Pattern recognition and mixed practice

Why it matters

Interview problems come from a large space, and you will rarely meet one you have practiced. What transfers is the ability to read a statement, notice its signals and test a few candidate techniques against their preconditions. Memorizing solutions, or lists of questions said to be used by particular companies, does not build that ability and fails as soon as a problem is altered. Learning research also generally favors interleaved practice and spaced review over blocked repetition: mixing problem types forces you to choose the technique, which is precisely the skill being tested.

Core ideas

Signals, candidates, conditions. Read the constraints and the question, map them to a short list of candidates, and check each candidate's condition before committing.

Signal	Candidate patterns	Condition or invariant to check
Complement lookup, duplicates, grouping	hashing, frequency counting	Equal keys identify equivalent items
Pairs in sorted input, in-place partitioning	two pointers	Each move discards candidates that cannot improve the answer
Contiguous subarray or substring	sliding window, prefix sums	Window validity is monotone (often needs non-negative values); otherwise prefix sums with a hash map
Sorted input, or smallest feasible value	binary search, binary search on the answer	A predicate that flips from false to true exactly once
Overlapping ranges, many range updates	sorting, interval merging, sweep line	Sort key matches the objective
Nested structure, matching, undo	stack	Top is the most recent unresolved item
Next greater or smaller, window extremes	monotonic stack or deque	Stored indices stay in monotone order
k largest, running median, merging streams	heap selection, two heaps	Heap top is the current extreme; size stays bounded
Reordering a list, cycles, middles	linked-list pointers, fast and slow pointers	Save next before rewiring; fast moves twice as far
Hierarchy, subtree answers	tree DFS or BFS, BST invariant	Each subtree returns a summary its parent can combine
Many prefix queries over words	trie	Each node represents one prefix
Minimum steps, grids, transformations	graph BFS or DFS, grid as implicit graph, multi-source BFS	Unit weights; mark visited on enqueue
Prerequisites, dependencies	topological ordering, cycle detection	An order exists only for a DAG
Groups merging over time	union-find	Undirected, unions only
Cheapest weighted route	shortest paths	Weight sign decides Dijkstra or Bellman-Ford
All combinations or placements	backtracking	Choose, recurse, undo; prune early
Count ways, best total over choices	DP (1D, grid, strings, knapsack)	Overlapping subproblems with a valid evaluation order
A locally best choice looks safe	greedy	An exchange argument exists
Independent halves, k-th order statistic	divide and conquer	Subproblems do not overlap; write the recurrence
Parity, small sets encoded as masks	bit manipulation	Width and signedness are explicit
Fast operations on changing data	stateful design	Map plus list, heap or sorted structure kept consistent

Constraints narrow the list: n up to about 20 allows exponential search, n around 10^3 allows O(n^2), and n around 10^5 or more needs O(n log n) or better.

Mixed practice routine. Once several topics are studied, stop practicing only by topic. A session is three unseen problems from different patterns, 25 to 35 minutes each, with the pattern unknown in advance; the official LeetCode 75 and Top Interview 150 plans and the extra items in each topic are good pools. Before coding, write the signals you noticed and two candidates with the reason for your choice. Afterwards, log the problem, the pattern, whether you recognized it unaided, the time taken and the bug class.

Review rotation. Revisit each pattern at growing intervals, for example after about 2, 7 and 21 days, through a different problem each time. If you learned sliding windows on Longest Substring Without Repeating Characters, review with Permutation in String, then Minimum Window Substring; for BFS, follow Number of Islands with Rotting Oranges, then 01 Matrix. Re-solving the same problem mostly tests memory of that answer. Patterns you recognized late or solved with bugs come back sooner.

Stateful design problems. LRU Cache, Time Based Key-Value Store, Snapshot Array and LFU Cache combine patterns: a hash map for direct access plus a linked list, heap or sorted list for order. State the invariant that links the structures and the cost of every operation.

Common misconceptions

• "The strategy is to memorize the questions a company uses." Such lists cannot be verified, and you will rarely see the same problem; recognition and clean execution transfer.
• "More problems solved means more progress." Many problems practiced by topic, always knowing the pattern beforehand, build less recognition than fewer problems practiced mixed.
• "Review means re-solving the same problem." That rehearses one answer; review the pattern through a new problem.
• "A signal determines the pattern." Signals suggest candidates; the applicability condition decides.

How to approach it

Start mixed practice once arrays and hashing, two pointers, trees, graph traversal and basic DP are studied, and run two or three mixed sessions per week alongside topic study. When a pattern fails repeatedly, return to its topic notes instead of grinding more problems of the same kind. Mid and senior loops typically add design-flavored problems and follow-up extensions, so keep stateful design in the rotation.