A Spark version in plan? about shap

[nickthegeek2877]

@QuentinAmbard First of all, I want to thank you for this excellent job. I'm also interested in pyspark. The problem is what @evah mentioned: you can build an explainer based on pyspark, but you cannot feed your explainer with a pyspark DataFrame. Theoretically, one could define a UserDefinedFunction, collect the features from spark to get a numpy or a pandas object, run your explainer and let spark do the dispatching. Unfortunately, to do so, objects needed for the computation must be serializable (through pickle) in order to be dispatched to every node, and an explainer is not.

A possible solution is to expose a serialize and a unserialize method for an explainer. The serialize should return an object made of dictionary, lists, whatever "basic" python types. The unserialize of course takes the serialized object and build from it a proper explainer.

In this way things might work: the serialized object could be dispatched, every node would build its explainer and the calculation could be carried out over the entire DataFrame.

Is it feasible to have those methods?

Thank you very much

PS: mostly interested in TreeExplainer if that helps in some way.

[techwithshadab]

Even I'm interested in PySpark implementation of it. I tried calculating SHAP values using the UDF code below:

X_df = X_df.repartition(10)
X_columns = X_df.columns

def add_shap(rows):
  rows_pd = pd.DataFrame(rows, columns=X_columns)
  shap_values = explainer.shap_values(rows_pd.drop(["Respondent"], axis=1))
  return [Row(*([int(rows_pd["Respondent"][i])] + [float(f) for f in shap_values[i]])) for i in range(len(shap_values))]

shap_df = X_df.rdd.mapPartitions(add_shap).toDF(X_columns)

Code Reference- DataBricks
But this code is throwing below error:

Traceback (most recent call last):
  File "/usr/lib/spark/python/pyspark/serializers.py", line 597, in dumps
    return cloudpickle.dumps(obj, 2)
  File "/usr/lib/spark/python/pyspark/cloudpickle.py", line 863, in dumps
    cp.dump(obj)
  File "/usr/lib/spark/python/pyspark/cloudpickle.py", line 260, in dump
    return Pickler.dump(self, obj)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 409, in dump
    self.save(obj)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 476, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 736, in save_tuple
    save(element)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 476, in save
    f(self, obj) # Call unbound method with explicit self
  File "/usr/lib/spark/python/pyspark/cloudpickle.py", line 400, in save_function
    self.save_function_tuple(obj)
  File "/usr/lib/spark/python/pyspark/cloudpickle.py", line 549, in save_function_tuple
    save(state)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 476, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 821, in save_dict
    self._batch_setitems(obj.items())
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 847, in _batch_setitems
    save(v)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 476, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 821, in save_dict
    self._batch_setitems(obj.items())
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 852, in _batch_setitems
    save(v)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 521, in save
    self.save_reduce(obj=obj, *rv)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 634, in save_reduce
    save(state)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 476, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 821, in save_dict
    self._batch_setitems(obj.items())
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 847, in _batch_setitems
    save(v)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 521, in save
    self.save_reduce(obj=obj, *rv)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 634, in save_reduce
    save(state)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 476, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 821, in save_dict
    self._batch_setitems(obj.items())
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 847, in _batch_setitems
    save(v)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 521, in save
    self.save_reduce(obj=obj, *rv)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 634, in save_reduce
    save(state)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 476, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 821, in save_dict
    self._batch_setitems(obj.items())
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 847, in _batch_setitems
    save(v)
  File "/opt/conda/anaconda/lib/python3.6/pickle.py", line 496, in save
    rv = reduce(self.proto)
  File "/usr/lib/spark/python/lib/py4j-0.10.7-src.zip/py4j/java_gateway.py", line 1257, in __call__
    answer, self.gateway_client, self.target_id, self.name)
  File "/usr/lib/spark/python/pyspark/sql/utils.py", line 63, in deco
    return f(*a, **kw)
  File "/usr/lib/spark/python/lib/py4j-0.10.7-src.zip/py4j/protocol.py", line 332, in get_return_value
    format(target_id, ".", name, value))
py4j.protocol.Py4JError: An error occurred while calling o3332.__getstate__. Trace:
py4j.Py4JException: Method __getstate__([]) does not exist
	at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:318)
	at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:326)
	at py4j.Gateway.invoke(Gateway.java:274)
	at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:132)
	at py4j.commands.CallCommand.execute(CallCommand.java:79)
	at py4j.GatewayConnection.run(GatewayConnection.java:238)
	at java.lang.Thread.run(Thread.java:748)



