table 2: Performance comparison of various recommendation techniques
Table 2: RMSE of predicted ratings for our method vs. standard MF across three applications. Table 1: Examples of words pertaining to aspects in the restaurant application.
Table 2: RMSE of predicted ratings for our method vs. standard MF across three applications. Table 1: Examples of words pertaining to aspects in the restaurant application.
former contains 1 million 1-5 stars ratings from 6,040 users on 3,883 movies. The LibraryThing dataset contains more than 2 millions ratings from 7,564 users on 39,515 books. As there are many duplicated ratings in the dataset, when a user has rated more than once the same item, we select her last rating. This choice brings to have 626,000 rat- ings in the range from 1 to 10. The user-item interactions contained in the datasets are enriched with side informa- tion thanks to the item mapping and linking to DBpedia technique detailed in [27], whose dump is available at http: //sisinflab.poliba.it/semanticweb/lod/recsys/datasets/. In the attempt to reduce the popularity bias from our final evaluation we decided to remove the top 1% most popular items from both datasets [5]. Moreover we keep out, from LibraryThing, users with less than five ratings and items rated less than five times, and to have a dataset character- ized by lower sparsity we retain for Movielens only users with at least fifty ratings, as already done in [10]. Table 1 contains the final statistics for our datasets. Table 1: Statistics about the two datasets
former contains 1 million 1-5 stars ratings from 6,040 users on 3,883 movies. The LibraryThing dataset contains more than 2 millions ratings from 7,564 users on 39,515 books. As there are many duplicated ratings in the dataset, when a user has rated more than once the same item, we select her last rating. This choice brings to have 626,000 rat- ings in the range from 1 to 10. The user-item interactions contained in the datasets are enriched with side informa- tion thanks to the item mapping and linking to DBpedia technique detailed in [27], whose dump is available at http: //sisinflab.poliba.it/semanticweb/lod/recsys/datasets/. In the attempt to reduce the popularity bias from our final evaluation we decided to remove the top 1% most popular items from both datasets [5]. Moreover we keep out, from LibraryThing, users with less than five ratings and items rated less than five times, and to have a dataset character- ized by lower sparsity we retain for Movielens only users with at least fifty ratings, as already done in [10]. Table 1 contains the final statistics for our datasets. Table 1: Statistics about the two datasets
Table 2: Results of the ItemKNN approach on Movielens dataset. P and R stand respectively for precision and recall, SG indicates the Skip-Gram model, and DB and WD represent DBpedia and Wikidata respectively.
Table 2: Results of the ItemKNN approach on Movielens dataset. P and R stand respectively for precision and recall, SG indicates the Skip-Gram model, and DB and WD represent DBpedia and Wikidata respectively.
Table 4: Examples of K-nearest-neighbor sets on Movielens, for the Skip-Gram model with depth of 4 and size vectors 200, on DBpedia.
Table 4: Examples of K-nearest-neighbor sets on Movielens, for the Skip-Gram model with depth of 4 and size vectors 200, on DBpedia.
Figure 1: Two-dimensional PCA projection of the 200-dimensional Skip-gram vectors of movies in Ta- ble 4.
Figure 1: Two-dimensional PCA projection of the 200-dimensional Skip-gram vectors of movies in Ta- ble 4.
Figure 1: Prediction accuracy and novelty of the algorithms evaluated
Figure 1: Prediction accuracy and novelty of the algorithms evaluated
Figure 2: The graph view of V for Vendetta detailed view, which contains three tabs: the first contain: brief textual description, the second presents a graph view f the resource in order to show the main properties and the hird summarizes the main information in a tabular form ‘he graph view is illustrated in Figure 2. The graph i: aginated and few properties per page are presented in or. er to avoid information overload, since the resource car ave a very high number of properties. This view can be seful also to explain the recommendation: for instance the ser can understand that the recommended resource has the ame director or the same main actor as the initial movie ‘he graph view is based on DBpedia Mobile Explorer [25], ¢ inked Data visualization framework for the mobile environ. aent, which enables the application to hide the underlying omplexity of the Linked Data to the users by processing he resources to be presented received from DBpedia.
