In the unprecedented era of Covid-19, an alteration in the landscape for online education is clearly manifested by the hundreds of thousands of educators and learners setting out to academic cyberspace and OLEs. This is a paradigmatic change, a ‘black swan’ moment47, 官网最新版telegram下载的入口是多少 as the unforeseen event of Covid-19 pandemic ushers the educational practices in video conferencing platforms (e.g., Zoom, WebEx, MS Teams) and LMS-based uses最新的中文版telegram下载的地址在哪呢. Surely, there is a high variability in the way educators act online (often for the first time) for offering remote instruction to their students outside the physical classroom. The abrupt ending of in-person classes leading to online settings can speed up the adoption of OLEs as learning mediators. Nevertheless, this instructional change, in such a compressed time frame, has the risk to solely create an insipid copy of today’s best online learning practices. Possibly, this is due, in part, to the lack of investment in online education modality by many educational institutions and/or to underestimation of online learning as a core aspect of their learner experience48. However, led by top universities, a noticeable change began a few years ago, as fully digital academic experiences started flourishing49. The current crisis due to Covid-19 is accelerating this trend, revealing the need for Higher Education Institutions (HEIs) to promote faculty’s digital skills无障碍中文版telegram的下载地方是多少. The latter can be facilitated by the construction of a technological backbone, to mitigate the effects of this crisis and to welcome the online teaching/learning within a digital era. These digital competences can amalgamate the short-term response to crisis into an enduring digital transformation of education contexts.
最新官网的telegram的下载的入口在哪里
In this disrupted educational landscape, the issues of how instructors and colleges treat student evaluation and how institutions treat student evaluations of professors have surfaced. Educators face a challenge/opportunity in trying to evaluate quality, even as the educational activity they are evaluating is mutating, in real time. DeepLMS comes into foreground as a means to offer quantitative metrics of the user’s LMS-based QoI, as an alternative to conventional evaluation metrics. The efficient predictive performance of the DeepLMS, as justified by the experimental results derived from three databases, involving pre- and during-Covid-19 pandemic data, establishes a reliable basis to construct a motivational, personalized feedback to the LMS users, so to readjust their interaction with the LMS, as an effort to increase the related QoI. The latter refers to the efficient engagement of the user with the online part of the learning process (nowadays almost the sole one), and provides educators with an evaluation path, in parallel to the content-related assessment, that could enrich their overall view about learner motivation and participation in the learning process. Moreover, based on the estimated dQoI(k) and its segmentation setting (see “Methods” section), personalized feedback can be provided to users that helps them get the most out of their interaction with the LMS and the related online material, and can have a significant impact on overall learning performance outcomes50. Many forms of representation of the segmented dQoI(k) value can be employed (e.g., text, graphs, audiovisual); in all of them, however, a personalization in the way the feedback is communicated to users should be incorporated. For example, when , a text message of ‘Serious alert! Your QoI is expected to significantly fall!’, can be used as an intense warning; alternatively, it can be more constructive in the form of ‘From now on, please try to be more focused and more active in the online part of your course’. The former textual feedback is a descriptive interpretation of the dQoI(k) value per se, whereas the latter is a proactive interpretation that motivates learners to act constructively. This need for personalised interpretation stems from the fact that learners, usually, chose different paths to respond to learning challenges. For example, the ones with a positive orientation view feedback (either positive or negative) as information to be assimilated and accommodated. However, the ones with a negative orientation, perceive negative feedback as a ‘crushing blow’ and reflection of their poor ability51; most of such learners easily give up50telegram中文最新版下载的地方是多少. Hence, the adaptation of the feedback path could better support the ultimate aim in feedback provision, which allows learners to take over the function of assessing themselves and others52.
Within the aforementioned context, yet from a more integrated perspective, the proposed DeepLMS approach can be augmented to become an ideal mechanism/feedback to support various stakeholder groups in the domain of education (including department heads, teachers, administrators, technical support staff, and learners). This can be achieved by aggregating the individual predictive user outputs. This process could lead to effective technology-enabled learning. Amongst its attributes, it should include a focus on enduring learning outcomes. This endurance is reinforced by the DeepLMS through its focusing on the QoI prediction, whose dynamic feedback to LMS users, gradually etches in them the optimized LMS interaction as an enduring learning outcome.