---------------------------------------------------------------------------
Py4JError                                 Traceback (most recent call last)
/usr/lib/spark/python/pyspark/serializers.py in dumps(self, obj)
    596         try:
--> 597             return cloudpickle.dumps(obj, 2)
    598         except pickle.PickleError:

/usr/lib/spark/python/pyspark/cloudpickle.py in dumps(obj, protocol)
    862         cp = CloudPickler(file,protocol)
--> 863         cp.dump(obj)
    864         return file.getvalue()

/usr/lib/spark/python/pyspark/cloudpickle.py in dump(self, obj)
    259         try:
--> 260             return Pickler.dump(self, obj)
    261         except RuntimeError as e:

/opt/conda/anaconda/lib/python3.6/pickle.py in dump(self, obj)
    408             self.framer.start_framing()
--> 409         self.save(obj)
    410         self.write(STOP)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    475         if f is not None:
--> 476             f(self, obj) # Call unbound method with explicit self
    477             return

/opt/conda/anaconda/lib/python3.6/pickle.py in save_tuple(self, obj)
    735             for element in obj:
--> 736                 save(element)
    737             # Subtle.  Same as in the big comment below.

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    475         if f is not None:
--> 476             f(self, obj) # Call unbound method with explicit self
    477             return

/usr/lib/spark/python/pyspark/cloudpickle.py in save_function(self, obj, name)
    399                 or themodule is None):
--> 400             self.save_function_tuple(obj)
    401             return

/usr/lib/spark/python/pyspark/cloudpickle.py in save_function_tuple(self, func)
    548             state['qualname'] = func.__qualname__
--> 549         save(state)
    550         write(pickle.TUPLE)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    475         if f is not None:
--> 476             f(self, obj) # Call unbound method with explicit self
    477             return

/opt/conda/anaconda/lib/python3.6/pickle.py in save_dict(self, obj)
    820         self.memoize(obj)
--> 821         self._batch_setitems(obj.items())
    822 

/opt/conda/anaconda/lib/python3.6/pickle.py in _batch_setitems(self, items)
    846                     save(k)
--> 847                     save(v)
    848                 write(SETITEMS)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    475         if f is not None:
--> 476             f(self, obj) # Call unbound method with explicit self
    477             return

/opt/conda/anaconda/lib/python3.6/pickle.py in save_dict(self, obj)
    820         self.memoize(obj)
--> 821         self._batch_setitems(obj.items())
    822 

/opt/conda/anaconda/lib/python3.6/pickle.py in _batch_setitems(self, items)
    851                 save(k)
--> 852                 save(v)
    853                 write(SETITEM)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    520         # Save the reduce() output and finally memoize the object
--> 521         self.save_reduce(obj=obj, *rv)
    522 

/opt/conda/anaconda/lib/python3.6/pickle.py in save_reduce(self, func, args, state, listitems, dictitems, obj)
    633         if state is not None:
--> 634             save(state)
    635             write(BUILD)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    475         if f is not None:
--> 476             f(self, obj) # Call unbound method with explicit self
    477             return

/opt/conda/anaconda/lib/python3.6/pickle.py in save_dict(self, obj)
    820         self.memoize(obj)
--> 821         self._batch_setitems(obj.items())
    822 