Figure 2: The graph view of V for Vendetta detailed view, which contains three tabs: the first contain: brief textual description, the second presents a graph view f the resource in order to show the main properties and the hird summarizes the main information in a tabular form ‘he graph view is illustrated in Figure 2. The graph i: aginated and few properties per page are presented in or. er to avoid information overload, since the resource car ave a very high number of properties. This view can be seful also to explain the recommendation: for instance the ser can understand that the recommended resource has the ame director or the same main actor as the initial movie ‘he graph view is based on DBpedia Mobile Explorer [25], ¢ inked Data visualization framework for the mobile environ. aent, which enables the application to hide the underlying omplexity of the Linked Data to the users by processing he resources to be presented received from DBpedia.
Google Play, but the Android Package (APK) of the first version is already available on the Web"*. Figure 3: The interactions between the main modules of the application
Google Play, but the Android Package (APK) of the first version is already available on the Web"*. Figure 3: The interactions between the main modules of the application
Figure 1 An example of quote usage in Twitter thread
Figure 1 An example of quote usage in Twitter thread
Table 1 number of contexts for each quote in datasets
Table 1 number of contexts for each quote in datasets
(* indicates that each of our algorithms outperform the best baseline algorithm with statistically significant increase at p < 0.01 in two-tailed t-tests) Table 2 Results of Recall@5 of different methods.
(* indicates that each of our algorithms outperform the best baseline algorithm with statistically significant increase at p < 0.01 in two-tailed t-tests) Table 2 Results of Recall@5 of different methods.
Table 3 number of correct cases of single algorithms in Twitter dataset
Table 3 number of correct cases of single algorithms in Twitter dataset
An overview of the proposed approach is depicted in Fig- ure 1. For the experimental setup only, we split user’s r interest set Pj,, into training and test sets Plain, Plest re- spectively. However, this step is dropped out in the non- experimental recommendation scenario and the ‘first step re- ceives, in this case, the complete interest set P/,,,.
An overview of the proposed approach is depicted in Fig- ure 1. For the experimental setup only, we split user’s r interest set Pj,, into training and test sets Plain, Plest re- spectively. However, this step is dropped out in the non- experimental recommendation scenario and the ‘first step re- ceives, in this case, the complete interest set P/,,,.
Figure 2: WBV and PBV compared with Logistic regression and Support Vector Machine
Figure 2: WBV and PBV compared with Logistic regression and Support Vector Machine
Figure 1: Purchase behavior prediction network.
Figure 1: Purchase behavior prediction network.
The 10-fold cross-validation results averaged over severa runs (due to random initialization) on the expert-labelec training data are shown in Table 5. As the baseline, we trained a FF model with sums of all tokens across all tweets for each user (*-sum) as the input. We observe better per- formance when tweet information is input sequentially to model with memory (*-seq) than when the sequence infor. mation is lost by summing (*-sum) tweets. That is, it is im- portant to capture information embedded in the sequence o: tweets. As expected, the LSTM consistently outperformed the RNN because it has the ability to retain information ove longer time spans. We also observe that the addition of pre- dicted relevance probabilities to the RNN model (*+Rel) im- proved performance over the simple RNN; however, adding tweets’ relevance only improved the LSTM’s performance for the harder combined product task, which may indicate that the vanilla LSTM is powerful enough to learn which tweets are relevant or not when trained on a single product. Table 5: Purchase prediction cross-validation.
The 10-fold cross-validation results averaged over severa runs (due to random initialization) on the expert-labelec training data are shown in Table 5. As the baseline, we trained a FF model with sums of all tokens across all tweets for each user (*-sum) as the input. We observe better per- formance when tweet information is input sequentially to model with memory (*-seq) than when the sequence infor. mation is lost by summing (*-sum) tweets. That is, it is im- portant to capture information embedded in the sequence o: tweets. As expected, the LSTM consistently outperformed the RNN because it has the ability to retain information ove longer time spans. We also observe that the addition of pre- dicted relevance probabilities to the RNN model (*+Rel) im- proved performance over the simple RNN; however, adding tweets’ relevance only improved the LSTM’s performance for the harder combined product task, which may indicate that the vanilla LSTM is powerful enough to learn which tweets are relevant or not when trained on a single product. Table 5: Purchase prediction cross-validation.
Figure 2: Histogram of the number of days between the first “want” and “bought” tweets by “buy” users.
Figure 2: Histogram of the number of days between the first “want” and “bought” tweets by “buy” users.
Related topics:
Computer ScienceSemanticsRecommender SystemsUser Generated ContentRecommender SystemImplicit Feedback
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved