TECHNIQUES FOR RECOMMENDING NEXT COMMANDS USING RECURRENT NEURAL NETWORKS AND HIDDEN STATE CLUSTERING
In example embodiments, techniques are provided for determining next command recommendations using a trained recurrent neural network model. A command prediction module of an application gathers command data and user characteristic data for a user, and cleans the command data to produce an input dataset. The command prediction module applies the input dataset to a trained recurrent neural network model, where the trained recurrent neural network model is configured to produce a separate next command prediction for each of a plurality of different values of one or more user characteristics. The command prediction module selects one or more recommended next commands from within the next command prediction produced for a value of one or more user characteristics that correspond to the user characteristic data for the user, and provides the one or more recommended next commands for display in a user interface of the application.
1 . A method for recommending one or more next commands to a user of an application, comprising:
gathering, by a command prediction module of the application executing on a computing device, command data and user characteristic data for the user;
cleaning the command data to produce an input dataset;
applying, by the command prediction module, the input dataset to a trained recurrent neural network model, the trained recurrent neural network model configured to produce a separate next command prediction for each of a plurality of different values of one or more user characteristics;
selecting, by the command prediction module, one or more recommended next commands from within the next command prediction produced for a value of the one or more user characteristic that corresponds to the user characteristic data for the user; and
displaying the one or more recommended next commands in a user interface of the application.
2 . The method of claim 1 , wherein the trained recurrent neural network model is configured to produce the separate next command predictions by clustering final hidden states from a last hidden layer of the trained recurrent neural network model, associating each cluster with a value of one or more user characteristics, and having the output layer of the trained recurrent neural network model produce the separate next command predictions based on the final hidden states from each cluster.
3 . The method of claim 2 , wherein the trained neural network model is a trained gated recurrent unit (GRU) neural network model and the last hidden layer is a last GRU layer.
4 . The method of claim 1 , wherein the applying further comprises:
extracting a past command sequence from the input dataset;
encoding the past command sequence as a plurality of vectors; and
providing the plurality of vectors to an input layer of the trained recurrent neural network model.
5 . The method of claim 1 , wherein the trained recurrent neural network model is configured to produce an associated confidence level for each command within each separate next command prediction, and the selecting further comprises selecting a command having a greatest confidence level or selecting each command having an associated confidence level above a given threshold
6 . The method of claim 1 , wherein the gathering further comprises:
processing the command data to determine the user characteristics data, the processing to include comparing aspects of the command data to one or more thresholds.
7 . The method of claim 1 , wherein the gathering further comprises:
soliciting the user to provide the user characteristics data in the user interface of the application.
8 . The method of claim 1 , wherein the cleaning further comprises:
removing commands on a predetermined list of commands from the command data;
removing instances of sequential commands that occur more than a threshold number of times from the command data; or removing commands that occur less frequently than a threshold from the command data.
9 . The method of claim 1 , further comprising:
comparing an actual next command selected by the user to the one or more recommended next commands; and
in response to the actual next command not matching any of the one or more recommended next commands, there being less than a threshold number of previous incorrect predictions, and one or more recommended next commands having a confidence level of above a threshold level, determining the user switched tasks.
10 . The method of claim 1 , further comprising:
comparing an actual next command selected by the user to the one or more recommended next commands; and
in response to the actual next command not matching any of the one or more recommended next commands, there being greater than a threshold number of previous incorrect predictions, and one or more recommended next commands having a confidence level of above a threshold level, determining the user is having difficulty operations the application.
11 . The method of claim 1 , further comprising:
comparing an actual next command selected by the user to the one or more recommended next commands; and
in response to the actual next command matching one of the one or more recommended next commands, there being greater than a threshold number of previous correct predictions, and one or more recommended next commands having a confidence level of above a threshold level, determining the user well-understands how to operate the application.
12 . A computing device configured to recommend one or more next commands to a user of an application, the computing device comprising:
a processor; and
a memory coupled to the processor, the memory configured to maintain a command prediction module of the application that when executed on the processor is operable to:
obtain an input dataset,
extract a past command sequence from the input dataset,
encode the past command sequence as a plurality of vectors, and
provide the plurality of vectors to an input layer of a trained recurrent neural network model,
produce using the trained recurrent neural network model a separate next command prediction for each of a plurality of different values of one or more user characteristics,
select one or more recommended next commands from within the next command prediction produced for a value of one or more user characteristic that corresponds to the user characteristic data for the user, and
provide the one or more recommended next commands.
13 . The computing device of claim 12 , wherein the trained recurrent neural network model is configured to produce the separate next command predictions by clustering final hidden states from a last hidden layer, associating each cluster with a value of one or more user characteristics, and having an output layer produce the separate next command 5 predictions based on the final hidden states from each cluster.
14 . The computing device of claim 12 , wherein the one or more user characteristics comprise a user skill level or a user industry sector.
15 . A non-transitory computing device readable medium having instructions stored thereon, the instructions when executed by one or more computing devices operable to:
gather command data and user characteristic data for a user;
clean the command data to produce a input dataset;
apply the input dataset and the user characteristic data to a trained recurrent neural network model, the trained recurrent neural network model configured to produce a next command prediction including one or more recommended next commands for a value of one or more user characteristics that corresponds to the user characteristic data for the user; and
display the one or more recommended next commands in a user interface.
16 . The non-transitory computing device readable medium of claim 15 , wherein the trained neural network model is a trained gated recurrent unit (GRU) neural network model and the last hidden layer is a last GRU layer.
17 . The non-transitory computing device readable medium of claim 15 , wherein the instructions that when executed are operable to apply further comprise instructions that when executed are operable to:
extract a past command sequence from the input dataset;
encode the past command sequence as a plurality of vectors; and
provide the plurality of vectors to an input layer of the trained recurrent neural network model.
18 . The non-transitory computing device readable medium of claim 15 , wherein the trained recurrent neural network model is configured to produce an associated confidence level for each of command with a separate next command prediction, and the instructions that when executed are operable to select further comprise instructions that when executed are operable to select a command having a greatest confidence level or select each command having an associated confidence level above a given threshold.
19 . The non-transitory computing device readable medium of claim 15 , further comprising instructions operable to:
process the command data to determine the user characteristics data.
20 . The non-transitory computing device readable medium of claim 15 , wherein the instructions operable to clean the command data comprise instructions operable to:
remove commands on a predetermined list of commands from the command data;
remove instances of sequential commands that occur more than a threshold number of times from the command data; or
remove commands that occur less frequently than a threshold from the command data.