DETAILED ACTION
This action is written in response to the RCE filed 4/17/26. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Arguments
The Applicants argue that the previous art of record does not anticipate or render obvious the claims as currently amended. The Examiner provides updated prior art rejections below necessitated by the current amendments.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103(a) which forms the basis for all obviousness rejections set forth in this Office action:
(a) A patent may not be obtained though the invention is not identically disclosed or described as set forth in section 102 of this title, if the differences between the subject matter sought to be patented and the prior art are such that the subject matter as a whole would have been obvious at the time the invention was made to a person having ordinary skill in the art to which said subject matter pertains. Patentability shall not be negatived by the manner in which the invention was made.
The following are the references relied upon in the rejections below:
Wang (EP 3836037 A1, cited by Applicant in IDS dated 7/21/23)
Karri (US 2023/0015531 A1)
Claims 1, 3, 5-6, 8, 10, 12-13, 15, 17 and 19-20 are rejected under 35 U.S.C. 103 as being unpatentable over Wang and Karri.
Regarding claims 1, 8 and 15, Wang discloses a data processing method, comprising:
acquiring a target directed acyclic graph (DAG) corresponding to a service processing logic of a model self-taught learning service, wherein the service processing logic comprises:
“[0058] Specifically, the directed acyclic graph (DAG diagram) shown in the middle part of FIG. 7 shows 6 nodes: "feedback data" node, "behavioral data" node, "data splitting" node, and "feature engineering" node, "LR (logistic regression) algorithm" node, "GBDT (gradient boosting decision tree) algorithm" node, "HE-TreeNet (high-dimensional discrete embedded tree network) algorithm" node, and "NN (neural network) algorithm" node.” (Emphasis added.)
PNG
media_image1.png
386
496
media_image1.png
Greyscale
Wang, fig. 7.
execution logic for acquiring service data generated by an online released service model, execution logic for training a to-be-trained service model based on the service data, and execution logic for releasing the trained service model online; and
[0022] “The model auto-training unit 130 may, according to a configured model updating scheme, generate updated training samples based on the collected prediction data and corresponding real results thereof and continuously obtain the updated machine learning models by using the updated training samples.” (Emphasis added.)
[0024] “As an example, before the model auto-training unit 130 continuously obtains the updated machine learning models, there is already an initial machine learning model in the system 100, and the initial machine learning model may be a machine learning model previously trained by the system 100 (for example, the model autotraining unit 130) by using a model training scheme, …”
[0067] “In step S205, the service providing unit 140 may automatically save the prediction data included in the prediction service request, and the data collecting unit 110 may continuously collect the prediction data from the service providing unit, wherein the collected prediction data (with corresponding real results) will be used to obtain the updated machine learning models by the model auto-training unit 130, which will be described in detail later. Through the step S205, the automatic backflow of data may be realized, thereby providing a necessary data source for the continuous loop of the automatic machine learning processes.” (Emphasis added.)
performing self-taught learning on the to-be-trained service model according to the target DAG;
[0026] “After the updated training samples are generated, the model auto-training unit 130 may further continuously obtain the update machine learning models by using the updated training samples according to settings regarding model training (for example, the model algorithms, the parameter adjusting and optimizing, etc.) defined in the configured model updating scheme. As described above, the configured model updating scheme may be generated by the model auto-training unit 130 on the basis of the model training scheme based on which the initial machine learning model is trained, or it may be any scheme for continuously training and obtaining the machine learning models, the model updating scheme herein aims to emphasize that the scheme may be used to more automatically and continuously generate models, but does not limit the manners of model generation to full retraining or incremental learning training.” (Emphasis added.)
wherein a service model is a resource recommendation model, and the service data is interactive data of a recommended resource or
[0031] “For the updating resource auto-configuration manner, the model auto-training unit 130 needs to know how to utilize system resources (for example, CPU, bus, bandwidth, memory and other resources) during the process of obtaining the updated machine learning models. Here, the auto-training unit 130 may configure the resources according to a data amount together with a rule, but the disclosure is not limited thereto.” (Emphasis added.)
