Sifarish is a suite of solutions for recommendation personalization implementaed on Hadoop and Storm. Various algorithms, including feature similarity based recommendation and collaborative filtering based recommendation using social rating data are available
- Providing complete business solutions, not just bunch of machine learning algorithms
- Simple to use
- Input output in CSV format
- Metadata defined in simple JSON file
- Extremely configurable with tons of configuration knobs
Please read ../resource/GentleIntroductionToSifarish.docx for a high level introduction and overview. The various tutorial documents in the resource directory are useful for running different example use cases.
The following blogs of mine are good source of details of sifarish. These are the only source of detail documentation
- http://pkghosh.wordpress.com/2011/10/26/similarity-based-recommendation-basics/
- http://pkghosh.wordpress.com/2011/11/28/similarity-based-recommendation-hadoop-way/
- http://pkghosh.wordpress.com/2011/12/15/similarity-based-recommendation-text-analytic/
- http://pkghosh.wordpress.com/2012/04/21/socially-accepted-recommendation/
- http://pkghosh.wordpress.com/2010/10/19/recommendation-engine-powered-by-hadoop-part-1/
- http://pkghosh.wordpress.com/2010/10/31/recommendation-engine-powered-by-hadoop-part-2/
- http://pkghosh.wordpress.com/2012/12/31/get-social-with-pearson-correlation/
- http://pkghosh.wordpress.com/2012/09/03/from-item-correlation-to-rating-prediction/
- http://pkghosh.wordpress.com/2014/02/10/from-explicit-user-engagement-to-implicit-product-rating/
- http://pkghosh.wordpress.com/2014/04/14/making-recommendations-in-real-time/
- http://pkghosh.wordpress.com/2014/05/26/popularity-shaken/
- http://pkghosh.wordpress.com/2014/06/23/novelty-in-personalization/
- http://pkghosh.wordpress.com/2014/09/10/realtime-trending-analysis-with-approximate-algorithms/
- http://pkghosh.wordpress.com/2014/12/22/positive-feedback-driven-recommendation-rank-reordering/
- https://pkghosh.wordpress.com/2015/01/20/diversity-in-personalization-with-attribute-diffusion/
- https://pkghosh.wordpress.com/2015/03/22/customer-service-and-recommendation-system/
In the absence of social rating data, the only options is a feature similarity based recommendation. Similarity is calculated based on distance between entities in a multi dimensional feature space. Some examples are - recommending jobs based on user's resume - recommending products based on user profile. These solutions are known as content based recommendation, because it's based innate features of some entity.
There are two different solutions as follows
- Similarity between entities of different types (e.g. user profile and product)
- Similarity between entities of same type (e.g. product)
Attribute meta data is defined in a json file. Both entities need not have the same set of attributes. Mapping between attributes values from one entity to the other can be defined in the config file.
The data type supported are numerical (integer), categorical, text, geo location, time. The distance algorithms can be chosed to be euclidian, manhattan or minkowski. The default algorithm is euclidian.
The distancs between different atrributes of different types are combined to find distance between two entity instances. Different weights can be assigned to the attributes to control the relative importance of different attributes.
The tutorial ../resource/product_similarity_tutorial.txt is a good starting point. The relevant blogs are useful to understand the inner workings.
These solutions are based user behavior data with respect to some product or service. these algorithms are also known as collaborative filtering.
User behavior data is defined in terms of some explicit rating by user or it's derived from user behavior in the site. The essential input to all these algorithms is a matrix of user and items. The value for a cell could be the ratingas an integer. It could also be boolean, if the user's interest in an item is expressed as a boolean
The tutorial ../resource/tutorial.txt is a good starting point. The relevant blogs are useful to understand the inner workings.
These solutions are used when enough social data is not avaialable.
- If data contains text attributes, use TextAnalyzer MR to convert text to token stream using lucene
- Find similar items based on user profile. Use DiffTypeSimilarity MR
- Use TopMatches MR to find top n matches for a profile
When limited amount of user behavior data is available, these solutuions are appropriate
- If data contains text attributes, use TextAnalyzer MR to convert text to token stream using lucene
- Find similar items by pairing items with one another using SameTypeSimilarity MR
- Use TopMatches MR to find top n matches for a product
When significant of user behavior data is available, these soltions can be used. In the order of complexity, the choices are as follows. They are all based on social data
There two phases for collaborative filetering based recommendation using social data
- Find correlation between items 2. Predict rating based on items alreadyv rated and result of 1
The process involved running multiple map reduce jobs. Some of them are optional. Please refer to the tutorial document tutorial.txt in the resource directory
Recommendations can be made real time based on user's current behavior in a pre defined time window. The solution is based on Storm, although Hadoop gets used to compute item correlation matrix from historical user behavior data.
For content based recommendation being able to find match between text field is an important factor. Text attributes are stemmed or normalized with Apache Lucene. Various languages, in addition to default of english are supported. They are german, french, italian, spanish, polish and brazilian portuguese. Text matching algorithms supported are cosine, jaccard and semantic. For semantic matching, RDF semantic graph is used
For content based recommendation, There is support for structured fields e.g., Location, Time Window, Event, Categorized Item etc. Many of these provide contextual dimensions to recommendation. They are particularly relevant for recommendation in the mobile space
Novelty for an item can be computed at individual user level or the whole user community as a whole. Novelty is blended into the final recommendation list by taking weighted average of predicted rating and novelty
Based on recent work in the academic world, I am working on implementing some algorithms to introduce
diversity in recommendation. Unlike novelty, diversity is group wise property. Diversity can be
defined either in terms item dissimilarity in a collaborative filtering sense or structural and content
sense
For content based recommendation, faceted match is supported as faceted search in Solr. Faceted fields are specified through a configuration parameter
Dithering effectively handles the problem users usually not browsing the first few items in a list. The dithering process shuffles the list little bit, every time recommended items are presented to the user.
Please use the tutorial.txt file in the resource directory for batch mode recommendation processing. For real time recommendation please use the tutorial document there is a separate tutorial document realtime_recommendation_tutorial.txt
If you use Apache mahout or some thing else for recommendation, you can bring your basic recommendation output (userID, itemID, predictedRating) to sifarish for additional postprocessing to improve the quality of the output. They are listed in the next section.
Just accuracy from the CF algorithm is not enough for a good recommender. There are various post processing plugins are essential. They improve the quality of results. Here is the list. Sifarsh supports most them. Some are under development.
- Business goal injection
- Adding novelty
- Adding diversity
- Rank reordering for explicit positive feedback
- Rank reordering for implicit negative feedback
- Dithering
Please refer to the wiki page for a detailed list of all configuration parameters https://github.com/pranab/sifarish/wiki/Configuration. Going through the tutorial documents in the resource directory, you can find sample configuration for various use cases.
Please read jar_dependency.txt in the resource directory for build and run time dependency
For Hadoop 1
- mvn clean install
For Hadoop 2 (non yarn), use the branch nuovo
- git checkout nuovo
- mvn clean install
For Hadoop 2 (yarn), use the branch nuovo
- git checkout nuovo
- mvn clean install -P yarn
Please feel free to email me at [email protected]
Contributors are welcome. Please email me at [email protected]