/opt/conda/anaconda/lib/python3.6/pickle.py in _batch_setitems(self, items)
    846                     save(k)
--> 847                     save(v)
    848                 write(SETITEMS)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    520         # Save the reduce() output and finally memoize the object
--> 521         self.save_reduce(obj=obj, *rv)
    522 

/opt/conda/anaconda/lib/python3.6/pickle.py in save_reduce(self, func, args, state, listitems, dictitems, obj)
    633         if state is not None:
--> 634             save(state)
    635             write(BUILD)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    475         if f is not None:
--> 476             f(self, obj) # Call unbound method with explicit self
    477             return

/opt/conda/anaconda/lib/python3.6/pickle.py in save_dict(self, obj)
    820         self.memoize(obj)
--> 821         self._batch_setitems(obj.items())
    822 

/opt/conda/anaconda/lib/python3.6/pickle.py in _batch_setitems(self, items)
    846                     save(k)
--> 847                     save(v)
    848                 write(SETITEMS)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    520         # Save the reduce() output and finally memoize the object
--> 521         self.save_reduce(obj=obj, *rv)
    522 

/opt/conda/anaconda/lib/python3.6/pickle.py in save_reduce(self, func, args, state, listitems, dictitems, obj)
    633         if state is not None:
--> 634             save(state)
    635             write(BUILD)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    475         if f is not None:
--> 476             f(self, obj) # Call unbound method with explicit self
    477             return

/opt/conda/anaconda/lib/python3.6/pickle.py in save_dict(self, obj)
    820         self.memoize(obj)
--> 821         self._batch_setitems(obj.items())
    822 

/opt/conda/anaconda/lib/python3.6/pickle.py in _batch_setitems(self, items)
    846                     save(k)
--> 847                     save(v)
    848                 write(SETITEMS)

/opt/conda/anaconda/lib/python3.6/pickle.py in save(self, obj, save_persistent_id)
    495             if reduce is not None:
--> 496                 rv = reduce(self.proto)
    497             else:

/usr/lib/spark/python/lib/py4j-0.10.7-src.zip/py4j/java_gateway.py in __call__(self, *args)
   1256         return_value = get_return_value(
-> 1257             answer, self.gateway_client, self.target_id, self.name)
   1258 

/usr/lib/spark/python/pyspark/sql/utils.py in deco(*a, **kw)
     62         try:
---> 63             return f(*a, **kw)
     64         except py4j.protocol.Py4JJavaError as e:

/usr/lib/spark/python/lib/py4j-0.10.7-src.zip/py4j/protocol.py in get_return_value(answer, gateway_client, target_id, name)
    331                     "An error occurred while calling {0}{1}{2}. Trace:\n{3}\n".
--> 332                     format(target_id, ".", name, value))
    333         else:

Py4JError: An error occurred while calling o3332.__getstate__. Trace:
py4j.Py4JException: Method __getstate__([]) does not exist
	at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:318)
	at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:326)
	at py4j.Gateway.invoke(Gateway.java:274)
	at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:132)
	at py4j.commands.CallCommand.execute(CallCommand.java:79)
	at py4j.GatewayConnection.run(GatewayConnection.java:238)
	at java.lang.Thread.run(Thread.java:748)



During handling of the above exception, another exception occurred:

PicklingError                             Traceback (most recent call last)
<ipython-input-370-4f054869b703> in <module>
----> 1 shap_df1 = shap_df.withColumn("shap_val", shapvalue_udf1(f.col("shap")))

/usr/lib/spark/python/pyspark/sql/udf.py in wrapper(*args)
    187         @functools.wraps(self.func, assigned=assignments)
    188         def wrapper(*args):
--> 189             return self(*args)
    190 
    191         wrapper.__name__ = self._name

/usr/lib/spark/python/pyspark/sql/udf.py in __call__(self, *cols)
    165 
    166     def __call__(self, *cols):
--> 167         judf = self._judf
    168         sc = SparkContext._active_spark_context
    169         return Column(judf.apply(_to_seq(sc, cols, _to_java_column)))