[The Examiner notes that only the first limitation of this Markush group is taught by Wang.]
wherein the target DAG comprises at least two DAG subgraphs, different DAG subgraphs are configured to implement different execution logic, and the different DAG subgraphs construct the target DAG based on a data flow direction of the service processing logic;
[0059] “Referring to FIG. 7, through corresponding configuration at the "data splitting" node in the DAG diagram, the model auto-training unit 130 may split the historical data into the training data and the verification data. Thereafter, through corresponding configuration at the "feature engineering" node in the DAG graph, the model autotraining unit 130 may perform automatic feature generation on the split training data/validation data to extract at least one feature, preferably, the model auto-training unit 130 may also perform automatic feature combination after automatic feature generation to obtain various features including combined features.”
[0109] “Referring to FIG. 7, through the corresponding configuration at the "data splitting" node in the DAG diagram, the training data obtained after the splicing of the behavioral data and the feedback data may be split into a training set and a validation set. Thereafter, through the corresponding configuration at the "feature engineering" node in the DAG diagram, automatic feature generation may be performed on the training set and the validation set to extract at least one feature to generate a training sample. At the three nodes corresponding to the lowest layer in the DAG diagram (i.e. "LR algorithm" node, "GBDT algorithm" node, "HE-TreeNet algorithm" node and "NN algorithm" node), the training samples is utilized to perform at least one round of training with respect to the four preset algorithms, respectively, and then the corresponding multiple machine learning models are trained.”
wherein in a case where the at least two DAG subgraphs comprise a training DAG subgraph that implements the execution logic for training the to-be-trained service model based on the service data, performing the self-taught learning on the to-be-trained service model according to the target DAG comprises:
operating the model training DAG subgraph to train the to-be-trained service model according to the service data in a case where a training condition is satisfied;
[0029] “the model auto-training unit 130 may update the machine learning model according to a certain model updating cycle (i.e., generate a new machine learning model). The model updating cycle may be pre-configured by the user, or may be modified in real time according to a specific condition based on a certain rule.” (Emphasis added.)
[0079] “Specifically, the service providing unit 140 may select one or more machine learning models as the online machine learning model from the machine learning models obtained and stored by the model auto-training unit 130 according to the model selecting rule included in the model application scheme, wherein the model selecting rule may include a rule for selecting the machine learning model with the highest AUC, a rule for selecting the newly generated machine learning model or the like. For example, the service providing unit 140 may select the machine learning model with the highest AUC from the stored machine learning models as the online machine learning model according to the AUC value.” (Emphasis added.)
wherein the satisfied training condition comprises at least one of the following: start training time of a preset training period being reached, and duration of acquisition of the service data reaching preset duration;
[The Examiner notes that this is a Markush group.]
[0105] “In Fig. 15, a configuration of a self-learning cycle is provided, the user may select the operating mode as "single run", "cyclic run" and "crontab expression", and select a task start time as "2019-06-17 11:38:43", and a self-learning data configuration is further provided, the users may perform selection of data source, data slices, model naming result, and task timeout duration, etc.” (Emphasis added.)
start training time :: “task start time”
duration of acquisition of the service data :: “task timeout duration”
Karri discloses the following further limitations which Wang does not disclose
wherein the preset training period is a period in which the to-be-trained service model performs training according to the service data, when the start training time of the preset training period is reached, the to-be-trained service model is controlled to perform iterative training;
[0077] “Service level management 84 provides cloud computing resource allocation and management such that required service levels are met. Service Level Agreement (SLA) planning and fulfillment 85 provide pre-arrangement for, and procurement of, cloud computing resources for which a future requirement is anticipated in accordance with an SLA.” (Emphasis added.)
[0030] “the computing resource allocator 134 extracts features computing tasks and their projected start times, completion times, durations, workflow sequences or dependencies, computing resources required, non-computing resources required, priority levels, etc. from the collected training data.” (Emphasis added.)
[0024] “machine learning methods, such as neural networks”.
The Examiner notes that neural networks are inherently trained iteratively.
wherein the to-be-trained service model is trained when the duration of the acquisition of the service data for training reaches the preset duration.
[0024] “In embodiments, such features may include features such as computing tasks and their projected start times, completion times, durations, workflow sequences or dependencies, computing resources required, non-computing resources required, priority levels, etc.” (Emphasis added.)
At the time of filing, it would have been obvious to a person of ordinary skill to apply the service level agreement / scheduling techniques disclosed by Karri with the Wang system because it provides for the efficient use of computing resources by anticipating bottlenecks (eg due to dependencies) and scheduling resource use accordingly.
