Pack: RCOMM2011 recommender systems workflow templates

Description:

This workflow takes user-item matrix A as a input. Then it calculates reduced SVD decomposition A_k by taking only k greatest singular values and corresponding singular vectors. This worfkflow calculates recommendations and predictions for particular user %{id} from matrix A. Particular row %{id} is taken from original matrix A and replaced with %{id} row in A_k matrix. Predictions are made for %{id} user based on another users A_k.

Note: This workflow uses R-script operator with R library 'Matrix'.

1. Install R-component in RapidMiner.

2. Install library 'Matrix' in R-console inside RapidMiner

Type :

> install.packages('Matrix')

inside R console and follow instructions.

Reference:

Bošnjak, M., Antulov-Fantulin, N. Šmuc, T. and Gamberger, D., Constructing recommender systems workflow templates in RapidMiner, Proceedings of the RapidMiner Community Meeting And Conference (RCOMM), 2011, Dublin, Ireland - In Press

Description:

Description:

Dataset description:

items

This is a concatenated train and test set from ECML/PKDD Discovery Challenge 2011. Only ID and name attributes were used, other attributes are discarded because of the size of the dataset. This example set represents the content information for each of the items represented by an ID.

user_history

This is an example set consisting of randomly sampled IDs from items dataset. It represents the user's history - all the items (in this case lectures) he has viewed.

user-item_user

This is a toy example of user-item matrix created for verification purposes of user-based CF approach.

user-item_item

This is a toy example of user-item matrix created for verification purposes of item-based CF approach.

INSTRUCTIONS: copy the contents of this file into your (file system!) RapidMiner repository, in the same directory where the recommender systems workflow templates are in order for workflows to see the data

Description:

The workflow for item-based collaborative filtering receives a user-item matrix for its input, and the same context defined macros as the user-based recommender template, namely %{id}, %{recommendation_no}, and %{number_of_neighbors}.

Although this process is in theory very similar to user-based technique, it differs in several processing steps since we are dealing with an item-user matrix, the transposed user-item example set. The first step of the workflow, after declaring zero values missing, is transposing the user-item matrix and calculating the item to %{number_of_neighbors} nearest items similarities. The resulting example set consists of IDs of items, IDs of their most similar items and similarity values. This is joined with IDs of items which our user of interest rated. As a result we obtain an example set of item IDs, their scores, and similarities to the items our user has rated. With the next join we filter only those items that our user had no contact with. The last step is to aggregate that data; group each item ID and calculate the average of that item's rating weighted by the similarity of that item to items our user of interest did rate. The operators following this aggregation, grouped in "Postprocessing" subprocess, only format the output and trim the number of examples in the example set to %{recommendation_no}.

Reference:

Bošnjak, M., Antulov-Fantulin, N. Šmuc, T. and Gamberger, D., Constructing recommender systems workflow templates in RapidMiner, Proceedings of the RapidMiner Community Meeting And Conference (RCOMM), 2011, Dublin, Ireland - In Press

Description:

The workflow for user-based collaborative filtering, takes only one example set as an input: a user-item matrix, where the attributes denote item IDs, and rows denote users. If a user i has rated an item j with a score s, the matrix will have the value s written in i-th row and j-th column. In the context of the process we define the ID of the user %{id}, desired number of recommendations %{recommendation_no}, and the number of neighbors used in calculating user to user distance %{number_of_neighbors}.

The first step of the workflow is to set all the zero values in the matrix to missing values in order to ignore them later in the aggregation process. Afterwards, we separate our user of interest from all the users in the example set and then calculate the similarities between our user and the %{number_of_neighbors} nearest users. We join the resulting similarities as weights with the example set of all users without our user of interest. As a result of this join we obtain a list of the most similar %{number_of_neighbors} users and their ratings with a weight that equals their similarity to our user of interest. Afterwards, we aggregate the ratings using these  similarities as weights and select only the items our user did not rate and for which we could calculate the rating. The final result is a list of items and their scores trimmed to %{recommendation_no} recommendations.

Reference:

Bošnjak, M., Antulov-Fantulin, N. Šmuc, T. and Gamberger, D., Constructing recommender systems workflow templates in RapidMiner, Proceedings of the RapidMiner Community Meeting And Conference (RCOMM), 2011, Dublin, Ireland - In Press

Description:

As an input, this workflow takes two distinct example sets: a complete set of items with IDs and appropriate textual attributes (item example set) and a set of IDs of items our user had interaction with (user example set). Also, a macro %{recommendation_no} is defined in the process context, as a required number of outputted recommendations.

The first steps of the workflow are to preprocess those example sets; select only textual attributes of item example set, and set ID roles on both of the example sets to the corresponding ID attributes. After the preprocessing, we exclude from the item example set all the items user had interaction with (i.e. we assume the user will not wish to see them recommended), and we analyze the rest using text mining analysis operators. The text is analyzed by first transforming it to the lower case, then tokenizing it by splitting on all non-letter characters, and then filtering the resulting tokens with respect to token length; filter out tokens shorter than 3 and longer than 25 letters. The last steps are stopword filtering (using English stopwords) and stemming using Porter stemmer algorithm. In the end, tokens are discarded if they show up in too few (less than 1%) or too many (more than 30%) documents. The result of the analysis is an example set of TF-IDF vectors of the items, and a bag of words obtained through this analysis.

This bag of words is then used for the textual analysis of the combined text from all the descriptions on the items the user had interaction with. The result of this analysis is also an example set of TF-IDF vectors, this time of our user's items, which is then used to calculate similarities to all the other items in the database. The final step, executed inside "Postprocessing" subprocess is to pick out the items that yield the highest similarity and format it into a new example set consisting of a recommendation and its similarity score.

Reference:

Bošnjak, M., Antulov-Fantulin, N. Šmuc, T. and Gamberger, D., Constructing recommender systems workflow templates in RapidMiner, Proceedings of the RapidMiner Community Meeting And Conference (RCOMM), 2011, Dublin, Ireland - In Press