/usr/lib/spark/python/pyspark/sql/udf.py in _judf(self)
    149         # and should have a minimal performance impact.
    150         if self._judf_placeholder is None:
--> 151             self._judf_placeholder = self._create_judf()
    152         return self._judf_placeholder
    153 

/usr/lib/spark/python/pyspark/sql/udf.py in _create_judf(self)
    158         sc = spark.sparkContext
    159 
--> 160         wrapped_func = _wrap_function(sc, self.func, self.returnType)
    161         jdt = spark._jsparkSession.parseDataType(self.returnType.json())
    162         judf = sc._jvm.org.apache.spark.sql.execution.python.UserDefinedPythonFunction(

/usr/lib/spark/python/pyspark/sql/udf.py in _wrap_function(sc, func, returnType)
     33 def _wrap_function(sc, func, returnType):
     34     command = (func, returnType)
---> 35     pickled_command, broadcast_vars, env, includes = _prepare_for_python_RDD(sc, command)
     36     return sc._jvm.PythonFunction(bytearray(pickled_command), env, includes, sc.pythonExec,
     37                                   sc.pythonVer, broadcast_vars, sc._javaAccumulator)

/usr/lib/spark/python/pyspark/rdd.py in _prepare_for_python_RDD(sc, command)
   2418     # the serialized command will be compressed by broadcast
   2419     ser = CloudPickleSerializer()
-> 2420     pickled_command = ser.dumps(command)
   2421     if len(pickled_command) > (1 << 20):  # 1M
   2422         # The broadcast will have same life cycle as created PythonRDD

/usr/lib/spark/python/pyspark/serializers.py in dumps(self, obj)
    605                 msg = "Could not serialize object: %s: %s" % (e.__class__.__name__, emsg)
    606             cloudpickle.print_exec(sys.stderr)
--> 607             raise pickle.PicklingError(msg)
    608 
    609 

PicklingError: Could not serialize object: Py4JError: An error occurred while calling o3332.__getstate__. Trace:
py4j.Py4JException: Method __getstate__([]) does not exist
	at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:318)
	at py4j.reflection.ReflectionEngine.getMethod(ReflectionEngine.java:326)
	at py4j.Gateway.invoke(Gateway.java:274)
	at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:132)
	at py4j.commands.CallCommand.execute(CallCommand.java:79)
	at py4j.GatewayConnection.run(GatewayConnection.java:238)
	at java.lang.Thread.run(Thread.java:748)

Anyone has faced a similar issue and able to resolve it? Or I'm happy to work together if anyone is working to resolving this out

[srowen]

Pardon for self-promotion, but you can most certainly use SHAP with Pyspark. It doesn't require any changes in SHAP. Here's my proof of existence: https://databricks.com/blog/2019/06/17/detecting-bias-with-shap.html

[ijoseph]

In case anyone's interested, there's finally an example notebook for our Monte Carlo implementation of Shapley values. cc @shadab-entrepreneur

[xboard]

Thanks @slundberg, as soon as I get something working properly I'll get back to you.

[slundberg]

Often with KernelExplainer you use a single background reference value (like the mean feature values), so that won't be a big issue. And you can then explain each sample individually if you like.

[evah]

@QuentinAmbard Thanks for offering your notebook, yet our answer did not solve the problem ...

• the python version, you code is using pandas data frame running single-node xgboost
• even if the scala version is parallel xgboost, it is scala and this issue's author is asking for pyspark
• people still need SHAP for spark models (random forest & gbt etc.) not for xgboost model

[evah]

@QuentinAmbard Actually I just tried it 🤦‍♀️ the head-aching part is, it takes an pyspark model, but the X (feature set) still needs to be in pandas. But I guess, to make it able to take pyspark dataframe, we need to re-implement the whole algorithm in pyspark.

