A datum represents an observation, category or value. Interpretation requires knowing what it describes and how it was obtained. Metadata provide context: origin, time, scale, method or transformations. Not all are needed in every project; choose them according to the question. A relative light reading and a measurement in lux cannot be compared as the same quantity. A file’s editing date is not necessarily the moment when the phenomenon was observed either.

Big Data involves datasets and problems whose management requires scalable resources and methods. Volume, variety and generation speed help explain the need: many records, different formats and continuous inputs can make timely storage and analysis difficult. No row count alone turns our table into a Big Data infrastructure. The school sample teaches a shared process: ask, describe provenance, prepare data, analyse and communicate limitations. Relate that process to a real service without confusing a small demonstration with a large platform’s capabilities.

A data dictionary defines each field: name, meaning, type, unit or scale and permitted values. A zone is categorical; a reading may be numerical; an instant expresses time. An identifier recognises a record but must not contain personal information. Equal readings are not necessarily duplicates: they may represent separate observations. Identify an exact duplicate using a rule covering identifier and fields, checking provenance before removing a copy.

Cleaning prepares a working copy without changing the original. Record action, reason and effect on rows or calculations. Missing information is not zero; an out-of-scale value requires checking the source rather than intuitive correction. For a mean, sum only valid values and divide by their count, reporting excluded missing values. Do not hide data contradicting expectations. Cleaning must make decisions reconstructable and distinguish corrected errors from remaining uncertainty. Another pair can then review whether the chart actually represents the described sample.

A chart represents selected relationships from a table. Bars support category comparisons; lines can show evolution when order is meaningful, such as time. An increasing identifier does not guarantee equal time intervals. Before choosing a type, write the question and what each axis represents. When comparing zone means, also report the valid observation count for each zone. A chart cannot recover context missing from the data.

Scale changes visual impressions. For bars, starting at zero usually supports proportional interpretation; a visible truncation may serve another purpose but must be justified without hiding its effect. Compare two representations of the same dataset to distinguish appearance from conclusion. Association does not demonstrate causation: quantities may vary together because of another factor or sample selection. Limit conclusions to observed conditions and identify additional data needed. Keep table and chart linked so ranges, filters and labels remain reviewable instead of submitting only an image without provenance.

Supervised learning starts with examples accompanied by labels identifying their reference classes. During training, the model adjusts relationships between input features and those labels. For another example, it uses what it has learned to produce a prediction, without understanding the object as a person does. Those relationships depend on the available examples: if a class always appears against the same background, the model may use that as a cue rather than distinguish the intended feature. Quantity, diversity and label consistency influence this learning.

Testing examines what happens with examples that were not used for training. Reserving them beforehand and documenting how they were obtained separates the two purposes. Renaming a photograph or extracting a nearly identical image from the same capture does not guarantee independence. The workshop uses twenty training examples per class and ten new tests, five per class. These quantities organise the exercise and support comparable records, but do not guarantee general performance. Materials consist of objects or non-identifying sounds, with no recognisable people, using the approved tool.

Test errors can also guide improvements to training. Once used for that purpose, the series has influenced the design: it no longer provides an independent check of the chosen improvement. Its exploratory role is therefore retained, and another independent series is prepared after the change is selected. This separation distinguishes what was learned by reviewing errors from what was observed when evaluating the model again.

A results matrix organises each test using two pieces of information: the actual class, known as the reference, and the class predicted by the model. In our representation, rows correspond to actual classes and columns to predictions; labelling them prevents reversed interpretations. Matches lie on the diagonal and the other cells count confusions. A row total shows how many actual examples of that class were tested. A column total counts how often the model answered with that class, rather than how many objects actually belonged to it.

Percentages answer different questions depending on their denominators. Overall accuracy divides all matches by the total number of tests. Class-level accuracy divides a row’s matches by the actual examples of that class. With different class sizes, a high overall result can hide failures concentrated in one class. Keeping the matrix, totals and denominators reveals that distribution and supports deciding which cases deserve review.

The score or confidence displayed for a prediction concerns that example, rather than the percentage of correct answers observed in the test. It does not guarantee truth: a high-scoring response still counts as an error if it contradicts the reference. The record retains those errors and their scores too. The direction of confusion also matters, because putting material in the wrong box and leaving it unsorted can have different consequences. The small sample guides improvements; it neither demonstrates performance under all future conditions nor justifies important decisions about people.

A system may work technically yet be inappropriate for a purpose. Ask which decision it supports, who is affected by error and who can review results. Classroom-object classification studies patterns without classifying people or allocating opportunities. Choose only necessary information and check that open material contains no personal data before reusing it. Removing names or replacing them with codes does not by itself prove that other columns or images cannot identify someone.

The report connects question, provenance, cleaning, representation, training and evaluation. A responsible conclusion explains two limitations and a usage condition, linking them to evidence. Bias may arise from collected examples, chosen labels or missing situations. Adding repeated data does not automatically remove it. Formulate a concrete hypothesis and testable improvement, such as diversifying a missing orientation. Retain human review outside tested conditions. Recognising uncertainty does not make the work worthless: it determines which part of the demonstration is useful and which cannot extend to other contexts.

The laboratory connects table analysis and classification through a shared question and reviewable record. There is no need to pretend both files contain the same data types: explain each contribution to the school need. Preserve original, cleaned copy, dictionary, charts and model results. A test sheet identifies example, actual class, prediction and capture conditions; another summarises the matrix and denominators without replacing individual records.

To study reduced examples in one class, maintain comparable conditions and document each version’s training. Using one series to compare versions may support exploration, but that series influences selection of the improvement. Then run a fresh independent evaluation. Do not attribute every variation to quantity if background, labels or tool also changed. A01 combines the question, plan and first completed cleaning, charting, training and testing actions. A02 adds full analysis, class results, evaluated improvement and limitations. Each partner explains a decision and a data-supported conclusion within the same two assignments.

Documentation checks terms and procedures without turning an example figure into your own result. Google’s dataset reference explains that tests can wear out when repeatedly used to adjust decisions. Its classification reference shows why overall accuracy depends on class distribution. Google Sheets help guides chart types. Consult each resource with a specific question and connect its explanation to the operations and conditions actually used in class.

Teachable Machine trains and runs models using browser technology, but that does not justify claiming that every configuration, save or export always keeps all data in one place. The teacher checks the approved environment and permitted examples are used. Preserve the project or model through authorised options and record reopening instructions; do not publish it automatically just to obtain evidence. Videos are optional and may be supported by text. To transfer the model to another object or condition, review classes, diversity, error consequences and fresh evaluation. Explain what the test would teach without promising that more data guarantee improvement.