From the results presented in Figs. 2, 3, 4, 5, 6, 7 and 8 and Table 1, the proposed DeepLMS seems independent of the group type, as it shows a similar predictive performance both in Professors’ and Students’ QoI prediction (Wilcoxon rank sum test for DB1: ). In addition, cross-country/scale/time-period statistical analysis has resulted in non-significantly statistical differences of the performance of DeepLMS for different sociocultural and temporal settings (Wilcoxon rank sum test for : p = 0.387). The same holds for the factors of sex and age, as linear regression tests did not show any influence of both on the prediction of QoI for Professors () and Students (). Note that the statistics related to Professors were estimated for DB1 only, as the number of Professors in DB2 (3) and DB3 (1) is limited. Moreover, DeepLMS seems insensitive to the sparsity of the interaction, as it efficiently models the user’s LMS-based various interaction patterns, as expressed in the QoI time-series morphology across time (Figs. 2, 3, 5, 7). These patterns are governed by various academic calendar activities, telegram最新的中文下载网站哪里有 e.g., lectures, mid-term exams, final exams, winter/spring/summer breaks, and/or external ones, especially for DB2 and DB3, as the lockdown due to Covid-19 pandemic (DB2: 26/3-24/4/2020; DB3: 11/3-4/5/2020) lies within their time duration (see Table 2). In spite of these, the DeepLMS acknowledges such data evolution, resulting in adequate predictive performance due to the ability of its embedded LSTM forecasting model to outperform classical time series methods in cases with long, interdependent and sparse time series53. Clearly, the aforementioned results show increased generalization power in the performance of DeepLMS. Extending the demographics analysis in the bias domain, as Table 2 shows, the distribution of Male/Female was quite balanced, both in Professors and Students, along with their age, without any heterogeneity that would potentially produce data bias in LSTM learning. Hence, no historical bias (as no socio-technical issues were involved), no representation bias (sufficient number of users were involved and the significant spread of QoI data comes from users across three countries, with five courses with 30-40 different disciplines each course (macro level: DB1), one course (meso level: DB2) and one discipline (micro level: DB3)), no measurement bias (data were captured from users that all had equal access to the LMS Moodle pages after logged in), no evaluation bias (the evaluation was performed on an equal basis and with the same objective measures of (RMSE, r) as in the baseline algorithm (FCM-QoI44)), no population bias (user population represented in the datasets is coming from a real-life setting (University) end expresses the original target population), no Simpson’s Paradox (the data were homogeneous and there were no subgroups in Professors’ and Students’ groups), no sampling bias (uniform sampling across all groups), no user-interaction bias (the LMS Moodle metrics involved in the production of the QoI are 112 (see Table S1) and provide a high variety to the user to interact with the LMS Moodle), and no self-selection bias (the data were analyzed after the users interacted with the LMS and they were totally unaware of the research; hence, they could not influence the results by selective self-participation). Consequently, there is no unfairness arising from biases in the data. Taking the bias issue further, another source of unfairness could potentially arise from the learning algorithm involved itself. To avoid such event, some techniques54 could be explored to try to transform the data (pre-processing), so that the underlying discrimination is removed, or incorporating changes into the objective function or imposing a constraint (in-processing), or accessing a holdout set, which was not involved during the training of the model, and reassign the initially assigned labels by the model based on a function (post-processing). In the DeepLMS case, although no data bias was identified, in a broader perspective, flexibly fair representation in DeepLMS learning could be introduced in its future edition by creating a layer that disentangles the information that relate with sensitive attributes (e.g., demographics) and create a targeted learning for such sensitive latent variables, which potentially can bias the model, and incorporate such knowledge in a debias process (e.g., as in55,56) at the higher QoI prediction layer.
When performing a relevant comparative analysis between the DeepLMS performance and the most related FCM-QoI model44, that it is also based on the same QoI data drawn from the FuzzyQoI model42, it seems that the proposed DeepLMS achieves higher overall performance, when compared to the testing output of FCM-QoI. In particular, based on the predictive performance of both the DeepLMS and the FCM-QoI44 tabulated in Table 1, it is apparent that the DeepLMS exhibits lower testing RMSE and higher r values in its predictive output, when compared with the ones from the FCM-QoI model44. From a structural comparison, DeepLMS overcomes the training limitation of the FCM-QoI, i.e., its training is based on the mean values of QoI across users provided by the FuzzyQoI model; this, however, merges the specific characteristics of each user to an average behavior44. On the contrary, the DeepLMS provides personalised predictions for the QoI of each user across the academic semesters.