[techwithshadab]

@shadab-entrepreneur that looks very similar to Niloy Gupta, my, and others' work here

I think you may need to yield rather than return, to attempt directly answer your question?

Thanks @ijoseph
I tried using Shparkley as well and getting the same error. Is there any demo notebook for the same?

[slundberg]

I don't have any plans for that since I am not a Spark user. What Spark workflow do you have in mind that you would like SHAP to integrate with? I might be able to suggest something in a specific case (though I don't know much about spark).

[richardxy]

Scott, Thanks for the reply. I think that could be alright because SHAP has a model-agnostic mode. But I found it would take a lot of memory and I couldn't run it through with a Sklearn gradient boosting tree model, while the integrated Xgboost solution works fine. Basically we have a model trained in Spark MLlib and deployed in Spark environment. We would love to provide explanability under the same environment. If the model-agnostic mode of SHAP would be memory friendly, that would be the solution for us. thanks, Richard

…

On Thu, Mar 1, 2018 at 1:46 PM, Scott Lundberg ***@***.***> wrote: I don't have any plans for that since I am not a Spark user. What Spark workflow do you have in mind that you would like SHAP to integrate with? I might be able to suggest something in a specific case (though I don't know much about spark). — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#38 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ACFv2Oo5BTwBNcOwzLk1M9UWUtMPAAx6ks5taEIDgaJpZM4SYyqK> .

[slundberg]

Ah. I have been thinking about releasing a standalone version of the C++ code we currently integrated into XGBoost. Then I could explain models from sklearn and other packages with Tree SHAP rather than waiting for them to integrate it. This is just an idea right now though so I don't have a timeline. As for memory, that can be minimized by providing just a single reference value as the background, but the right answer in your situation is really to enable Tree SHAP with sklearn gradient boosting somehow (or use XGBoost in spark). Using the model agnostic version is always going to slower and less accurate than the exact formulation. - Scott

…

On Thu, Mar 1, 2018 at 12:11 PM richardxy ***@***.***> wrote: Scott, Thanks for the reply. I think that could be alright because SHAP has a model-agnostic mode. But I found it would take a lot of memory and I couldn't run it through with a Sklearn gradient boosting tree model, while the integrated Xgboost solution works fine. Basically we have a model trained in Spark MLlib and deployed in Spark environment. We would love to provide explanability under the same environment. If the model-agnostic mode of SHAP would be memory friendly, that would be the solution for us. thanks, Richard On Thu, Mar 1, 2018 at 1:46 PM, Scott Lundberg ***@***.***> wrote: > I don't have any plans for that since I am not a Spark user. What Spark > workflow do you have in mind that you would like SHAP to integrate with? I > might be able to suggest something in a specific case (though I don't know > much about spark). > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#38 (comment)>, or mute > the thread > < https://github.com/notifications/unsubscribe-auth/ACFv2Oo5BTwBNcOwzLk1M9UWUtMPAAx6ks5taEIDgaJpZM4SYyqK > > . > — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#38 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADkTxTscToRsRigMUFeL4DyFY8SHQvyBks5taFXTgaJpZM4SYyqK> .

[pavitrasrinivasan]

Hi Scott,

I'm would also like to understand if there is integration of SHAP method with SparkML lib ?
Any updates in this regard ?

Thanks,
Pavitra

[slundberg]

@pavitrasrinivasan The TreeExplainer now works for sklearn, so that part is done. As for spark integration I don't know what that would look like since I don't use spark. Do you need something special to run a python package in spark?

[pavitrasrinivasan]

Thanks for your response !

Yes, I will not be able to use the python based package directly in spark. I would need a spark compatible version.

Thanks,
Pavitra

[slundberg]

Okay, if anyone wants to tackle adding spark support that would be great. I am happy to help, but having not used spark myself I can't promise I'll get to it anytime soon by myself.

