201). The dimensions of the multi-dimensional representation correspond to characteristics of the document, such as document attributes or user activity in or on the document. The multi-dimensional representation of the document includes values indicative of characteristics of the document. In an implementation, the characteristics include quantifiable metrics, and the multi-dimensional representation includes values for the corresponding document.

Next, the data logging service generates from the multi-dimensional representation of the document a reduced dimension representation of the document (step 203). In an implementation, the reduced dimension representation includes document data in a format of fewer dimensions than the multi-dimensional representation. For example, the data logging service converts the values of the multi-dimensional representation to a single value, such as a binary number, binary string, or hashed value.

To generate the reduced dimension representation of the document, in an implementation, the data logging service first reduces the resolution of each of the data values of the multi-dimensional representation. For each data value of the multi-dimensional representation, the data logging service identifies a subset of values to which the value belongs from among a range of possible values for the corresponding document characteristic. Having identified a subset for the value, the data logging service then converts the value to a subset identifier which maps to otherwise represents the identified subset. In performing this type of conversion for every data value of the multi-dimensional representation, the data logging service forms a combination of the subset identifiers.

Next, the data logging service uses the combination of the subset identifiers to generate a single value which is uniquely representative of the combination. In an implementation, the data logging service maps the combination of subset identifiers to a unique identifier from a set of combination identifier values. The set of combination identifiers uniquely identifies every possible combination of subset identifiers. Thus, for any given combination identifier, a particular combination of subset identifiers can be identified.

In an implementation, a binary string can be used to represent the combination of subset identifiers by setting the number of positions of the string equal to the total number of possible combinations. Each of the positions of the binary string corresponds to a particular combination of subset identifiers. For a brief example, a dataset may include six document metrics, and the numbers of subsets for each metric may vary according to the range of possible values of the metric. If the number of possible subsets of the metrics is 5, 3, 4, 4, and 2, then the total number of possible combinations is 480. Thus, a binary string of 480 positions can be used to uniquely represent each of the 480 possible combinations by setting the position corresponding to a given combination to “1” and all other positions to “0.”

Having created a binary string to represent the reduced dimension representation, the data logging service adds noise to the single value to generate a noisy representation of the document (step 205). In an implementation, the data logging service randomly selects and flips some of the “0” values of the binary string to “1” according to a predetermined probability. Adding noise to the binary representation further enhances protection of privacy of the document data by obscuring the true “1” from the randomly generated “1s.”

The data logging service generates telemetry data based at least on the noisy representation of the document (step 207). For example, the data logging service may create a JSON object containing the noisy representation for transmission via an API (application programming interface) to a telemetry service. The data logging service then transmits the noisy representation of the document to a telemetry service (step 209). The telemetry service collects other similarly configured noisy representations for the document and for other documents from the data logging service and from other data logging services.

FIG. 3 illustrates telemetry process 300 in an implementation, hereinafter referred to as process 300. Process 300 may be implemented in program instructions in the context of any of the software applications, modules, components, or other such elements of one or more computing devices. The program instructions direct the computing device(s) to operate as follows, referred to in the singular for the sake of clarity.

In an implementation, a telemetry service executing on a server computer receives telemetry associated with an electronic document from a data logging service executing on a computing device (step 301). In an implementation, the telemetry service interfaces with the data logging service via an API. The data logging service may be triggered to configure and transmit the telemetry data when a particular action is performed on or in the document, such as when the document is opened or saved, or when the user performs a particular action, such as adding a comment or inserting a hyperlink.

The telemetry service receives similarly configured telemetry data from multiple other data logging services executing on the same or other computing devices. To extract the document data from the noisy representations of the telemetry data from the various sources, the telemetry service aggregates the noisy representations to suppress or filter out the noise and reveal the true reduced dimension representations (step 303). In an implementation, the noisy representations are in the form of binary strings each containing an indication of the true combination (of subset identifiers) along with false indications of combinations pseudo-randomly added to the strings according to a predetermined probability. To remove or diminish the effects of the added noise, for each position of the binary strings, the telemetry service generates a sum of the values of the binary strings at the position. The telemetry service then identifies positions whose sums exceed a threshold value. The identified positions are indications of the true combinations (step 305). Having identified the true combinations of subset identifiers, the telemetry service can extract the subsets according to the subset identifiers. The combinations correlate the subsets of metric data and can be used to identify trends in the data.

In various implementations, the telemetry service or a related service specifies configuration requirements for the telemetry data according to the application. For example, the telemetry service may transmit to the data logging services a set of document characteristics for data captured from word processing application documents, and a different set of document characteristics for data captured from spreadsheet application documents. In addition to specifying the document characteristics, the telemetry service specifies the subsets for each metric. The subsets may be defined by boundary values. The telemetry service also specifies what event caused the data logging service to capture the document data.

In some implementations, the telemetry service or a related service updates the configuration requirements according to the telemetry data received. For example, the telemetry service may determine that the subsets fail to provide sufficient resolution to analyze a metric. Alternatively, the telemetry service may determine that the subsets are too granular to reveal a distribution of the data (i.e., the subsets flatten the data distribution). The telemetry service may update the subsets for the metric and deploy the updated subsets (or updated boundary values of the subsets) to the data logging services.

