I am using Spark to build a recommendation system prototype. After going through some tutorials, I have been able to train a MatrixFactorizationModel from my data.
However, the model trained by Spark mllib is just a Serializable. How can I use this model to do recommendation for real users? I mean, how can I persist the model into some sort of database or update it if the user data has been incremented?
For example, the model trained by Mahout recommendation library can be stored into databases like Redis, then we can query for the recommended item list later. But how can we do similar stuff in Spark? Any suggestion?
First, the "model" you're referring to from Mahout is not a model, but a pre-computed list of recommendations. You could also do this with Spark, and compute in batch recommendations for users, and persist them anywhere you like. This has nothing to do with serializing a model. If you don't want to do real-time updates or scoring, you can stop there and just use Spark for batch just like you do Mahout.
But I agree that in a lot of cases you do want to ship the model somewhere else and serve it. As you can see, other models in Spark are Serializable, but not MatrixFactorizationModel. (Yes, even though it's marked as such, it won't serialize.) Likewise, there is a standard serialization for predictive models called PMML but it contains no vocabulary for a factored matrix model.
The reason is actually the same. Whereas many predictive models, like an SVM or logistic regression model, are just a small set of coefficients, a factored matrix model is huge, containing two matrices with potentially billions of elements. That is why I think PMML doesn't have any reasonable encoding for it.
Likewise, in Spark, that means the actual matrices are RDDs that can't be serialized directly. You can persist these RDDs to storage, re-read them elsewhere using Spark, and recreate a MatrixFactorizationModel by hand that way.
You can't serve or update the model using Spark though. For this you are really looking at writing some code to perform updates and calculate recommendations on the fly.
I don't mind suggesting here the Oryx project, since its point is to manage exactly this aspect, particularly for ALS recommendation. In fact, the Oryx 2 project is based on Spark and although in alpha, already contains the complete pipeline to serialize and serve the output of MatrixFactorizationModel. I don't know if it meets your needs, but may at least be an interesting reference point.
Another method for creating recs with Spark is the search engine method. This is basically a cooccurrence recommender served by Solr or Elasticsearch. Comparing factorized to cooccurrence is beyond this question so I'll just describe the latter.
You feed interactions (user-id,item-id) into Mahout's spark-itemsimilarity. This produces a list of similar items for every item seen in the interaction data. It will come out by default as a csv and so can be stored anywhere. But it needs to be indexed by a search engine.
In any case when you want to fetch recs you use the user's history as the query, you get back an ordered list of items as recs.
One benefit of this method is that indicators can be calculated for as many user actions as you want. Any action the user takes that correlates to what you want to recommend can be used. For instance if you want to recommend a purchase but you record product-views as well. If you treated product-views the same as purchases you would likely get worse recs (I've tried it). However if you calculate an indicator for purchases and another (actually cross-cooccurrence) indicator for product-views they are equally predictive of purchases. This has the effect of increasing the data used for recs. The same type of thing can be done with user locations to blend in location information into purchase recs.
You can also bias your recs based on context. If you are in the "electronics" section of a catalog, you may want recs to be skewed towards electronics. Add electronics to the query against the item's "category" metadata field and give it a boost in the query and you have biased recs.
Since all of the biasing and mixing of indicators happens in the query it makes the recs engine easily tuned to multiple contexts while maintaining only one multi-field query made through a search engine. We get scalability from Solr or Elasticsearch.
One other benefit of either factorization or the search method is that entirely new users and new history can be used to create recs where the older Mahout recommenders could only recommend to users and interactions known when the job was run.
Descriptions here:
Mahout docs
Slides
Mahout on Spark: What’s New in Recommenders, part 1
Mahout on Spark: What’s New in Recommenders, part 2
Practical Machine Learning ebook
You should run model.predictAll() on a reduced RDD set of (user,product) pairs like in the Mahout Hadoop Job and store the results for online usage...
https://github.com/apache/mahout/blob/master/mrlegacy/src/main/java/org/apache/mahout/cf/taste/hadoop/item/RecommenderJob.java
You can use the function .save(sparkContext, outputFolder) to save the model to a folder of your choice. While giving the recommendations in realtime, you just have to use MatrixFactorizationModel.load(sparkContext, modelFolder) function to load it as a MatrixFactorizationModel object.
A question to #Sean Owen: Doesn't the MatrixFactorizationObject contain the Factorization matrices: user-feature and item-feature matrices instead of recommendations/predicted ratings.
Related
I am struggling with the implementation of a performant version of a SOM Batch algorithm on Spark / Pyspark for a huge dataset with > 100 features.
I have the feeling that I can either use RDDs where I can/have to specifiy the parallization on my own or I use Dataframe which should be more performant but I see no way how to use something like a local accumulation variable for each worker when using dataframes.