[xboard]

Okay, if anyone wants to tackle adding spark support that would be great. I am happy to help, but having not used spark myself I can't promise I'll get to it anytime soon by myself.

Hi Scott, I've been using Spark for quite some time and would love to help adding support for Spark tree based models.

[slundberg]

@xboard great! If you take a look at a current example of parsing a tree model it should help you get started. Basically we build a TreeEnsemble object from many Tree objects, and each Tree object is built by parsing the trees from the original model and saving them in a form similar to how sklearn stores tree data: https://github.com/slundberg/shap/blob/5346fd4001ebb5e536fae68afb4097658def300b/shap/explainers/tree.py#L427-L439

[haribaskar]

@xboard You got something?
for randomforestclassifier we can use following parameter, but this is internally represented as java object. what can we do?

• model.trees
• model.minInfoGain

elif str(type(model)).endswith("pyspark.ml.classification.RandomForestClassifier'>"):
            scaling = 1.0 / len(model.estimators_) # output is average of trees
            self.trees = [Tree(e.tree, normalize=True, scaling=scaling) for e in **model.trees**]
            self.objective = objective_name_map.get(**model.minInfoGain**, None)
            self.tree_output = "probability"

[jbertscher]

@slundberg thanks for developing this great package! I'm also interested in a Spark implementation but as a first step would like to try using the model-agnostic version on a model trained with Spark.

As far as I can tell, it shouldn't matter where the model is trained, as long as I can pass a function that makes predictions based on the given features KernelExplainer. The problem would be that the dataset would be too large to fit in memory - which is why I'm using Spark. One suggestion I've had is to pass a down-sampled (but hopefully representative) version of the training data. I'm curious to know whether you think that would work?

[Chakri-V-V]

Hi ,

Is there any stable release of Shap in Pyspark till date? Sorry if my question doesn't make sense, but I am new to Pyspark & have a requirement of SHAP with a clustering analysis I am trying to solve.

[slundberg]

Not that I know of yet. Would be a great to have.

[fivejjs]

What about the random forest in Spark ML? It is also better to have a scala version to calculate shap values.

We can use sparkContext to get the tree objects as in https://github.com/fivejjs/spark-tree-plotting

[QuentinAmbard]

I've got a draft version of the ExplainerTree with a simple spark decision tree using pyspark, I'll try improve it to support random forest and commit it. @slundberg can you confirm that categorical splits aren't supported?

[QuentinAmbard]

I've created #721, feedback is welcome.

[ppakawatk]

Does SHAP support PySpark models yet?
I have tried using SHAP with RandomForestRegressionModel (PySpark ML). But got this error: "NotImplementedError: CategoricalSplit are not yet implemented".

Is there the list of supported model/ libs for SHAP?

Thank you.

[QuentinAmbard]

Yes it's supported, but my understanding is that shap doesn't support categorical split so you can't have categorical features in your model (for example a stringIndexer will create a categorical split)

[ppakawatk]

Yes it's supported, but my understanding is that shap doesn't support categorical split so you can't have categorical features in your model (for example a stringIndexer will create a categorical split)

Thanks. @QuentinAmbard. I tried with another simple model and it can run properly.
I wonder if there are cases where I need to use categorical features in my model.

However, there is another error occurs, "
AttributeError: 'DataFrame' object has no attribute 'shape'". My guess is that I'm using Spark Dataframe as an input for shap_values?

[QuentinAmbard]

Categorical features are often required in trees (using a one hot encoder isn't a good alternative)...

There is an example in this thread: #884
As per this thread you also need to disable the additivity check for now, #905 will solve this issue (I still need to make spark work on Windows in the appveyor build)

Yes I assume you're using a Dataframe.
If you need to distribute it you can implement a pandas UDF and call spah_values inside each batch, and merge the output in a spark dataframe.
Here is a scala example with a distributed shap computation: https://databricks.com/blog/2019/06/17/detecting-bias-with-shap.html

[sacmax]

Yes it's supported, but my understanding is that shap doesn't support categorical split so you can't have categorical features in your model (for example a stringIndexer will create a categorical split)

Any solution for NotImplementedError: CategoricalSplit are not yet implemented" in pyspark?