Returning to FIG. 1, the following is a brief example of processes 200 and 300 with respect to operational architecture 100. In an implementation, data logging service 120 executing on client computing device 110 generates a multi-dimensional representation of an electronic document opened in an application executing on client computing device 110. In a simplified illustration of the processes disclosed herein, the multi-dimensional representation comprises five metrics A, B, C, D, and E each of which may be a document attribute, or an action performed by the application or a user on the document. In generating the multi-dimensional representation, data logging service 120 captures raw data values 21, 7, 4, 2, and 9 for the five metrics, respectively, as shown in view 121.

Data logging service 120 generates a reduced dimension representation of the document based at least on the multi-dimensional representation of the document, as shown in view 122. In an implementation, data logging service 120 converts the data values to subset identifiers in a process of which reduces the resolution of the data. For each of the five data values, the subset identifier identifies a subset of values to which the data value belongs. The identified subset is one of multiple, non-overlapping subsets each representing a subset of values from the set of all possible values of the corresponding metric. In the simplified illustration, the raw data values are converted to subset identifiers 2, 1, 2, 1, and 3, which indicates that, for example, A=21 falls in the second subset for metric A, B=7 falls in the first subset of metric B, and so on.

Having converted the raw data values to subset identifiers, data logging service 120 determines a combination identifier (“612”) for the particular combination of subset identifiers, that is, for subset identifiers 2, 1, 2, 1, 3. In an implementation, data logging service 120 maps the set of subset identifiers to a unique, single-valued combination identifier (“612”) which uniquely identifies the combination from all the possible combinations of subset identifiers for the five metrics. Thus, a telemetry service receiving the combination identifier 612 can determine the subset identifiers for the document dataset.

In an implementation, the combination identifier is converted to a binary representation, such as a binary string or binary number. In an implementation, a binary string is configured to represent every possible combination of subset identifiers, that is to say, to have the same number of positions as there are possible combinations. To uniquely represent a combination, the position of the binary string corresponding to the combination is set to “1” while all the other positions are set to “0.” In this way, the combination can be identified from the string by determining the position of the only “1” in the string.

Because the reduced resolution values (i.e., subset identifiers) can be obtained from the binary string, differential privacy methods can be used to further strengthen the privacy protection of the data. In an implementation, data logging service 120 randomly selects some positions of the binary string and flips the value of those positions from “0” to “1” according to a predetermined probability, as shown in view 123. The position of the string corresponding to the combination, already set to “1,” is not changed. In an implementation, data logging service 120 ensures that some minimum number of additional positions is flipped. Because the randomized string contains multiple indications of a combination being represented, the true representation is obscured.

Data logging service 120 generates telemetry data based at least on the noisy representation of the document and transmits the telemetry data to telemetry service 140 executing on remote server computer 130. Remote server computer receives the telemetry data in the form of the noisy representation of the document, as shown in view 141. Telemetry service 140 aggregates the noisy representation with other, similarly configured noisy representations of other documents from other computing devices. In an implementation, to aggregate the noisy representations, telemetry service 140 generates a sum for each position of the binary strings to which noise was added. With each position corresponding to a particular combination of subset identifiers, the aggregation generates a frequency of occurrence of 1's for all the combinations. As shown in view 142, the frequencies of all the combinations will vary according to the captured document data but also according to the randomly added noise. Telemetry service 140 determines the combinations, for which the frequencies exceed a specified threshold value, as being indicative of true combinations of metric values. Outlier combinations of metric values will generally be obscured by the noise. The subsets of the combinations identified as being true can be extracted (as shown in view 143) for further analysis.

FIG. 4 illustrates telemetry data processes performed by data logging service 420, of which data logging service 120 in FIG. 1 is representative. Data logging service 420 executes on a client computing device in an implementation. Data logging service 420 captures data relating to a document which is open in an application executing on the computing device. Capturing the data may be triggered by an event which occurs in connection with the document, such as the selection by a user of an application command or function which has been instrumented to trigger data capture or by an action performed automatically by the application on the document, such as when the document is auto-saved.

In view 421, data logging service 420 has captured document data including five metrics for the electronic document. The metrics are characteristics of the document, such as document attributes or actions performed on the document by the user or by the application. In a simplified example of the concepts, five data values (12, 5, 0, 8, 16) are captured for five characteristics of the document (V, W, X, Y, and Z). The captured data set forms a multi-dimensional representation of the document.

In view 422, data logging service 420 converts the captured data values to new, reduced resolution values. The reduced resolution values are subset identifiers which identify subsets of data values into which the data values fall. For example, for metric V, the data value of 12 falls into the first subset of values for metric V. For metric Z, the data value 16 falls into the fourth subset of values for metric Z.

Data logging service 420 forms a reduced dimension representation of the document metrics data shown in view 421 based on the combination of subset identifiers (1, 2, 1, 3, 4). In an implementation, every possible combination of subset identifiers maps is uniquely identified. For example, any given combination of subset identifiers may be uniquely identified according to a serially numbered listing of every possible combination, where the serial numbers form a reduced dimension representation of the corresponding combination. Alternatively, a unique combination identifier for the combination may be algorithmically determined for any given combination based on the subset identifiers, forming another type of reduced dimension representation of the combination.

In an implementation, the combination of subset identifiers (e.g., 1, 2, 1, 3, 4) maps to a position in a binary string, where every position of the binary string maps to a particular combination of subset identifiers in a one-to-one correlation with every combination represented. For example, the combinations may be numbered from 1 to n, where n represents the total number of possible combinations of subset identifiers. A binary string of n positions will uniquely represent the ith combination (where i is number in the inclusive range 1 to n) by setting the ith position of the string to “1” and all the other n−1 positions to “0.”