Ideas:
Using Accumulators. Parallelize the calculations by creating a UDF which takes the observations as input, calculates the impacts on the net and sends the impacts to an accumulator in the driver. (Implemented this version already, but seems rather slow (I think accumulator updates take to long))
Store results in a new column of Dataframe and then sum it together in the end. (Would have to store a whole neural net in the each row (e.g. 20*20*130) tho) Are spark optimization algorithms realizing, that it does not need to save each net but only sum them together?
Create an custom parallized algorithms using RDDs similar to that: https://machinelearningnepal.com/2018/01/22/apache-spark-implementation-of-som-batch-algorithm/ (but with more performant calculation algorithms). But I would have to use some kind of loop to loop over each row and update the net -> sounds like that would be rather unperformant.)
Any thoughts on the different options? Is there an even better option?
Or are all ideas not that good and I should just preselect a maximum variety subset of my dataset and train a SOM locally on that.
Thanks!
This is exactly what I have done last year, so I might be in a good position to give you an answer.
First, here is my Spark implementation of the batch SOM algorithm (it is written in Scala, but most things will be similar in Pyspark).
I needed this algorithm for a project, and every implementation I found had at least one of these two problems or limitations:
they did not really implement the batch SOM algorithm, but used a map averaging method that gave me strange results (abnormal symmetries in the output map)
they did not use the DataFrame API (pure RDD API) and were not in the Spark ML/MLlib spirit, i.e. with a simple fit()/transform() API operating over DataFrames.
So, there I went on to code it myself: the batch SOM algorithm in Spark ML style. The first thing I did was looking how k-means was implemented in Spark ML, because as you know, the batch SOM is very similar to the k-means algorithm. Actually, I could re-use a large portion of the Spark ML k-means code, but I had to modify the core algorithm and the hyperparameters.
I can summarize quickly how the model is built:
A SOMParams class, containing the SOM hyperparameters (size, training parameters, etc.)
A SOM class, which inherits from spark's Estimator, and contains the training algorithm. In particular, it contains a fit() method that operates on an input DataFrame, where features are stored as a spark.ml.linalg.Vector in a single column. fit() will then select this column and unpack the DataFrame to obtain the unerlying RDD[Vector] of features, and call the run() method on it. This is where all the computations happen, and as you guessed, it uses RDDs, accumulators and broadcast variables. Finally, the fit() method returns a SOMModel object.
SOMModel is a trained SOM model, and inherits from spark's Transformer/Model. It contains the map prototypes (center vectors), and contains a transform() method that can operate on DataFrames by taking an input feature column, and adding a new column with the predictions (projection on the map). This is done by a prediction UDF.
There is also SOMTrainingSummary that collects stuff such as the objective function.
Here are the take-aways:
There is not really an opposition between RDD and DataFrames (or rather Datasets, but the difference between those two is of no real importance here). They are just used in different contexts. In fact, a DataFrame can be seen as a RDD specialized for manipulating structured data organized in columns (such as relational tables), allowing SQL-like operations and an optimization of the execution plan (Catalyst optimizer).
For structured data, select/filter/aggregation operations, DO USE Dataframes, always.
...but for more complex tasks such as a machine learning algorithm, you NEED to come back to the RDD API and distribute your computations yourself, using map/mapPartitions/foreach/reduce/reduceByKey/and so son. Look at how things are done in MLlib: it's only a nice wrapper around RDD manipulations!
Hope it will solve your question. Concerning performance, as you asked for an efficient implementation, I did not make any benchmarks yet but I use it at work and it crunches 500k/1M-rows datasets in a couple of minutes on the production cluster.
I have built a Machine Learning model based on clustering, & now just want to update it with new data periodically (on daily basis). I am using PySpark Mlib, & not able to find any method in Spark for this need.
Note, required method 'partial_fit' available in scikit-learn, but not in Spark.
I am not in favor of appending new data & then re-built the model on daily basis, as it will increase data size & will be computationally expensive.
Please suggest me an effective way for model update or on-line learning using Spark Mllib?
You cannot update arbitrary models.
On a few select models this works. On some it works if you accept some loss in a accuracy. But on other models, the only way is to rebuild it completely.
For example support vector machines. The model only stores the support vectors. When updating, you would also need all the non-support vectors in order to find the optimal model.
That is why it is fairly common to build new models every night, for example.
Streaming is quite overrated. In particular k-means. Total nonsense to do online k-meand with "big" (lol) data. Because the new point have next to zero effect, you may just as well do a batch every night. These are just academic toys with no relevancy.