[QuentinAmbard]

@slundberg is there a way to support categorical split with the current shap implementation?
I'm not sure how this is implemented in XGBoost, is it because there is not categorical split and just numerical one (so it's actually not implemented)?

[sacmax]

I am using RandomForest and GBT in pyspark and I have to transform some categorical features using StringIndexer and some using oneHotEncoder. Is there a way/workaround to use SHAP on the transformed categorical features?

[slundberg]

A categorical multi-way split is something that is not yet in the C++ version of TreeExplainer. So that would need to be added to support this. I would like to say I’ll just take a few days and write it, but I am not sure when I will have those free days. I’ll try to get it in sooner rather than later...

[QuentinAmbard]

@evah I think it does solve the problem ;)
The notebook just show you how you can compute your explanations using spark on multiple node (explanation of any model, in this case a single node xgboost).
If you want to get the explanation of your spark Tree-based model, you can find an example here:
https://github.com/QuentinAmbard/shap/blob/%23866_spark_regressor_support/tests/explainers/test_tree.py#L190
(As sacmax mentioned, categorical features are not implemented)

[ijoseph]

@shadab-entrepreneur that looks very similar to Niloy Gupta, my, and others' work here

I think you may need to yield rather than return, to attempt directly answer your question?

[ijoseph]

Hm, interesting. The readme has example usage but there's no example notebook as of yet.

Looking again at your stack trace, are you calling df.rdd.mapParittions inside of shapvalue_udf1? Mostly guessing based on seeing what error you got means: py4j.Py4JException: Method __getstate__([]) does not exist and this syntax:

shap_df1 = shap_df.withColumn("shap_val", shapvalue_udf1(f.col("shap")))

[davezhouwa]

@shadab-entrepreneur have you solved the issue? I'm facing the same problem.

[techwithshadab]

@davezhouwa Nope, still I'm facing the same issue. Were you able to solve it?

[QuentinAmbard]

It's working fine in my notebook, make sure you can serialize the explainer:
pickle.dumps(explainer)
If you can't then something is wrong (you didn't removed the line)
You need to edit it manually as it's not yet released in the last shap version from pypi

model = pipeline.stages[-1].bestModel
print(model)
 ==> Out[22]: GBTClassificationModel: uid = GBTClassifier_367df9840c81, numTrees=5, numClasses=2, numFeatures=8
explainer = shap.TreeExplainer(bestModel)
#you can check if we can serialize the explainer: 
#pickle.dumps(explainer)

And now it's working in a pandas_udf, or using the new mapInPandas in spark 3:

def compute_shap_values(iterator):
  for X in iterator:
      yield pd.DataFrame(explainer.shap_values(X, check_additivity=False))

display(dataset.mapInPandas(compute_shap_values, schema=dataset.schema).toDF(*[x+"_shap_value" for x in test.columns]))

[davezhouwa]

@shadab-entrepreneur, I followed @QuentinAmbard suggestions and it worked

[boechat107]

Hoping someone might benefit, my approach to the problems described in this issue was:

1. randomly sample the target Spark DataFrame (to make sure the data fits the master node)
2. convert the DF to a numpy array
3. calculate SHAP

def df2numpy(df: DataFrame, cols: List[str], nsamples: int) -> np.ndarray:
    """
    Converts a DataFrame into a Numpy array (matrix). "cols" gives the order of
    the matrix columns; "nsamples" is the approximate number of records.
    """
    # This is not efficient; "count" may be expensive to compute.
    fraction = nsamples / df.count()
    samp_df = df.sample(fraction=fraction, seed=6969) if fraction < 1 else df
    return np.array(samp_df.select(*cols).collect())