From a more general perspective, DeepLMS aligns with the previous efforts that incorporate LSTM-based predictions in the context of online education, yet not at the exact same specific problem settings as in DeepLMS. Hence, the latter is well-positioned with the approaches related to: i) cross-domains analysis, e.g., MOOCs impact in different contexts57, as DeepLMS could be easily adapted to a micro analysis of the QoI per discipline/course and transfer learning from one discipline to another at the same course (or courses with comparable content), as shown here with the application of DeepLMS to DB1-DB3, in a similar manner that was applied in MOOCs from different domains57; ii) combination of learning patterns in the context domain with the temporal nature of the clickstream data58, and identification of students at risk59, as DeepLMS could be combined with an auto-encoder to capture both the underlying behavioral patterns and the temporal nature of the interaction data at various levels of the predicted QoI (e.g., low (<0.5) QoI (at risk level)); iii) predicting learning gains by incorporating skills discovery60,61, as DeepLMS could provide the predicted QoI as an additional source of the user profile to his/her skills and learning gains; iv) user learning states and learning activities prediction from wearable devices62, as DeepLMS could easily be embedded in the expanded space of affective (a-) learning, and inform a more extended predictive model that would incorporate the learning state with the estimated QoI; v) increasing the communication of the instructional staff to learners based on individual predictions of their engagement during MOOCs63,64, as DeepLMS could facilitate the coordination of the instructor with the learner based on the informed predicted QoI; and vi) predicting the learning paths/performance65 and the teaching paths66, as the DeepLMS could be extended in the context of affecting the learning/teaching path by the predicted QoI.
Despite the promising results of the proposed DeepLMS towards prediction of the user’s LMS-based QoI, certain limitations exist. In particular, no correlation analysis with the content evaluation outcome from, e.g., quizzes, mid-/final exams, was undertaken. In fact, this was left for a future endeavor, as the focus here was to explore the predictive performance of the DeepLMS in LMS-based QoI prediction, fostering the role of the latter as an additional, to conventional grading, assessment field. Moreover, the data used here refer to one (2009/2010) or half (2020) academic year; thus, exploration of the DeepLMS application and further validation of its predictive performance upon follow-up data, i.e., monitoring of the same users across sequential academic years, would shed light upon the consistency in the predictive performance of the DeepLMS across longer time periods.
The efficient performance of the DeepLMS was validated on real data, incorporating adequate number of users and LMS data logs from different countries and educational settings. Since the structure and training of the proposed DeepLMS are not restricted to a specific course content, actually they were tested on human kinetics (DB1), engineering design (DB2), and advanced signal processing (DB3) educational contents, it could easily be expanded to the analysis of LMS data coming from various fields, e.g., Social Sciences, Medical and/or Engineering Education67最新的官方的telegram下载的网址是什么. This would allow for the exploration of any dis/similarities and correlations in the LMS users’ QoI, from an institutional perspective. Moreover, as the Covid-19 pandemic shifted the use of LMS Moodle to Secondary Education Institutions (SEIs), as well, the DeepLMS could be used for comparing the LMS-based QoI across younger student groups and explore the age-related trends in LMS-based interaction.
As part of our future work on DeepLMS, we aim to perform a fusion of other measures of user’s quality in the online learning context at both SEIs and HEIs. This includes prediction of the Quality of Collaboration (QoC)68 and Quality of Affective Engagement (QoAE)1,69, in an effort to predict, in a holistic way, the various components that play significant role in the learning process, i.e., interaction, collaboration and affectiveness1. The incorporation of Deep Learning-based predictions of QoC and QoAE, in parallel to the QoI ones, extends the work of the authors70,71,72 from the concept of affective/blended/collaborative-teaching/learning (a/b/c-TEACH, http://abcteach.fmh.ulisboa.pt/) to the a/b/c/d(eep)-TEACH one. In the midst of the Covid-19 pandemic, such an AI-based scaffolding helps educators and learners move from quick fixes, and their possible consequence of regressing to poor practice, to maximum efficiency of the online learning tools available and truly support learning. Finally, as distinct time periods of pre-, during- and post-Covid-19 lockdown have been formed, the analysis of LMS data that emerged during these three periods seems promising, in particular for the identification of any effect on the QoI per se and its related prediction via the proposed DeepLMS model. This analysis will allow for further evaluation of the DeepLMS model predictive robustness against effects caused by time-related disruptors, such as Covid-19, in the context of education; ongoing efforts towards such direction are reported in73.