In one hand I want to use spark capability to compute TF-IDF for a collection of documents, on the other hand, the typical definition of TF-IDF (that Spark implementation is based on that) is not fit in my case. I want the TF to be term frequency among all documents, but in the typical TF-IDF, it's for each pair of (word, document). The IDF definition is the same as the typical definition.
I implemented my customized TF-IDF using Spark RDDs, but I was wondering if there any way to customize the source of the Spark TF-IDF so that I can use the capability of that, like Hashing.
Actually, I need something like :
public static class newHashingTF implements Something<String>
Thanks
It is pretty simple to implement different hashing strategies, as you can see by the simplicity of HashingTF:
(modern) Dataset version
(old) RDD version
This talk and its slides can help and there are many others online.
I have a predictive model (Logistic Regression) built in Spark 1.6 that has been saved to disk for later reuse with new data records. I want to invoke it with multiple clients with each client passing in single data record. It seems that using a Spark job to run single records through would have way too much overhead and would not be very scalable (each invocation will only pass in a single set of 18 values). The MLlib API to load a saved model requires the Spark Context though so am looking for suggestions of how to do this in a scalable way. Spark Streaming with Kafka input comes to mind (each client request would be written to a Kafka topic). Any thoughts on this idea or alternative suggestions ?
Non-distributed (in practice it is majority) models from o.a.s.mllib don't require an active SparkContext for single item predictions. If you check API docs you'll see that LogisticRegressionModel provides predict method with signature Vector => Double. It means you can serialize model using standard Java tools, read it later and perform prediction on local o.a.s.mllib.Vector object.
Spark also provides a limited PMML support (not for logistic regression) so you share your models with any other library which supports this format.
Finally non-distributed models are usually not so complex. For linear models all you need is intercept, coefficients and some basic math functions and linear algebra library (if you want a decent performance).
o.a.s.ml models are slightly harder to handle but there are some external tools which try to address that. You can check related discussion on the developers list, (Deploying ML Pipeline Model) for details.
For distributed models there is really no good workaround. You'll have to start a full job on distributed dataset one way or another.
I'm trying to find out if it is possible to have "incremental training" on data using MLlib in Apache Spark.
My platform is Prediction IO, and it's basically a wrapper for Spark (MLlib), HBase, ElasticSearch and some other Restful parts.
In my app data "events" are inserted in real-time, but to get updated prediction results I need to "pio train" and "pio deploy". This takes some time and the server goes offline during the redeploy.
I'm trying to figure out if I can do incremental training during the "predict" phase, but cannot find an answer.
I imagine you are using spark MLlib's ALS model which is performing matrix factorization. The result of the model are two matrices a user-features matrix and an item-features matrix.
Assuming we are going to receive a stream of data with ratings or transactions for the case of implicit, a real (100%) online update of this model will be to update both matrices for each new rating information coming by triggering a full retrain of the ALS model on the entire data again + the new rating. In this scenario one is limited by the fact that running the entire ALS model is computationally expensive and the incoming stream of data could be frequent, so it would trigger a full retrain too often.
So, knowing this we can look for alternatives, a single rating should not change the matrices much plus we have optimization approaches which are incremental, for example SGD. There is an interesting (still experimental) library written for the case of Explicit Ratings which does incremental updates for each batch of a DStream:
https://github.com/brkyvz/streaming-matrix-factorization
The idea of using an incremental approach such as SGD follows the idea of as far as one moves towards the gradient (minimization problem) one guarantees that is moving towards a minimum of the error function. So even if we do an update to the single new rating, only to the user feature matrix for this specific user, and only the item-feature matrix for this specific item rated, and the update is towards the gradient, we guarantee that we move towards the minimum, of course as an approximation, but still towards the minimum.
The other problem comes from spark itself, and the distributed system, ideally the updates should be done sequentially, for each new incoming rating, but spark treats the incoming stream as a batch, which is distributed as an RDD, so the operations done for updating would be done for the entire batch with no guarantee of sequentiality.
In more details if you are using Prediction.IO for example, you could do an off line training which uses the regular train and deploy functions built in, but if you want to have the online updates you will have to access both matrices for each batch of the stream, and run updates using SGD, then ask for the new model to be deployed, this functionality of course is not in Prediction.IO you would have to build it on your own.
Interesting notes for SGD updates:
http://stanford.edu/~rezab/classes/cme323/S15/notes/lec14.pdf
For updating Your model near-online (I write near, because face it, the true online update is impossible) by using fold-in technique, e.g.:
Online-Updating Regularized Kernel Matrix Factorization Models for Large-Scale Recommender Systems.
Ou You can look at code of:
MyMediaLite
Oryx - framework build with Lambda Architecture paradigm. And it should have updates with fold-in of new users/items.
It's the part of my answer for similar question where both problems: near-online training and handling new users/items were mixed.