I also had problems with categorical features (NotImplementedError: CategoricalSplit are not yet implemented). I had forgotten that Spark implicitly uses column metadata to identify indexed features and make "categorical splits". To solve this issue, I just created another DF without any metadata:

    # "gbt" is an instance of GBTClassifier.
    # "df" contains numeric and categorical/indexed features.
    pipe = Pipeline(stages=[assembler, gbt]).fit(
        # Remove metadata of the input DF. We don't want to handle categorical
        # values when training a GBT model (Spark models implicitly use
        # metadata to handle categorical features).
        df.rdd.toDF()
    )

[zwag20]

Hoping someone might benefit, my approach to the problems described in this issue was:

1. randomly sample the target Spark DataFrame (to make sure the data fits the master node)
2. convert the DF to a numpy array
3. calculate SHAP

def df2numpy(df: DataFrame, cols: List[str], nsamples: int) -> np.ndarray:
    """
    Converts a DataFrame into a Numpy array (matrix). "cols" gives the order of
    the matrix columns; "nsamples" is the approximate number of records.
    """
    # This is not efficient; "count" may be expensive to compute.
    fraction = nsamples / df.count()
    samp_df = df.sample(fraction=fraction, seed=6969) if fraction < 1 else df
    return np.array(samp_df.select(*cols).collect())

I also had problems with categorical features (NotImplementedError: CategoricalSplit are not yet implemented). I had forgotten that Spark implicitly uses column metadata to identify indexed features and make "categorical splits". To solve this issue, I just created another DF without any metadata:

    # "gbt" is an instance of GBTClassifier.
    # "df" contains numeric and categorical/indexed features.
    pipe = Pipeline(stages=[assembler, gbt]).fit(
        # Remove metadata of the input DF. We don't want to handle categorical
        # values when training a GBT model (Spark models implicitly use
        # metadata to handle categorical features).
        df.rdd.toDF()
    )

I didn't not find that the categorical features fixed the CategoricalSplit are not yet implemented.
Has anyone found a fix for this?

[demirbilek95]

Hey everyone,
Is there any update on CategoricalSplit implementation?
Thanks in advance!

Hoping someone might benefit, my approach to the problems described in this issue was:

1. randomly sample the target Spark DataFrame (to make sure the data fits the master node)
2. convert the DF to a numpy array
3. calculate SHAP

def df2numpy(df: DataFrame, cols: List[str], nsamples: int) -> np.ndarray:
    """
    Converts a DataFrame into a Numpy array (matrix). "cols" gives the order of
    the matrix columns; "nsamples" is the approximate number of records.
    """
    # This is not efficient; "count" may be expensive to compute.
    fraction = nsamples / df.count()
    samp_df = df.sample(fraction=fraction, seed=6969) if fraction < 1 else df
    return np.array(samp_df.select(*cols).collect())

I also had problems with categorical features (NotImplementedError: CategoricalSplit are not yet implemented). I had forgotten that Spark implicitly uses column metadata to identify indexed features and make "categorical splits". To solve this issue, I just created another DF without any metadata:

    # "gbt" is an instance of GBTClassifier.
    # "df" contains numeric and categorical/indexed features.
    pipe = Pipeline(stages=[assembler, gbt]).fit(
        # Remove metadata of the input DF. We don't want to handle categorical
        # values when training a GBT model (Spark models implicitly use
        # metadata to handle categorical features).
        df.rdd.toDF()
    )

I didn't not find that the categorical features fixed the CategoricalSplit are not yet implemented.
Has anyone found a fix for this?

[varunbanda]

Hello!

Will CategoricalSplit be implemented in the next release? Any information?

[olbapjose]

Hi, any news on this? I have seen the contributionPredictionCol but I am not sure whether SHAP values can now be calculated on the Spark version of XGBoost or there are still cornercases or workarounds.

[venser12]

Hi, is CategoricalSplit yet implemented?