Regarding independent claims 8 and 15, the recited computing components (ie “at least one processor”, “a memory”, and a “non-transitory computer-readable storage medium” are inherent throughout the Wang disclosure.
Regarding claims 3, 10 and 17, Wang discloses the further limitations wherein in a case where the at least two DAG subgraphs comprise an acquisition DAG subgraph that implements the execution logic for acquiring the service data generated by the online released service model, performing the self-taught learning on the to-be-trained service model according to the target DAG comprises:
operating the acquisition DAG subgraph to acquire the service data in a case where an acquisition condition is satisfied when the online released service model generates the service data in response to a service request.
[0029] “the model auto-training unit 130 may update the machine learning model according to a certain model updating cycle (i.e., generate a new machine learning model). The model updating cycle may be pre-configured by the user, or may be modified in real time according to a specific condition based on a certain rule.” (Emphasis added.)
[0079] “Specifically, the service providing unit 140 may select one or more machine learning models as the online machine learning model from the machine learning models obtained and stored by the model auto-training unit 130 according to the model selecting rule included in the model application scheme, wherein the model selecting rule may include a rule for selecting the machine learning model with the highest AUC, a rule for selecting the newly generated machine learning model or the like. For example, the service providing unit 140 may select the machine learning model with the highest AUC from the stored machine learning models as the online machine learning model according to the AUC value.” (Emphasis added.)
Regarding claims 5, 12 and 19, Wang discloses the further limitations wherein in a case where the at least two DAG subgraphs comprise a model online DAG subgraph that implements the execution logic for releasing the trained service model online, performing the self-taught learning on the to-be-trained service model according to the target DAG comprises:
operating the model online DAG subgraph to release the trained service model online in a case where a releasing online condition is satisfied.
[0029] “the model auto-training unit 130 may update the machine learning model according to a certain model updating cycle (i.e., generate a new machine learning model). The model updating cycle may be pre-configured by the user, or may be modified in real time according to a specific condition based on a certain rule.” (Emphasis added.)
[0073] “For example, the model updating cycle may be set to 1 week, the data selecting rule may be set to select data according to a time range (for example, a data range is set to "last 7 days"), and the model storage location may be set to the model center inside the system 100, and the updating resource auto-configuration manner is set to configure the resources according to the data amount in conjunction with a rule.”
Regarding claims 6, 13 and 20, Wang discloses the further limitations wherein the model online DAG subgraph comprises a model releasing DAG subgraph and a model push DAG subgraph; and
operating the model online DAG subgraph to release the trained service model online when the releasing online condition is satisfied comprises:
operating the model releasing DAG subgraph to release the trained service model to a model center when a releasing condition is satisfied; and
[0029] “the model auto-training unit 130 may update the machine learning model according to a certain model updating cycle (i.e., generate a new machine learning model). The model updating cycle may be pre-configured by the user, or may be modified in real time according to a specific condition based on a certain rule.” (Emphasis added.)
[0030] “For the model storage location, due to the continuous updating of the machine learning model, multiple machine learning models will be obtained, in order to enable the service providing unit 140 to select an online machine learning model used to provide an online prediction service from the multiple machine learning models, the model auto-training unit 130 needs to determine locations for storing the updated machine learning models which are continuously obtained. For example, the machine learning models may be stored in a model center inside the system 100, which may also enable the user to view model-related interpretations and reports.” (Emphasis added.)
operating the model push DAG subgraph to control to push the trained service model from the model center to an online platform according to a preset push requirement in a case where a push condition is satisfied.
Id.
Additional Relevant Prior Art
The following references were identified by the Examiner as being relevant to the disclosed invention, but are not relied upon in any particular prior art rejection:
Jin discloses a system for graph-based analysis of job flow data which includes directed acyclic graphs (DAGs). See eg fig. 12. (US 10,380,185 B2)
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Vincent Gonzales whose telephone number is (571) 270-3837. The examiner can normally be reached on Monday-Friday 7 a.m. to 4 p.m. MT. If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Miranda Huang, can be reached at (571) 270-7092.
Information regarding the status of an application may be obtained from the USPTO Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov.
/Vincent Gonzales/Primary Examiner, Art Unit 2124
Prosecution Insights